This is an industry we're[0] in. Owning is at one end of the spectrum, with cloud at the other, and a broadly couple of options in-between:
1 - Cloud – This is minimising cap-ex, hiring, and risk, while largely maximising operational costs (its expensive) and cost variability (usage based).
2 - Managed Private Cloud - What we do. Still minimal-to-no cap-ex, hiring, risk, and medium-sized operational cost (around 50% cheaper than AWS et al). We rent or colocate bare metal, manage it for you, handle software deployments, deploy only open-source, etc. Only really makes sense above €$5k/month spend.
3 - Rented Bare Metal – Let someone else handle the hardware financing for you. Still minimal cap-ex, but with greater hiring/skilling and risk. Around 90% cheaper than AWS et al (plus time).
4 - Buy and colocate the hardware yourself – Certainly the cheapest option if you have the skills, scale, cap-ex, and if you plan to run the servers for at least 3-5 years.
A good provider for option 3 is someone like Hetzner. Their internal ROI on server hardware seems to be around the 3 year mark. After which I assume it is either still running with a client, or goes into their server auction system.
Options 3 & 4 generally become more appealing either at scale, or when infrastructure is part of the core business. Option 1 is great for startups who want to spend very little initially, but then grow very quickly. Option 2 is pretty good for SMEs with baseline load, regular-sized business growth, and maybe an overworked DevOps team!
I think the issue with this formulation is what drives the cost at cloud providers isn't necessarily that their hardware is too expensive (which it is), but that they push you towards overcomplicated and inefficient architectures that cost too much to run.
A core at this are all the 'managed' services - if you have a server box, its in your financial interest to squeeze as much per out of it as possible. If you're using something like ECS or serverless, AWS gains nothing by optimizing the servers to make your code run faster - their hard work results in less billed infrastructure hours.
This 'microservices' push usually means that instead of having an on-server session where you can serve stuff from a temporary cache, all the data that persists between requests needs to be stored in a db somewhere, all the auth logic needs to re-check your credentials, and something needs to direct the traffic and load balance these endpoint, and all this stuff costs money.
I think if you have 4 Java boxes as servers with a redundant DB with read replicas on EC2, your infra is so efficient and cheap that even paying 4x for it rather than going for colocation is well worth it because of the QoL and QoS.
These crazy AWS bills usually come from using every service under the sun.
The complexity is what gets you. One of AWS's favorite situations is
1) Senior engineer starts on AWS
2) Senior engineer leaves because our industry does not value longevity or loyalty at all whatsoever (not saying it should, just observing that it doesn't)
3) New engineer comes in and panics
4) Ends up using a "managed service" to relieve the panic
5) New engineer leaves
6) Second new engineer comes in and not only panics but
outright needs help
7) Paired with some "certified AWS partner" who claims to help "reduce cost" but who actually gets a kickback from the extra spend they induce (usually 10% if I'm not mistaken)
Calling it it ransomware is obviously hyperbolic but there are definitely some parallels one could draw
On top of it all, AWS pricing is about to massively go up due to the RAM price increase. There's no way it can't since AWS is over half of Amazon's profit while only around 15% of its revenue.
Just this week a friend of mine was spinning up some AWS managed service, complaining about the complexity, and how any reconfiguration took 45 minutes to reload. It's a service you can just install with apt, the default configuration is fine. Not only is many service no longer cheaper in the cloud, the management overhead also exceed that of on-prem.
> If you're using something like ECS or serverless, AWS gains nothing by optimizing the servers to make your code run faster - their hard work results in less billed infrastructure hours.
If ECS is faster, then you're more satisfied with AWS and less likely to migrate. You're also open to additional services that might bring up the spend (e.g. ECS Container Insights or X-Ray)
I don’t understand why most cloud backend designs seem to strive for maximizing the number of services used.
My biggest gripe with this is async tasks where the app does numerous hijinks to avoid a 10 minute lambda processing timeout. Rather than structure the process to process many independent and small batches, or simply using a modest container to do the job in a single shot - a myriad of intermediate steps are introduced to write data to dynamo/s3/kinesis + sqs/and coordination.
A dynamically provisioned, serverless container with 24 cores and 64 GB of memory can happily process GBs of data transformations.
Fully agree to this. I find the cost of cloud providers is mostly driven by architecture. If you're cost conscious, cloud architectures need to be up-front designed with this in mind.
Microservices is a killer with cost. For each microservices pod
- you're often running a bunch of side cars - datadog, auth, ingress
- you pay massive workload separation overhead with orchestration, management, monitoring and ofc complexity
I am just flabbergasted that this is how we operate as a norm in our industry.
It's about fitting your utilization to the model that best serves you.
If you can keep 4 "Java boxes" fed with work 80%+ of the time, then sure EC2 is a good fit.
We do a lot of batch processing and save money over having EC2 boxes always on. Sure we could probably pinch some more pennies if we managed the EC2 box uptime and figured out mechanisms for load balancing the batches... But that's engineering time we just don't really care to spend when ECS nets us most of the savings advantage and is simple to reason about and use.
Agreed. There is a wide price difference between running a managed AWS or Azure MySQL service and running MySQL on a VM that you spin up in AWS or Azure.
Great comment. I agree it's a spectrum and those of us who are comfortable on (4) like yourself and probably us at Carolina Cloud [0] as well, (4) seems like a no brainer. But there's a long tail of semi-technical users who are more comfortable in 2-3 or even 1, which is what ultimately traps them into the ransomware-adjacent situation that is a lot of the modern public cloud. I would push back on "usage-based". Yes it is technically usage-based but the base fee also goes way up and there are also sometimes retainers on these services (ie minimum spend). So of course "usage-based" is not wrong but what it usually means is "more expensive and potentially far more expensive".
The problem is that clouds have easily become 3 or 5 times the price of managed services, 10x the price of option 3, and 20x the price of option 4. To say nothing of the fact that almost all businesses can run fine on "pc under desk" type situations.
So in practice cloud has become the more expensive option the second your spend goes over the price of 1 engineer.
Hetzner is definitely an interesting option. I’m a bit scared of managing the services on my own (like Postgres, Site2Site VPN, …) but the price difference makes it so appealing. From our financial models, Hetzner can win over AWS when you spend over 10~15K per month on infrastructure and you’re hiring really well. It’s still a risk, but a risk that definitely can be worthy.
> I’m a bit scared of managing the services on my own
I see it from the other direction, when if something fails, I have complete access to everything, meaning that I have a chance of fixing it. That's down to hardware even. Things get abstracted away, hidden behind APIs and data lives beyond my reach, when I run stuff in the cloud.
Security and regular mistakes are much the same in the cloud, but I then have to layer whatever complications the cloud provide comes with on top. If cost has to be much much lower if I'm going to trust a cloud provider over running something in my own data center.
You sum it up very neatly. We've heard this from quite a few companies, and that's kind of why we started our ours.
We figured, "Okay, if we can do this well, reliably, and de-risk it; then we can offer that as a service and just split the difference on the cost savings"
(plus we include engineering time proportional to cluster size, and also do the migration on our own dime as part of the de-risking)
I've just shifted my SWE infrastructure from AWS to Hetzner (literally in the last month). My current analysis looks like it will be about 15-20% of the cost - £240 vs 40-50 euros.
Expect a significant exit expense, though, especially if you are shifting large volumes of S3 data. That's been our biggest expense. I've moved this to Wasabi at about 8 euros a month (vs about $70-80 a month on S3), but I've paid transit fees of about $180 - and it was more expensive because I used DataSync.
Retrospectively, I should have just DIYed the transfer, but maybe others can benefit from my error...
I’m wondering if it makes sense to distribute your architecture so that workers who do most of the heavy lifting are in hetzner, while the other stuff is in costly AWS. On the other hand this means you don’t have easy access to S3, etc.
I don't know. I rent a bare metal server for $500 a month, which is way overkill. It takes almost no time to manage -- maybe a few hours a year -- and can handle almost anything I throw at it. Maybe my needs are too simple though?
Dead on. Recently, 3 and 4 have been compelling. Cloud costs have rocketed up. I started my casual transition to co-lo 2 years ago and just in december finished everything. I have more capacity at about 30% of the cost. If you go option 3, you even get the benefit of 6+ month retro pricing for RAM/storage. I'm running all DDR4, but I have so much of it I don't know what to do with it.
The flip side is that compliance is a little more involved. Rather than, say, carve out a whole swathe of SOC-2 ops, I have to coordinate some controls. It's not a lot, and it's still a lot lighter than I used to do 10+ years ago. Just something to consider.
5 - Datacenter (DC) - Like 4, except also take control of the space/power/HVAC/transit/security side of the equation. Makes sense either at scale, or if you have specific needs. Specific needs could be: specific location, reliability (higher or lower than a DC), resilience (conflict planning).
There are actually some really interesting use cases here. For example, reliability: If your company is in a physical office, how strong is the need to run your internal systems in a data centre? If you run your servers in your office, then there's no connectivity reliability concerns. If the power goes out, then the power is out to your staff's computers anyway (still get a UPS though).
Or perhaps you don't need as high reliability if you're doing only batch workloads? Do you need to pay the premium for redundant network connections and power supplies?
If you want your company to still function in the event of some kind of military conflict, do you really want to rely on fibre optic lines between your office and the data center? Do you want to keep all your infrastructure in such a high-value target?
I think this is one of the more interesting areas to think about, at least for me!
> 4 - Buy and colocate the hardware yourself – Certainly the cheapest option if you have the skills, scale, cap-ex, and if you plan to run the servers for at least 3-5 years.
Is it still the cheapest after you take into account that skills, scale, cap-ex and long term lock-in also have opportunity costs?
An interesting question, so time for some 100% speculation.
It sounds like they probably have revenue in the €500mm range today. And given that the bare metal cost of AWS-equivalent bills tends to be a 90% reduction, we'll say a €10mm+ bare metal cost.
So I would say a cautious and qualified "yes". But I know even for smaller deployments of tens or hundreds of servers, they'll ask you what the purpose is. If you say something like "blockchain," they're going to say, "Actually, we prefer not to have your business."
I get the strong impression that while they naturally do want business, they also aren't going to take a huge amount of risk on board themselves. Their specialism is optimising on cost, which naturally has to involve avoiding or mitigating risk. I'm sure there'd be business terms to discuss, put it that way.
Netflix might be spending as much as $120m (but probably a little less), and I thought they were probably Amazon's biggest customer. Does someone (single-buyer) spend more than that with AWS?
Hertzner's revenue is somewhere around $400m, so probably a little scary taking on an additional 30% revenue from a single customer, and Netflix's shareholders would probably be worried about risk relying on a vendor that is much smaller than them.
Sometimes if the companies are friendly to the idea, they could form a joint venture or maybe Netflix could just acquire Hertzner (and compete with Amazon?), but I think it unlikely Hertzner could take on Netflix-sized for nontechnical reasons.
However increasing pop capacity by 30% within 6mo is pretty realistic, so I think they'd probably be able to physically service Netflix without changing too much if management could get comfortable with the idea
This space of #2 like Lithus is not something I'm very familiar with, so thank you for the comment that piqued my interest!
If you're willing to share, I'm curious who else you would describe as being in this space.
My last decade and a half or so of experience has all been in cloud services, and prior to that it was #3 or #4. What was striking to me when I went to the Lithus website was that I couldn't figure out any details without hitting a "Schedule a Call" button. This makes it difficult for me to map my experiences in using cloud services onto what Lithus offers. Can I use terraform? How does the kubernetes offering work? How does the ML/AI data pipelines work? To me, it would be nice if I could try it out in a very limited way as self-service, or at least read some technical documentation. Without that, I'm left wondering how it works. I'm sure this is a conscious decision to not do this, and for good reasons, but I thought I'd share my impressions!
Hello! I think this is a fair question, and improving the communication on the website is something that is steadily climbing up our priority list.
We're not really that kind of product company; we're more of a services company. What we do is deploy Kubernetes clusters onto bare metal servers. That's the core technical offering. However, everything beyond that is somewhat per-client. Some clients need a lot of compute. Some clients need a custom object storage cluster. Some clients need a lot of high-speed internal networking. Which is why we prefer to have a call to figure out specifically what your needs are. But I can also see how this isn't necessarily satisfying if you're used to just grabbing the API docs and having a look around.
What we will do is take your company's software stack and migrate it off AWS/Azure/Google and deploy it onto our new infrastructure. We will then become (or work with) your DevOps team to supporting you. This can be anything from containerising workloads to diagnosing performance issues to deploying a new multi-region Postgres cluster. Whatever you need done on your hardware that we feel we can reasonably support. We are the ones on-call should NATS fall over at 4am.
Your team also has full access to the Kubernetes cluster to deploy to as you wish.
I think the pricing page is the most concrete thing on our website, and it is entirely accurate. If you were to phone us and say, "I want that exact hardware," we would do it for you. But the real value we also offer is in the DevOps support we provide, actually doing the migration up-front (at our own cost), and being there working with your team every week.
Unfortunately, (successful) startups can quickly get trapped into this option. If they're growing fast, everyone on the board will ask why you'd move to another option at the first place. The cloud becomes a very deep local minimum that's hard to get out off.
Can someone explain 2 to me. How is a managed private cloud different from full cloud? Like you are still using AWS or Azure but you are keeping all your operation in a bundled, portable way, so you can leave that provider easily at any time, rather than becoming very dependent on them? Is it like staying provider-agnostic but still cloud based?
To put it plainly: We deploy a Kubernetes cluster on Hetzner dedicated servers and become your DevOps team (or a part thereof).
It works because bare metal is about 10% the cost of cloud, and our value-add is in 1) creating a resilient platform on top of that, 2) supporting it, 3) being on-call, and 4) being or supporting your DevOps team.
This starts with us providing a Kubernetes cluster which we manage, but we also take responsibility for the services run on it. If you want Postgres, Redis, Clickhouse, NATS, etc, we'll deploy it and be SLA-on-call for any issues.
If you don't want to deal with Kubernetes then you don't have to. Just have your software engineers hand us the software and we'll handle deployment.
Everything is deployed on open source tooling, you have access to all the configuration for the services we deploy. You have server root access. If you want to leave you can do.
Our customers have full root access, and our engineers (myself included) are in a Slack channel with you engineers.
And, FWIW, it doesn't have to be Hetzner. We can colocate or use other providers, but Hetzner offer excellent bang-per-buck.
Edit: And all this is included in the cluster price, which comes out cheaper than the same hardware on the major cloud providers
Instead of using the Cloud's own Kubernetes service, for example, you just buy the compute and run your own Kubernetes cluster. At a certain scale that is going to be cheaper if you have to know how. And since you are no longer tied to which services are provided and you just need access to compute and storage. you can also shop around for better prices than Amazon or Azure since you can really go to any provider of a VPS.
Getting rid of bureaucratic internal IT department is a game changer for productivity. That alone is worth 10x infra costs, especially for big companies where work can grind to a halt dealing with obstructionists through service now. Good leaders understand this.
Sadly true. Or, the so-called internal IT Dept. can be a shambolic mess of PHB's, Brunchlords, Catberts, metric maximizers, and micromanagers, presiding over the hollowed-out and burned out remains of the actual workforce that you'd need to reliably do the job.
I am using something inbetween 2 and 3, a hosted Web-site and database service with excellent customer support. On shared hardware it is 22 €/month. A managed server on dedicated hardware starts at about 50 €/month.
Where do AWS reserved instances come into your hierarchy? What if there existed a “perpetual” reserved instance? Is cap-ex vs. op-ex really the key distinction?
Been using Hetzner Cloud for Kubernetes and generally like it, but it has its limitations. The network is highly unpredictable. You at best get 2Gbit/s, but at worst a few hundreds of Mbit/s.
Is that for the virtual private network? I heard some people say that you actually get higher bandwidth if you're using the public network instead of the private network within Hetzner, which is a little bit crazy.
> Buy and colocate the hardware yourself – Certainly the cheapest option if you have the skills
back then this type of "skill" was abundant. You could easily get sysadmin contractors who would take a drive down to the data-center (probably rented facilities in a real-estate that belonged to a bank or insurance) to exchange some disks that died for some reason. such a person was full stack in a sense that they covered backups, networking, firewalls, and knew how to source hardware.
the argument was that this was too expensive and the cloud was better. so hundreds of thousands of SME's embraced the cloud - most of them never needed Google-type of scale, but got sucked into the "recurring revenue" grift that is SaaS.
If you opposed this mentality you were basically saying "we as a company will never scale this much" which was at best "toxic" and at worst "career-ending".
The thing is these ancient skills still exist. And most orgs simply do not need AWS type of scale. European orgs would do well to revisit these basic ideas. And Hetzner or Lithus would be a much more natural (and honest) fit for these companies.
I wonder how much companies pay yearly in order to avoid having an employee pick up a drive from a local store, drive to the data center, pull the disk drive, screw out the failing hard drive and put in the new one, add it in the raid, verify the repair process has started, and then return to the office.
It baffles me that my career trajectory somehow managed to insulate me from ever having to deal with the cloud, while such esoteric skills as swapping a hot swap disk or racking and cabling a new blade chassis are apparently on the order of finding a COBOL developer now. Really?
I can promise you that large financial institutions still have datacenters. Many, many, many datacenters!
if someone on the DevOps team knows Nix, option 3 becomes a lot cheaper time-wise! yeah, Nix flakes still need maintenance, especially on the `nixos-unstable` branch, but you get the quickest disaster recovery route possible!
plus, infra flexibility removes random constraints that e.g. Cloudflare Workers have
There are a bunch of ways to manage bare metal servers apart from Nix. People have been doing it for years. Kickstart, theforeman, maas, etc, [0]. Many to choose from according to your needs and layers you want them to manage.
Reality is these days you just boot a basic image that runs containers
Indeed! We've yet to go down this route, but it's something we're thinking on. A friend and I have been talking about how to bring Nix-like constructs to Kubernetes as well, which has been interesting. (https://github.com/clotodex/kix, very much in the "this is fun to think about" phase)
Option 4 as well, that's how we do it at work and it's been great. However, it can't really be "someone on the team knows Nix", anyone working on Ops will need Nix skills in order to be effective.
Everything comes circle. Back in my day, we just called it a "data center". Or on-premise. You know, before the cloud even existed. A 1990s VP of IT would look at this post and say, what's new? Better computing for sure. Better virtualization and administration software, definitely. Cooling and power and racks? More of the same.
The argument made 2 decades ago was that you shouldn't own the infrastructure (capital expense) and instead just account for the cost as operational expense (opex). The rationale was you exchange ownership for rent. Make your headache someone else's headache.
The ping pong between centralized vs decentralized, owned vs rented, will just keep going. It's never an either or, but when companies make it all-or-nothing then you have to really examine the specifics.
There's a very interesting insight from your message.
The Cloud providers made a lot of sense to finance departments since aside from the promised savings, you would take that cloud expense now and lower your tax rate.
After the passing of the One Beautiful Bill ("OBB"), the law allows you to accelerate CapEx to instead expense it[1], similar to the benefit given by cloud service providers.
This puts way more wind on the sails of the on-prem movement, for sure
> you shouldn't own the infrastructure (capital expense) and instead just account for the cost as operational expense (opex)
That was part of the reason.
The real reason was the internal infrastructure team in many orgs got nowhere. There was a huge queue and many teams instead had to find infinite workarounds including standing up their own. The "cloud" provided a standardized way to at least deal with this mess e.g. single source of billing.
> A 1990s VP of IT would look at this post and say, what's new?
Speed. The US lives in luxury but outside of that it often takes a LONG time to get proper servers. You don't just go online. There are many places where you have to talk to a vendor with no list price and the drama continues. Being out of capacity can mean weeks to months before you get anywhere.
Yep! The biggest win for me when AWS came out was that I could self-serve what I needed and put it on a credit card, rather than filing a ticket and waiting some number of days / weeks / months to get a new VM approved and deployed.
I agree - my reference to the 1990s VP of IT was looking at the post, which is about on-premise data centers... not the cloud. I don't think there's a speed advantage for on-premise data centers now vs the 1990s, but if there is let me know. Otherwise, indeed, it's a 1990s-era blast from the past.
Agreed. Also, a realistic assessment should not downplay the very real overhead and headache of managing your on-premise data center. It comes at a cost in engineering/firefighting hours, it's not painless. There's a reason this eternal ping pong keeps going on!
Yeah, I think the major improvement of cloud services was the rationalization of them into services with a cost instead of "ask that person for a whatsit" and "hopefully the associate goomba will approve."
All teams will henceforth expose their data and functionality through service interfaces
>San Diego has a mild climate and we opted for pure outside air cooling. This gives us less control of the temperature and humidity, but uses only a couple dozen kW. We have dual 48” intake fans and dual 48” exhaust fans to keep the air cool. To ensure low humidity (<45%) we use recirculating fans to mix hot exhaust air with the intake air. One server is connected to several sensors and runs a PID loop to control the fans to optimize the temperature and humidity.
Oh man, this is bad advice. Airborn humidity and contaminants will KILL your servers on a very short horizon in most places - even San Diego. I highly suggest enthalpy wheel coolers (kyotocooling is one vendor - switch datacenters runs very similar units on their massive datacenters in the Nevada desert) as they remove the heat from the indoor air using outdoor air (+can boost slightly with an integrated refrigeration unit to hit target intake temps) without switching the air from one side to the other. This has huge benefits for air control quality and outdoor air tolerance and a single 500KW heat rejection unit uses only 25KW of input power (when it needs to boost the AC unit's output). You can combine this with evaporative cooling on the exterior intakes to lower the temps even further at the expense of some water consumption (typically far cheaper than the extra electricity to boost the cooling through an hvac cycle).
Not knocking the achievement just speaking from experience that taking outdoor air (even filtered + mixed) into a datacenter is a recipe for hardware failure and the mean time to failure for that is highly dependant on your outdoor air conditions. I've run 3MW facilities with passive air cooling and taking outdoor air directly into servers requires a LOT more conditioning and consideration than is outlined in this article.
Yes, it's easy to destroy the servers with a lot of dust and/or high humidity. But with filtering and ensuring humidity never exceeds 45% we've had pretty good results.
I remember visiting a small data center (about half the size of the Comma one) where shoe covers were required. Apparently they were worried about people’s shoes bringing in dust and other contamination.
It's not a static number as it's also based on ambient air temperature in the form of dew point - 45% RH at low temps can be far more dangerous than 65% RH at warm ambient.
Likewise the impact on server longevity is not a finite boundary but rather "exposure over time" gradient that, if exceeding the "low risk" boundary (>-12'C/10'f dew point or >15'C/59'f dry bulb temp) results in lower MTBF than design. This is defined (and server equipment manufacturers conform and build to) ASHRAE TC 9.9. This mean - if you're running your servers above high risk curve for humidity and temperature, you're shortening the life considerably compared to low risk curve.
Generally, 15% RH is considered suboptimal and can be dangerous near freezing temperatures - in San Diego in January there were several 90%+RH scenarios that would have been dangerous for servers even when mixed down with warm exhaust air - furthermore, the outdoor air at 76'f during that period means you have limited capacity to mix in warm exhaust air (which btw came from that same 99%RH input air) without getting into higher-than-ideal intake temps.
Any dew points above 62.5'f are considered high risk for servers - as are any intake temps exceeding 32'C/90'f. You want to be on the midpoint between those and 16'C/65'f temps & -12'C/10'f dew point to have no impact on server longevity or MTBF rates.
As a recent example:
KCASANDI6112 - January 2, 2026
High Low Average
Temperature 73.4 °F 59.9 °F 63.5 °F
Dew Point 68.0 °F 60.0 °F 62.6 °F
Humidity 99 % 81 % 96 %
Precipitation 0.12 in -- --
Lastly, air contaminants - in the form of dust (that can be filtered out) and chemicals (which can't without extensive scrubbing) are probably the most detrimental to server equipment if not properly managed, and require very intentional and frequent filter changes (typically high MERV pleated filters changed on a time or pressure drop signal) to prevent server degradation and equipment risks.
The last consideration is fire suppression - permitted datacenters usually require compliance with separate fire code, such that direct outdoor air exchange without active shutdown and dry suppression is not permitted - this is to prevent a scenario where your equipment catches on fire and a constant supply of fresh oxygen-rich outdoor air turns that into an inferno. Smoke detection systems don't operate well with outdoor-mixed air or any level of airborn particulates.
So - for those reasons - among a few others - open air datacenters are not recommended unless you're doing them at google or meta scale, and in those scenarios you typically have much more extensive systems and purpose-designed hardware in order to operate for the design life of the equipment without issues.
LOL’ed IRL at “ In a future blog post I hope I can tell you about how we produce our own power and you should too.” Producing own power as a pre-requisite for running on-prem is a non-starter for many.
I would suggest to use both on-premise hardware and cloud computing. Which is probably what comma is doing.
For critical infrastructure, I would rather pay a competent cloud provider than being responsible for reliability issues. Maintaining one server room in the headquarters is something, but two servers rooms in different locations, with resilient power and network is a bit too much effort IMHO.
For running many slurm jobs on good servers, cloud computing is very expensive and you sometimes save money in a matter of months. And who cares if the server room is a total loss after a while, worst case you write some more YAML and Terraform and deploy a temporary replacement in the cloud.
Another thing between is colocation, where you put hardware you own in a managed data center. It’s a bit old fashioned, but it may make sense in some cases.
I can also mention that research HPCs may be worth considering. In research, we have some of the world fastest computers at a fraction of the cost of cloud computing. It’s great as long as you don’t mind not being root and having to use slurm.
I don’t know in USA, but in Norway you can run your private company slurm AI workloads on research HPCs, though you will pay quite a bit more than universities and research institutions. But you can also have research projects together with universities or research institutions, and everyone will be happy if your business benefits a lot from the collaboration.
> but two servers rooms in different locations, with resilient power and network is a bit too much effort IMHO
I worked in a company with two server farms (a main and a a backup one essentially) in Italy located in two different regions and we had a total of 5 employees taking care of them.
We didn't hear about them, we didn't know their names, but we had almost 100% uptime and terrific performance.
There was one single person out of 40 developers who's main responsibility were deploys, and that's it.
It costed my company 800k euros per year to run both the server farms (hardware, salaries, energy), and it spared the company around 7-8M in cloud costs.
Now I work for clients that spend multiple millions in cloud for a fraction of the output and traffic, and I think employ around 15+ dev ops engineers.
> I would rather pay a competent cloud provider than being responsible for reliability issues.
Why do so many developers and sysadmins think they're not competent for hosting services. It is a lot easier than you think, and its also fun to solve technical issues you may have.
The point was about redundancy / geo spread / HA. It’s significantly more difficult to operate two physical sites than one. You can only be in one place at a time.
If you want true reliability, you need redundant physical locations, power, networking. That’s extremely easy to achieve on cloud providers.
> Why do so many developers and sysadmins think they're not competent for hosting services. It is a lot easier than you think, and its also fun to solve technical issues you may have.
It is a different skillset. SRE is also an under-valued/paid (unless one is in FAANGO).
Maybe you find it fun. I don’t, I prefer building software not running and setting up servers.
It’s also nontrivial once you go past some level of complexity and volume. I have made my career at building software and part of that requires understanding the limitations and specifics of the underlying hardware but at the end of the day I simply want to provision and run a container, I don’t want to think about the security and networking setup it’s not worth my time.
Because when I’m running a busy site and I can’t figure out what went wrong, I freak out. I don’t know whether the problem will take 2 hours or 2 days to diagnose.
> Why do so many developers and sysadmins think they're not competent for hosting services.
Because those services solve the problem for them. It is the same thing with GitHub.
However, as predicted half a decade ago with GitHub becoming unreliable [0] and as price increases begin to happen, you can see that self-hosting begins to make more sense and you have complete control of the infrastructure and it has never been more easier to self host and bring control over costs.
> its also fun to solve technical issues you may have.
What you have just seen with coding agents is going to have the same effect on "developers" that will have a decline in skills the moment they become over-reliant on coding agents and won't be able to write a single line of code at all to fix a problem they don't fully understand.
At a previous job, the company had its critical IT infrastructure on their own data center. It was not in the IT industry, but the company was large and rich enough to justify two small data centers. It notably had batteries, diesel generators, 24/7 teams, and some advanced security (for valid reasons).
I agree that solving technical issues is very fun, and hosting services is usually easy, but having resilient infrastructure is costly and I simply don't like to be woken up at night to fix stuff while the company is bleeding money and customers.
> Maintaining one server room in the headquarters is something, but two servers rooms in different locations, with resilient power and network is a bit too much effort IMHO.
Speaking as someone who does this, it is very straightforward. You can rent space from people like Equinix or Global Switch for very reasonable prices. They then take care of power, cooling, cabling plant etc.
Yes, we still use the azure for user-facing services and the website. They don't need GPUs and don't need expensive resources, so it's not as worth it to bring those in-house.
We also rely on github. It has historically been good a service, but getting worth it.
I don't get why most everyone insists on comparing cloud to on-premises and not to dedicated. Why would anyone run own DC infra when there's Hetzner and many others?
Unfortunately we experienced an issue where our Slurm pool was contaminated by a misconstrained Postgres Daemon. Normally the contaminated slurm pool would drain into a docker container, but due to Rust it overloaded and the daemon ate its own head. Eventually we returned it to a restful state so all's well that ends well.
(hardware engineer trying to understand wtaf software people are saying when they speak)
This is cool. Yet, there are levels of insanity and those depend on your inability to estimate things.
When I'm launching a project it's easier for me to rent $250 worth of compute from AWS. When the project consumes $30k a month, it's easier for me to rent a colocation.
My point is that a good engineer should know how to calculate all the ups and downs here to propose a sound plan to the management. That's the winning thing.
It goes further than this first order, though. If you're trying to build a business that attracts the types of talent who wants to know the stack up and down, starting with an AWS instance might give you a better shot at funding (and thus a better overall shot), but it's not clear that it gives you a shot a building the business you're aiming for. For the things that "don't make your beer better", sure, but we're talking about training ML models at an ML shop. Here it makes sense for this reason.
That last part is exactly it and I while I know the intro sentence nails it I don’t think compute resonates with people (everyone uses compute). If you are 24/7 running work at scale it absolutely makes sense past the initial first couple years to build out your own DC like this.
Feels like I’ve lived through a full infrastructure fashion cycle already. I started my career when cloud was the obvious answer and on-prem was “legacy.”
Now on-prem is cool again.
Makes me wonder whether we’re already setting up the next cycle 10 years from now, when everyone rediscovers why cloud was attractive in the first place and starts saying “on-prem is a bad idea” again.
> Makes me wonder whether we’re already setting up the next cycle 10 years from now, when everyone rediscovers why cloud was attractive in the first place and starts saying “on-prem is a bad idea” again.
My entire career I’ve encountered people passionately pushing for on-prem and railing against anything cloud. I can’t remember a time when Hacker News comments leaned pro-cloud because it’s always been about self-hosting.
The few times the on-prem people won out in my career never went exactly as they imagined. Buying a couple servers and setting them up at the colo is easy enough, but the slow and steady drag of maintaining your own infrastructure starts to work its way into every development cycle after that. In my experience, every team has significantly underestimated how all the little things add up to a drag on available time for other work.
The best case for on-prem that I saw was when a company was basically in maintenance mode. Engineers had a lot of extra time to optimize, update. maintain, and cost reduce without subtracting from feature development or bug fixes.
The worst cases for on-prem I’ve seen have been funded startups. In this situation it’s imperative that everyone focus on feature development and rapid iteration. Letting some of the engineers get sidetracked with setting up and maintaining their own hosting to save a dollar amount that barely hires 1-2 more engineers but sets the schedule back by many months was a huge mistake.
In my experience, most engineers become less enchanted with rolling their own on premises hosting as they get older. Their work becomes more about getting the job done quickly and to budget, not hyper-optimizing the hosting situation at the expense of inviting more complexity and miscellaneous tasks into their workload.
If this were cyclical, I'd be inclined to agree, but this seems to be more of a wave. I also think the push back is more than just one against rented compute. It is tied to a societal ennui that comes from the feeling that we no longer own anything, be it music, housing, movies, land, tools, phones, or cars. Everything is moving to either being rented or on credit. There's a push back against this self-made feudal revival, and that scales all the way from individuals through to corporations; in this case, against the idea that a mega-corporation gets to decide how and when you get to use your compute, and at what variable price.
This is cyclical and I see the main axis of contention as centralized vs de-centralized computing.
Mainframes (network) gave way to mini and microcomputers (PCs). PCs gave way to server farms and web-based applications. Private servers and data centers gave way to the Cloud. Edge computing is again a push towards a more decentralized model.
Like all good engineering problems, where data and applications are hosted involve tradeoffs. Priorities change. Technologies change. But oftentimes, what works in one generation doesn't in the next. Part of it is the slow march of progress. But I think some of it is just not wanting to use your parent's technology stack and wanting to build your own.
The cloud vs. on-prem tradeoff is one of flexibility, capacity, maintenance, and capex vs opex.
It's a similar story in application development. At one point, we're navigating text forms on a mainframe, the next it's a GUI local application, followed by Electron or Web applications with remote data. We'll cycle back to local-first data (likely on-phone local models).
When you start to hear about the network being the computer again, you'll know we've started to swing back the other way again.
Sometimes, I feel like this is indicative of the incredible waste present in IT and development. Granted the cost of this kind of infrastructure upheaval is orders of magnitude cheaper than something like manufacturing - but still, it feels ridiculous that established companies can swap back and forth on a whim.
The problem was always the platform. For me, I saw very early on that kubernetes was exactly what I wanted after reading about how Google "treats the datacenter like one large computer." And I've been very happily running my own side projects on my own home cluster for 10 ish years (my kube-system namespace is 9y old). But selling any of my employers on this was a very hard proposition until enough people had shown it working at that scale.
On premises isn't only about saving money (that's not always clear). The article neglects the most important benefits which are freedom (control) and privacy. It's basically the same considerations that apply to owning vs renting a house.
The lowest grade I got in my business degree was in the "IT management" course. That's because the ONLY acceptable answer to any business IT problem is to move everything to the cloud. Renting is ALWAYS better than owning because you transfer cost and risk to a 3rd party.
That's pretty much the dogma of the 2010s.
It doesn't matter that my org runs a line-of-business datacentre that is a fraction of the cost of public cloud. It doesn't matter that my "big" ERP and admin servers take up half a rack in that datacentre. MBA dogma says that I need to fire every graybeard sysadmin, raze our datacentre facility to the ground, and move to AWS.
Fun fact, salaries and hardware purchases typically track inflation, because switching cost for hardware is nil and hiring isn't that expensive. Whereas software is usually 5-10% increases every year because they know that vendor lock-in and switching costs for software are expensive.
I was an on-prem maxi (if thats a thing) for a long time. I've run clusters that costed more than $5M, but these days I am a changed man. I start with PaaS like Vercel and work my way down to on-prem depending on how important and cost conscious that workload is.
Pains I faced running BIG clusters on-prem.
1. Supply chain Management -- everything from power supplies all the way to GPUs and storage has to be procured, shipped, disassembled and installed. You need labor pool and dedicated management.
2. Inventory Management -- You also need to manage inventory on hand for parts that WILL fail. You can expect 20% of your cluster to have some degree of issues on an ongoing basis
3. Networking and security -- You are on your own defending your network or have to pay a ton of money to vendors to come in and help you. Even with the simplest of storage clusters, we've had to deal with pretty sophisticated attacks.
When I ran massive clusters, I had a large team dealing with these. Obviously, with PaaS, you dont need anyone.
> I was an on-prem maxi (if thats a thing) for a long time. I've run clusters that costed more than $5M, but these days I am a changed man.
I have had a similar transformation. I still host non-critical services on-prem. They are exceptionally cheap to run. Everything else, I host it on Hetzner.
In addition to those sorts of non-first-hardware-purchase costs, the person writing the check needs to think long and hard about how bad an outage would be, and how much money it makes sense to budget simply to "avoiding outages." And the more important it is not to have any downtime, the more it's gonna cost to build up some sort of substitute for cross-datacenter cloud functionality. (You are also likely not going to be as good at either managing and configuring those networks, or hiring people to do so, as AWS, either.)
At scale (like comma.ai), it's probably cheaper. But until then it's a long term cost optimization with really high upfront capital expenditure and risk. Which means it doesn't make much sense for the majority of startup companies until they become late stage and their hosting cost actually becomes a big cost burden.
There are in between solutions. Renting bare metal instead of renting virtual machines can be quite nice. I've done that via Hetzner some years ago. You pay just about the same but you get a lot more performance for the same money. This is great if you actually need that performance.
People obsess about hardware but there's also the software side to consider. For smaller companies, operations/devops people are usually more expensive than the resources they manage. The cost to optimize is that cost. The hosting cost usually is a rounding error on the staffing cost. And on top of that the amount of responsibilities increases as soon as you own the hardware. You need to service it, monitor it, replace it when it fails, make sure those fans don't get jammed by dust puppies, deal with outages when they happen, etc. All the stuff that you pay cloud providers to do for you now becomes your problem. And it has a non zero cost.
The right mindset for hosting cost is to think of it in FTEs (full time employee cost for a year). If it's below 1 (most startups until they are well into scale up territory), you are doing great. Most of the optimizations you are going to get are going to cost you in actual FTEs spent doing that work. 1 FTE pays for quite a bit of hosting. Think 10K per month in AWS cost. A good ops person/developer is more expensive than that. My company runs at about 1K per month (GCP and misc managed services). It would be the wrong thing to optimize for us. It's not worth spending any amount of time on for me. I literally have more valuable things to do.
This flips when you start getting into the multiple FTEs per month in cost for just the hosting. At that point you probably have additional cost measured in 5-10 FTE in staffing anyway to babysit all of that. So now you can talk about trading off some hosting FTEs for modest amount of extra staffing FTEs and make net gains.
> At scale (like comma.ai), it's probably cheaper. But until then it's a long term cost optimization with really high upfront capital expenditure and risk. Which means it doesn't make much sense for the majority of startup companies until they become late stage and their hosting cost actually becomes a big cost burden.
You rent a dataspace, which is OPEX not CAPEX, and you just lease the servers, which turns big CAPEX into monthly OPEX bill
Running your own DC is "we have two dozen racks of servers" endeavour, but even just renting DC space and buying servers is much cheaper than getting same level of performance from the cloud.
> This flips when you start getting into the multiple FTEs per month in cost for just the hosting. At that point you probably have additional cost measured in 5-10 FTE in staffing anyway to babysit all of that. So now you can talk about trading off some hosting FTEs for modest amount of extra staffing FTEs and make net gains.
YOU NEED THOSE PEOPLE TO MANAGE CLOUD TOO. That's what always get ignore in calculations, people go "oh, but we really need like 2-3 ops people to cover datacenter and have shifts on the on-call", but you need same thing for cloud too, it is just dumped on programmers/devops guys in the team rather than having separate staff.
We have few racks and the part related to hardware is small part of total workload, most of it is same as we would (and do for few cloud customers) in cloud, writing manifests for automation.
Finally, some sense! "Cloud" was meant to make ops jobs disappear, but they just increased our salary by turning us into "DevOps Engineers" and the company's hosting bill increased fivefold in the process. You will never convince even 1% of devs to learn the ops side properly, therefore you'll still end up hiring ops people and we will cost you more now. On top of that, everyone that started as a "DevOps Engineer" knows less about ops than those that started as ops and transitioned into being "DevOps Engineers" (or some flavour of it like SREs or Platform Engineers).
If you're a programmer scared into thinking AI is going to take away your job, re-read my comment.
I'm not disagreeing... but it depends on how you shift the complexity/work and how you lean into or don't lean into the services a given cloud provider offer or not.
Just database management is a pretty specialized skill, separate from development or optimizing the structures of said data... For a lot of SaaS providers, if you aren't at a point where you can afford a dedicated DBA/Ops staff just for data, that's one reason you might lean into cloud operations or hybrid ops just for dbms management, security and backups. This is a low hanging fruit in terms of cloud offerings evem... but can shift a lot of burden in terms of operational overhead.
Again, depending on your business and data models.
To be fair, I think people are vastly over estimating the work they would have and the power they would need. Yes, if you have to massively scale up, then it'll take some work, but most of it is one-time work. You do it, and when it runs, you only have a fraction of work over the next months to maintain it. And with fraction, I mean below 5%. And keep in mind that >99% of startups who think of "yeah we need this and that cloud, because we need to scale" will never scale. Instead they are happily locking themselves into a cloud service. And if they actually scale at some point, this service will be massively more expensive.
We have two on site servers that we use. For various reasons (power cuts, internet outages, cleaners unplugging them) I’d say we have to intervene with them physically about once a month. It’s a total pain in the ass, especially when you don’t have _an_ it person sitting in the office to mind it. I’m in the Uk and our office is in Spain…
You should also calculate the cost of getting it up and running. With Google Cloud (I don't actually use AWS), I mainly worry about building docker containers in CI and deploying them to vms and triggering rolling restarts as those get replaced with new ones. I don't worry about booting them. I don't worry about provisioning operating systems or configuration to them. Or security updates. They come up with a lot of pre-provisioned monitoring and other stuff. No effort required on my side.
And for production setups. You need people on stand by to fix the server in case of hardware issues; also outside office hours. Also, where does the hardware live? What's your process when it fails? Who drives to wherever the thing is and fixes it? What do you pay them to be available for that? What's the lead time for spare components? Do you actually keep those in supply? Where? Do you pay for security for wherever all that happens? What about cleaning, AC, or a special server room in your building. All that stuff is cost. Some of it is upfront cost. Some of it is recurring cost.
The article is a about a company that owns its own data center. The cost they are citing (5 million) is substantial and probably a bit more complete. That's one end of the spectrum.
If you do it only a few hours every 6 months, you are not maintaining your infrastructure, you are letting it die (until the need arises and everything must be done and this is a massive project)
On the software side... depending on your business model, you can factor in a lot of the cost structures into your structure. Especially for say B2B arrangements.
Cloud integrations, for example, allow you to simply use a different database instance altogether per customer, while you can share services that utilize a given db connection. But actually setting up and managing that type of database infrastructure yourself may be much more resource intensive from a head count perspective.
I mention this, because having completely separate databases is an abstraction that cloud operations have already solved... while you can choose other options, such as more complex data models to otherwise isolate or share resources how does this complexity affect your services down-stream and the overall data complexities across one or all clients.
Harder still, if your data/service is centered around b2b clients of yours that have direct consumer interactions... then what if the industry is health or finance where there are even more legal concerns. Figuring a minimal (off the top) cost of each client of yours and scaling to the number of users under them isn't too hard to consider if you're using a mix of cloud services in concert with your own systems/services.
So yeah.. there's definitely considerations in either direction.
> it doesn't make much sense for the majority of startup companies until they become late stage
Here's what TFA says about this:
> Cloud companies generally make onboarding very easy, and offboarding very difficult. If you are not vigilant you will sleepwalk into a situation of high cloud costs and no way out.
and I think they're right. Be careful how you start because you may be stuck in the initial situation for a long time.
> But until then it's a long term cost optimization with really high upfront capital expenditure and risk.
The upfront capex does not need to be that high, unless you're running your own AI models. Other than leasing new ones, as a sibling comment stated, you can buy used. You can get a solid Dell 2U with a full service contract (3 years) for ~$5-10K depending on CPU / memory / storage configuration. Or if you don't mind going older - because honestly, most webapps aren't doing anything compute-heavy - you can drop that to < $1K/node. Replacement parts for those are cheap, so buy an extra of everything.
And if each of your clients is in the Healthcare industry and dealing with end-user medical data? Or financial data? Are you prepared for appropriate data isolation/sharding and controls? Do you have a strategy for scaling database operations per client or across all clients?
It really depends on the business model as to how well you might support your own infrastructure vs. relying on a new backend instance per client in a cloud infrastructure that has already solved many of the issues at play.
>At scale (like comma.ai), it's probably cheaper. But until then it's a long term cost optimization with really high upfront capital expenditure and risk.
The issue with comma.ai is that the company is HEAVILY burdened with Geohotz ideals, despite him no longer even being on the board. I used to be very much into his streams and he rants about it plenty. A large reason of why they run their own datacenter is that they ideologically refuse to give money to AWS or Google (but I guess Microsoft passes their non woke test).
Which is quite hilarious to me because they live in a very "woke" state and complain about power costs in the blog post. They could easily move to Wyoming or Montana and with low humidity and colder air in the winter run their servers more optimally.
Our preference for training in our own datacenter has nothing to do with wokeness. Did you read the blog post? The reasons are clearly explained.
The climate in Wyoming and Montana are actually worse in terms of climate. San Diego's climate extremes are less extreme than those places. Though moving out of CA is a good idea for power cost reasons, also addressed in the blog.
The reason companies don’t go with on premises even if cloud is way more expensive is because of the risk involved in on premises.
You can see it quite clearly here that there’s so many steps to take. Now a good company would concentrate risk on their differentiating factor or the specific part they have competitive advantage in.
It’s never about “is the expected cost in on premises less than cloud”, it’s about the risk adjusted costs.
Once you’ve spread risk not only on your main product but also on your infrastructure, it becomes hard.
I would be vary of a smallish company building their own Jira in house in a similar way.
Software companies have higher margins so these decisions are lower stakes. Unless on premises helps the bottom line of the main product that the company provides, these decisions don't really matter in my opinion.
Think of a ~5000 employee startup. Two scenarios:
1. if they win the market, they capture something like ~60% margin
2. if that doesn't happen, they just lose, VC fund runs out and then they leave
In this dynamic, costs associated with infrastructure don't change the bottomline of profitability. The risk involved with rolling out their on infrastructure can hurt their main product's existence itself.
Yes, the idea is that you focus on the things that differentiate you from the competition. If you’re a factory that makes nails, a better data centre won’t make you any more money. It won’t help you sell more nails. So you should leave the data centres to the experts, and focus on work which improves your actual product.
If you don’t, you’ll be stuck trying to figure out data centres. Hiring tons of infrastructure experts, trying to manage power consumption. And for what? You won’t sell any more nails.
If you’re a company like Google, having better data centres does relate to your products, so it makes sense to focus on them and build your own.
I think it wins because opex is seen as stable recurring cost and capex is seen as the money you put in your primary differentiation for long term gains.
Note that they're running R630/R730s for storage. Those are 12-year old servers, and yet they say each one can do 20 Gbps (2.5 GBps) of random reads. In comparison, the same generation of hardware at AWS ({c,m,r}4) instance maxes out at 50% of that for EBS throughput on m4, and 70% on r4 - and that assumes carefully tuned block sizes.
Old hardware is _plenty_ powerful for a lot of tasks today.
I’m on a project at work replacing our R430s and R730s. They’ve been absolute tanks with very few hardware failures. That said, my company chooses to have OEM support for replacing failed components and keeping firmware/bios/idrac updated. You can absolutely run these if you’re OK with 3rd party replacements or parting out spare machines. Some industries are more tolerant to this than others.
I ran 3x R620s 24/7/365 in my homelab for ~6 years (well, other than when I moved, or shut one down for a clean-and-inspect, or lost power in excess of what my UPS could handle... thanks, Texas). The only things that failed during that time were a couple of sticks of RAM, and a PSU.
The own vs rent calculus for compute is starting to mirror the market value vs replacement cost divergence we see in physical assets.
Cloud is convenient because it lowers OpEx initially, but you lose control over the long-term CapEx efficiency. Once you reach a certain scale, paying the premium for AWS flexibility stops making sense compared to the raw horsepower of owned metal.
Using "big" cloud providers is often a mistake. You want to use rented assets to bootstrap and then start deploying on instances that are more and more under your control. With big cloud providers, it is easy to just succumb to their service offerings rather than do the right thing. Do your PoC on Hetzner and DigitalOcean then scale with purpose.
No, low isn't good perse. I worked in a datacenter which in winters had less than 40%, ram was failing all over the place. Low humidity causes static electricity.
Low is good if you are also adding more humidity back in. If you want to maintain 45-50% (guessing), then you would want <45% environmental humidity so that you can raise it to the level you want. You're right about avoiding static, but you'd still want to try to keep it somewhat consistent.
It is much cheaper to use external air for cooling if you can.
Yeah but the article makes it sound as if lower is better, which it is definitely not. And yeah you need to control humidity, that might mean sometimes lowering, and sometimes increase it by whatever solution you have.
Also this is where cutting corner indeed results in lower cost, which was the reason for the OP to begin with. It just means you won't get as good a datacenter as people who are actually tuning this whole day and have decades of experience.
The distinction between rent/own is kind of a false dichotomy. You never truly own your platform - you just "rent" it in a more distributed way that shields you from a single stress point. The tradeoff is that you have to manage more resources to take care of it, but you have much greater flexibility.
I have a feeling AI is going to be similar in the future. Sure, you can "rent" access to LLM's and have agents doing all your code. And in the future, it'll likely be as good as most engineers today. But the tradeoff is that you are effectively renting your labor from a single source instead of having a distributed workforce. I don't know what the long-term ramifications are here, if any, but I thought it was an interesting parallel.
I think this is how IBM is making tons of money on mainframes. A lot of what people are doing with cloud can be done on premises with the right levels of virtualization.
For ML it makes sense, because you’re using so much compute that renting it is just burning money.
For most businesses, it’s a false economy. Hardware is cheap, but having proper redundancy and multiple sites isn’t. Having a 24/7 team available to respond to issues isn’t.
What happens if their data centre loses power? What if it burns down?
I fully lost three small VPS there, and their response was poor: they didn't even refund time lost, they didn't compensate for time lost (e.g. a couple of months of free VPS), I got better updates from the news than from them (news were saying "almost total loss", while them were trying to convince me that I had the incredible bad luck that my three VPS were in the very small zone affected by the fire). The only way I had to recover what I lost was backups in local machines.
When someone point out how safe are cloud providers, as if they have multiple levels of redundancy and are fully protected against even an alien invasion, I remember the OVH fire.
They use the datasenter for model training, not to serve online users. Presumably even if it will be offline for a week or even a month it will not be a total disaster as long as they have, for example, offsite tape backups.
Flooding due to burst frozen pipe, false sprinkler trigger, or many others.
Something very similar happened at work. Water valve monitoring wasn’t up yet. Fire didn’t respond because reasons. Huge amount of water flooded over a 3 day weekend. Total loss.
why build one when you can have two at twice the price?
But, if you're building a datacenter for $5M, spending $10-15M for redundant datacenters (even with extra networking costs), would still be cheaper than their estimated $25M cloud costs.
Ah Slurm, so good to see it still being used. As soon as I touched it in ~2010 I realized this was finally the solid queue management system we needed. Things like Sun Grid Engine or PBS were always such awful and burdensome PoS.
IIRC, Slurm came out of LLNL, and it finally made both usage and management of a cluster of nodes really easy and fun.
Compare Slurm to something like AWS Batch or Google Batch and just laugh at what the cloud has created...
Not nearly on the article's level, but I've been operating what I call a fog machine (itsy bitsy personal cloud) for about 15 years. It's just a bunch of local and off-site NAS boxes. It has kinda worked out great. Mostly Synology, but probably won't be when their scheduled retirement comes up. The networking is dead simple, the power use is distributed, and the size of it all is still a monster for me - back in the day, I had to use it for a very large audio project to keep backups of something like 750,000 albums and other audio recordings along with their metadata and assets.
The cloud requires expertise in company-specific APIs and billing systems. A data center requires knowledge of Watts, bits, and FLOPs. I know which one I rather think about.
Re: the "hobby" part is where I agree with you the most. Where you say it's not solving genuine problems is where I differ the most.
It really feels to me like Comma is staffed by people who recognize that they never stopped enjoying playing with Lego -- their bricks just grew up, and they realized they can:
1) solve real-world problems
2) not be jerks about it
3) get paid to do it
Not everything has to be about optimizing for #3.
I'm a happy paying customer of Comma.ai (Comma four, baby!) -- their product is awesome, extremely consumer-friendly, and I hope they can grow in their success!
To me it sounds more like a return to vertical integration.
This is becoming increasingly common as far as I can tell.
There are benefits either direction, and I think that each company needs to evaluate the pros and cons themselves. Emotional pros/cons are something companies need to evaluate as employee morale can make or break a company. If the company is super technical in culture and they gain something intangible that is boosting the bottom line, having a datacenter as a "cool" factor is probably worth it.
> Self-reliance is great, but there are other benefits to running your own compute. It inspires good engineering.
It's easy to inspire people when you have great engineers in the first place. That's a given at a place like comma.ai, but there are many companies out there where administering a datacenter is far beyond their core competencies.
I feel like skilled engineers have a hard time understanding the trade-offs from cloud companies. The same way that comma.ai employees likely don't have an in-house canteen, it can make sense to focus on what you are good at and outsource the rest.
> I feel like skilled engineers have a hard time understanding the trade-offs from cloud companies.
They spend too much time on yet another cloud native support group call, learning for ThatOneCloudProvider certificates, figuring out that single implementation caveats, standardizing security procedures between cloud teams, and so on.
Yet people in the article just throw a 1000 lines of code KV store mkv [0] on a huge raw storage server and call it a day. And it's a legit choice, they did actual study beforehand and concluded: we don't need redundancy in most cases. At all. I respect that.
You can also buy the hardware and hire an IT vendor to rack and help manage it as smart hands so you never need to visit the datacenter. With modern beefy hardware, even large web services only need a few racks so most orgs don't even to manage a large footprint.
Sure you have to schedule your own hardware repairs or updates but it also means you don't need to wrangle with the ridiculous cost-engineering, reserved instances, cloud product support issues or API deprecations, proprietary configuration languages, etc.
Bare metal is better for a lot of non-cost reasons too, as the article notes it's just easier/better to reason about the lower level primitives and you get more reliable and repeatable performance.
That’s called managed servers or managed services.
I have run bare metal and manage services you just have to be clear on what you have coverage for when disaster strikes or be willing to proactively replace hard drives before they die.
If your business relies on compute, and you run that compute in the cloud, you are putting a lot of trust in your cloud provider. Cloud companies generally make onboarding very easy, and offboarding very difficult. If you are not vigilant you will sleepwalk into a situation of high cloud costs and no way out. If you want to control your own destiny, you must run your own compute.
This is not a valid reason for running your own datacenter, or running your own server.
Self-reliance is great, but there are other benefits to running your own compute. It inspires good engineering. Maintaining a data center is much more about solving real-world challenges. The cloud requires expertise in company-specific APIs and billing systems. A data center requires knowledge of Watts, bits, and FLOPs. I know which one I rather think about.
This is not a valid reason for running your own datacenter, or running your own server.
Avoiding the cloud for ML also creates better incentives for engineers. Engineers generally want to improve things. In ML many problems go away by just using more compute. In the cloud that means improvements are just a budget increase away. This locks you into inefficient and expensive solutions. Instead, when all you have available is your current compute, the quickest improvements are usually speeding up your code, or fixing fundamental issues.
This is not a valid reason for owning a datacenter, or running your own server.
Finally there’s cost, owning a data center can be far cheaper than renting in the cloud. Especially if your compute or storage needs are fairly consistent, which tends to be true if you are in the business of training or running models. In comma’s case I estimate we’ve spent ~5M on our data center, and we would have spent 25M+ had we done the same things in the cloud.
This is one of only two valid reasons for owning a datacenter, and one of several valid reasons for running your own server.
The only two valid reasons to build/operate a datacenter: 1) what you're doing is so costly that building your own factory is the only profitable way for your business to produce its widgets, 2) you can't find a datacenter with the location or capacity you need and there is no other way to serve your business needs.
There's many valid reasons to run your own servers (colo), although most people will not run into them in a business setting.
This also depends so much on your scaling needs. If you need 3 mid-sized ECS/EC2 instances, a load balancer, and a database with backups, renting those from AWS isn’t going to be significantly more expensive for a decent-sized company than hiring someone to manage a cluster for you and dealing with all the overhead of keeping it maintained and secure.
If you’re at the scale of hundreds of instances, that math changes significantly.
And a lot of it depends on what type of business you have and what percent of your budget hosting accounts for.
I also thinks it’s risk model too. Every time I see these kind of posts I think it misses the point there is a balance not only on cost like you describe but risk as well. You are paying to offload some of the risk from yourself.
The issue is that they have already paid off their datacenter 5x over compared to cloud. For offline, batch training, I don't ses how any amount of risk could offset the savings.
> The cloud requires expertise in company-specific APIs and billing systems.
This is one reason I hate dealing with AWS. It feels like a waste of time in some ways. Like learning a fly-by-night javascript library - maybe I'm better off spending that time writing the functionality on my own, to increase my knowledge and familiarity?
Naive comment from a hobbyist with nothing close to $5M: I'm curious about the degree to which you build a "home lab" equivalent. I mean if "scaling" turned out to be just adding another Raspberry Pi to the rack (where is Mr. Geerling when you need him?) I could grow my mini-cloud month by month as spending money allowed.
The degree is whatever you want to deal with. I had a rack at my last house (need to redesign the space for it at new house) with 3x Dell R620s in a Proxmox cluster, running K8s, serving Ceph from NVMe drives over Infiniband (for the mesh traffic), and 2x Supermicros running independent ZFS pools.
It was fun to build - especially Infiniband - but my next iteration is going to be a single beefy server, maybe with storage attached externally. What I had had outstanding uptime, but ultimately it was massively overkill, noisy, hot, and sucked power down.
I paid 150€ for a Mini PC with an Intel N100, 16 GB of DDR5 memory, and a 500 GB SSD.
While I have no intention to scale up low spec hardware like this, it at least seems to beat the Azure VMs we use at work with "4 CPUs", which corresponds to two physical cores on an AMD EPYC CPU.
And that super slow machine I understand costs more than $100 per month, and that's without charges for disk space slower than the SSD, or network traffic.
Renting at Azure seems to be a terrible decision, particularly for desktop use.
Working at a non-tech regional bigco, where ofc cloud is the default, I see everyday how AWS costs get out of hand, it's a constant struggle just to keep costs flat. In our case, the reality is that NONE of our services require scalability, and the main upside of high uptime is nice primarily for my blood pressure.. we only really need uptime during business hours, nobody cares what happens at night when everybody is sleeping.
On the other hand, there's significant vendor lockin, complexity, etc. And I'm not really sure we actually end up with less people over time, headcount always expands over time, and there's always cool new projects like monitoring, observability, AI, etc.
My feeling is, if we rented 20-30 chunky machines and ran Linux on them, with k8s, we'd be 80% there. For specific things I'd still use AWS, like infinite S3 storage, or RDS instances for super-important data.
If I were to do a startup, I would almost certainly not base it off AWS (or other cloud), I'd do what I write above: run chunky servers on OVH (initially just 1-2), and use specific AWS services like S3 and RDS.
A bit unrelated to the above, but I'd also try to keep away from expensive SaaS like Jira, Slack, etc. I'd use the best self-hosted open source version, and be done with it. I'd try Gitea for git hosting, Mattermost for team chat, etc.
And actually, given the geo-political situation as an EU citizen, maybe I wouldn't even put my data on AWS at all and self-host that as well...
Same thing. I was previously spending 5-8K on DigitalOcean, supposedly a "budget" cloud. Then the company was sold, and I started a new company on entirely self-hosted hardware. Cloudflare tunnel + CC + microk8s made it trivial! And I spend close to nothing other than internet that I already am spending on. I do have solar power too.
> Cloud companies generally make onboarding very easy, and offboarding very difficult. If you are not vigilant you will sleepwalk into a situation of high cloud costs and no way out. If you want to control your own destiny, you must run your own compute.
Cost and lock-in are obvious factors, but "sovereignty" has also become a key factor in the sales cycle, at least in Europe.
Handing health data, Juvoly is happy to run AI work loads on premise.
On top of that, now when the US cloud act is again a weapon against EU, most European companies know better and are migrating in droves to colo, on-prem and EU clouds. Bye bye US hyperscalers!
The #1 reason I would advocate for using AWS today is the compliance package they bring to the party. No other cloud provider has anything remotely like Artifact. I can pull Amazon's PCI-DSS compliance documentation using an API call. If you have a heavily regulated business (or work with customers who do), AWS is hard to beat.
If you don't have any kind of serious compliance requirement, using Amazon is probably not ideal. I would say that Azure AD is ok too if you have to do Microsoft stuff, but I'd never host an actual VM on that cloud.
Compliance and "Microsoft stuff" covers a lot of real world businesses. Going on prem should only be done if it's actually going to make your life easier. If you have to replicate all of Azure AD or Route53, it might be better to just use the cloud offerings.
I used to colocate a 2U server that I purchased with a local data center. It was a great learning experience for me. Im curious why a company wouldn't colocate their own hardware? Proximity isnt an issue when you can have the datacenter perform physical tasks. Bravo to the comma team regardless. It'll be a great learning experience and make each person on their team better.
Ps... bx cable instead of conduit for electrical looks cringe.
The main reason not to colocate is if you're somewhere with high real estate costs... E.g
Hetzner managed servers competes on price w/co-location for me because I'm in London.
I colocate in London, a single server / firewall comes to around £5k a year. I also colocate two other servers in some northern UK location in some industrial estate for £2k as my backups. I've never enjoyed the cloud and dedicated server's have their own caveats too.
Budget hosts such as Hetzner/OVH have been known to suddenly pull the plug for no reason.
My kit is old, second hand old (Cisco UCS 220 M5, 2xDell somethings) and last night I just discovered I can throw in two NVIDIA T4's and turn it in to a personal LLM.
I'm quite excited having my own colocated server with basic LLM abilities. My own hardware with my own data and my own cables. Just need my own IP's now.
This is hackernews, do the math for the love of god.
There are good business and technical reasons to choose a public cloud.
There are good business and technical reasons to choose a private cloud.
There are good business and technical reasons to do something in-between or hybrid.
The endless "public cloud is a ripoff" or "private clouds are impossible" is just a circular discussion past each other. Saying to only use one or another is textbook cargo-culting.
The company I work for used to have a hybrid where 95% was on-prem, but became closer to 90% in the cloud when it became more expensive to do on-prem because of VMware licensing. There are alternatives to VMware, but not officially supported with our hardware configuration, so the switch requires changing all the hardware, which still drives it higher than the cloud. Almost everything we have is cloud agnostic, and for anything that requires resilience, it sits in two different providers.
Now the company is looking at doing further cost savings as the buildings rented for running on-prem are sitting mostly unused, but also the prices of buildings have gone up in recent years, notably too, so we're likely to be saving money moving into the cloud. This is likely to make the cloud transition permanent.
> Cloud companies generally make onboarding very easy, and offboarding very difficult.
I reckon most on-prem deployments have significantly worse offboarding than the cloud providers. As a cloud provider you can win business by having something for offboarding, but internally you'd never get buy-in to spend on a backup plan if you decide to move to the cloud.
> As a cloud provider you can win business by having something for offboarding, but internally you'd never get buy-in to spend on a backup plan if you decide to move to the cloud.
Its the other way around. How do you think all businesses moved to the cloud in the first place?
15-years ago or so a spreadsheet was floating around where you could enter server costs, compute power, etc and it would tell you when you would break-even by buying instead of going with AWS. I think it was leaked from Amazon because it was always three-years to break-even even as hardware changed over time.
Azure provides their own "Total Cost of Ownership" calculator for this purpose [0]. Notably, this makes you estimate peripheral costs such as cost of having a server administrator, electricity, etc.
If you buy, maybe. Leasing or renting tends to be cheaper from day one. Tack on migration costs and ca. 6 months is a more realistic target. If the spreadsheet always said 3 years, it sounds like an intentional "leak".
Well, somebody should recreate it. I smell a potential startup idea somewhere. There's a ton of "cloud cost optimizers" software but most involve tweaking AWS knobs and taking a cut of the savings. A startup that could offload non critical service from AWS to colo and traditional bare metal hosting like Hetzner has a strong future.
One thing to keep in mind is that the curve for GPU depreciation (in the last 5 years at least) is a little steeper than 3 years. Current estimates is that the capital depreciation cost would plunge dramatically around the third year. For a top tier H100 depreciation kicks in around the 3rd year but they mentioned for the less capable ones like the A100 the depreciation is even worse.
Now this is not factoring cost of labour. Labor at SF wages is dreadfully expensive, now if your data center is right across the border in Tijuana on the other hand..
> Maintaining a data center is much more about solving real-world challenges. The cloud requires expertise in company-specific APIs and billing systems. A data center requires knowledge of Watts, bits, and FLOPs. I know which one I rather think about.
I find this to be applicable on a smaller scale too! I'd rather setup and debug a beefy Linux VPS via SSH than fiddle with various propietary cloud APIs/interfaces. Doesn't go as low-level as Watts, bits and FLOPs but I still consider knowledge about Linux more valuable than knowing which Azure knobs to turn.
I'm thinking about doing a research project at my university looking into distributed "data centers" hosted by communities instead of centralized cloud providers.
The trick is in how to create mostly self-maintaining deployable/swappable data centers at low cost...
Realistically, it's the speed with which you can expand and contract. The cloud gives unbounded flexibility - not on the per-request scale or whatever, but on the per-project scale. To try things out with a bunch of EC2s or GCEs is cheap. You have it for a while and then you let it go. I say this as someone with terabytes of RAM in servers, and a cabinet I have in the Bay Area.
Cloud, in terms of "other company's infrastructure" always implies losing the competence to select, source and operate hardware. Treating hardware as commodity will eventually treat your very own business as commodity: Someone can just copy your software/IP and ruin your business. Every durable business needs some kind of intellectual property and human skills that are not replaceable easily. This sounds binary, but isn't. You can build long-lasting partnerships. German Mittelstand did that over decades.
I just read about Railway doing something similar, sadly their prices are still high compared to other bare metal providers and even VPS such as Hetzner with Dokploy, very similar feature set yet for the same 5 dollars you get way more CPU, storage and RAM.
Billing per used or not idle cpu cycle would be quite interesting. Number of cores would just effectively be your cost cap. Efficiency would be even more important. And if the provider over subscribes cores you just pay less. Actually that's probably why they don't do it...
Heavy ML workloads make this more worthwhile since you get to design it to squeeze value out of every facet. For a basic web server and database it’s definitely overkill and something like a colocation makes much more sense
Electricity cost in California is generally more expensive than most other US states, except Hawaii. Not sure why.
Perhaps Comma needed the datacenter to be in San Diego for latency or other reasons, but if they need it mostly for compute, it would have been cheaper to operate their datacenter elsewhere... but if we keep going down that path, maybe it actually becomes cheaper to rent a cloud after all.
The cloud is a psyop, a scam. Except at the tiniest free-tier / near free-tier use cases, or true scale to zero setups.
I've helped a startup with 2.5M revenue reduce their cloud spend from close to 2M/yr to below 1M/yr. They could have reached 250k/yr renting bare-metal servers. Probably 100k/yr in colos by spending 250k once on hardware. They had the staff to do it but the CEO was too scared.
Cloud evangelism (is it advocacy now?) messed up the minds of swaths of software engineers. Suddenly costs didn't matter and scaling was the answer to poor designs. Sizing your resource requirements became a lost art, and getting into reaction mode became law.
Welcome to "move fast and get out of business", all enabled by cloud architecture blogs that recommend tight integration with vendor lock-in mechanisms.
Use the cloud to move fast, but stick to cloud-agnostic tooling so that it doesn't suck you in forever.
I've seen how much cloud vendors are willing to spend to get business. That's when you realize just how massive their margins are.
> Suddenly costs didn't matter and scaling was the answer to poor designs.
It did.
Did you know that cloud cost less than what the internal IT team at a company would charge you?
Let's say you worked on product A for a company and needed additional VM. Besides paperwork, the cost to you (for your cost center) would be more than using the company credit card for the cloud.
> Sizing your resource requirements became a lost art
In what way? We used to size for 2-4x since getting additional resources (for the in-house team) would be weeks to months. Same old - just cloud edition.
> Did you know that cloud cost less than what the internal IT team at a company would charge you?
Yes. Internal IT teams ran old-school are inefficient. And that's what the vendor tells you while they create shadow IT inside your company. Skip ITSM and ITIL... do it the SRE way.
Until the cloud economist (real role) comes in and finds a way to extract more rent out of their customer base (like GCP's upcoming doubling rates on CDN Interconnect). And until internal IT kills shadow IT and regains management of cloud deployments. Cybersecurity and stuff...
Back to square one. ITIL with cloud deployments. Some use cases will be way cheaper... but for your 100s of PBs of enterprise data, that's another story. And data gravity will kill many initiatives just based on bit movement costs.
> Besides paperwork, the cost to you (for your cost center) would be more than using the company credit card for the cloud.
To some extent. One is hard dollars the other is funny money. But I thought paying for cloud with the company credit card was a 2016 thing. Now it's paid through your internal IT cost center, with internal IT markup.
I've seen petabytes of data move to the cloud and then we couldn't perform some queries on it anymore as that store wouldn't support it, and we'd need to spend 7 figures to move to another cloud database to query it. And that's hard dollars.
Yes, during early cloud days it was lean and aimed at startups. Now it's aimed at enterprise, and for some reason lots of startups still think it's optimized for them. It's not and it hasn't been for a long time.
Even at the personal blog level, I'd argue it's worth it to run your own server (even if it's just an old PC in a closet). Gets you on the path to running a home lab.
Absolutely. I don't have a blog but run my own email, several game servers, Matrix instance, Nextcloud and other internal services on a retired gaming PC. The total cost of my cloud subscriptions is $0, and no one is snooping on me. It's a great setup when combined with Linux machines and GrapheneOS phones, completely private and free of Big Tech.
I cancelled my digital ocean server of almost a decade late last year and replaced it with a raspberry pi 3 that was doing nothing. We can do it, we should do it.
Microsoft made the TCO argument and won. Self-hosting is only an option if you can afford expensive SysOps/DevOps/WhateverWeAreCalledTheseDays to manage it.
IT dinosaur here, who has run and engineered the entire spectrum over the course of my career.
Everything is a trade-off. Every tool has its purpose. There is no "right way" to build your infrastructure, only a right way for you.
In my subjective experience, the trade-offs are generally along these lines:
* Platform as a Service (Vercel, AWS Lambda, Azure Functions, basically anything where you give it your code and it "just works"): great for startups, orgs with minimal talent, and those with deep pockets for inevitable overruns. Maximum convenience means maximum cost. Excellent for weird customer one-offs you can bill for (and slap a 50% margin on top). Trade-off is that everything is abstracted away, making troubleshooting underlying infrastructure issues nigh impossible; also that people forget these things exist until the customer has long since stopped paying for them or a nasty bill arrives.
* Infrastructure as a Service (AWS, GCP, Azure, Vultr, etc; commonly called the "Public Cloud"): great for orgs with modest technical talent but limited budgets or infrastructure that's highly variable (scales up and down frequently). Also excellent for everything customer-facing, like load balancers, frontends, websites, you name it. If you can invoice someone else for it, putting it in here makes a lot of sense. Trade-off is that this isn't yours, it'll never be yours, you'll be renting it forever from someone else who charges you a pretty penny and can cut you off or raise prices anytime they like.
* Managed Service/Hosting Providers (e.g., ye olde Rackspace): you don't own the hardware, but you're also not paying the premium for infrastructure orchestrators. As close to bare metal as you can get without paying for actual servers. Excellent for short-term "testing" of PoCs before committing CapEx, or for modest infrastructure needs that aren't likely to change substantially enough to warrant a shift either on-prem or off to the cloud. You'll need more talent though, and you're ultimately still renting the illusion of sovereignty from someone else in perpetuity.
* Bare Metal, be it colocation or on-premises: you own it, you decide what to do with it, and nobody can stop you. The flip side is you have to bootstrap everything yourself, which can be a PITA depending on what you actually want - or what your stakeholders demand you offer. Running VMs? Easy-peasy. Bare metal K8s clusters? I mean, it can be done, but I'd personally rather chew glass than go without a managed control plane somewhere. CapEx is insane right now (thanks, AI!), but TCO is still measured in two to three years before you're saving more than you'd have spent on comparable infrastructure elsewhere, even with savings plans. Talent needs are highly variable - a generalist or two can get you 80% to basic AWS functionality with something like Nutanix or VCF (even with fancy stuff like DBaaS), but anything cutting edge is going to need more headcount than a comparable IaaS build. God help you if you opt for a Microsoft stack, as any on-prem savings are likely to evaporate at your next True-Up.
In my experience, companies have bought into the public cloud/IaaS because they thought it'd save them money versus the talent needed for on-prem; to be fair, back when every enterprise absolutely needed a network team and a DB team and a systems team and a datacenter team, this was technically correct. Nowadays, most organizational needs can be handled with a modest team of generalists or a highly competent generalist and one or two specialists for specific needs (e.g., a K8s engineer and a network engineer); modern software and operating systems make managing even huge orgs a comparable breeze, especially if you're running containers or appliances instead of bespoke VMs.
As more orgs like Comma or Basecamp look critically at their infrastructure needs versus their spend, or they seriously reflect on the limited sovereignty they have by outsourcing everything to US Tech companies, I expect workloads and infrastructure to become substantially more diversified than the current AWS/GCP/Azure trifecta.
> In comma’s case I estimate we’ve spent ~5M on our data center, and we would have spent 25M+ had we done the same things in the cloud.
IMO, that's the biggie. It's enough to justify paying someone to run their datacenter. I wish there was a bit more detail to justify those assumptions, though.
That being said, if their needs grow by orders of magnitude, I'd anticipate that they would want to move their servers somewhere with cheaper electricity.
What redundancy are we talking about? AWS has proven to the world on multiple occasions that redundancy across geo locations is useless, because if us-east-1 is down, their whole cloud is done, causing a big chunk of the world to be down.
Half sarcasm of course, but it goes to show that the world is not going to fall apart in many cases when it comes to software. Sure, it's not ideal in lots of cases, but we'll survive without redundancy.
This is a great solution for a very specific type of team but I think most companies with consistent GPU workloads will still just rent dedicated servers and call it a day.
I agree, and cloud compute is poised to become even more commoditized in the coming years (gazillion new data centers + AI plateauing + efficiency gains, the writing is on the wall). There’s no way this makes sense for most companies.
The advantage of renting vs. owning is that you can always get the latest gen, and that brings you newer capabilities (i.e. fp8, fp4, etc) and cheaper prices for current_gen-1. But betting on something plateauing when all the signs point towards the exact opposite is not one of the bets i'd make.
Other benefits: easy access to reliable infrastructure and latest hardware which you can swap as you please. There are cases where it makes sense to navigate away from the big players (like dropbox going from aws to on-prem), but again you make this move when you want to optimize costs and are not worried about the trade-offs.
Not long ago Railway moved from GCP to their own infrastructure since it was very expensive for them. [0] Some go for a Oxide rack [1] for a full stack solution (both hardware and software) for intense GPU workloads, instead of building it themselves.
It's very expensive and only makes sense if you really need infrastructure sovereignty. It makes more sense if you're profitable in the tens of millions after raising hundreds of millions.
It also makes sense for governments (including those in the EU) which should think about this and have the compute in house and disconnected from the internet if they are serious about infrastructure sovereignty, rather than depending on US-based providers such as AWS.
The observation about incentives is underappreciated here. When your compute is fixed, engineers optimize code. When compute is a budget line, engineers optimize slide decks. That's not really a cloud vs on-prem argument, it's a psychology-of-engineering argument.
One thing I don't really understand here is why they're incurring the costs of having this physically in San Diego, rather than further afield with a full-time server tech essentially living on-prem, especially if their power numbers are correct. Is everyone being able to physically show up on site immediately that much better than a 24/7 pair of remote hands + occasional trips for more team members if needed?
Clouds suck. But so does “on premises”. Or co-location.
In the future, what you will need to remain competitive is computing at the edge. Only one company is truly poised to deliver on that at massive scale.
I like Hotz’s style: simply and straightforwardly attempting the difficult and complex. I always get the impression: “You don’t need to be too fancy or clever. You don’t need permission or credentials. You just need to go out and do the thing. What are you waiting for?”
And finally we reach the point where you're not shot for explaining if you invest in ownership after everything is over you have something left that has intrinsic value regardless of what you were doing with it.
Otherwise, well just like that gym membership, you get out what you put into it...
Am I the only one that is simply scared of running your own cloud? What happens if your administrator credentials get leaked? At least with Azure I can phone microsoft and initiate a recovery. Because of backups and soft deletion policies quite a lot is possible. I guess you can build in these failsafe scenarios locally too? But what if a fire happens like in South Korea? Sure most companies run more immediate risks such as going bankrupt, but at least Cloud relieves me from the stuff of nightmares.
Except now I have nightmares that the USA will enforce the patriot act and force Microsoft to hand over all their data in European data centers and then we have to migrate everything to a local cloud provider. Argh...
Do you have a computer at home? Are you scared of its credentials leaking? A server is just another computer with a good internet connection.
You can equip your server with a mouse, keyboard and screen and then it doesn't even need credentials. The credential is your physical access to the mouse and keyboard.
I mean people are nowadays are really scared of using microwave oven too. What happens if I heat my coffee 1 min too long. Could be near death experience. Thats why I always drive down to Starbucks for coffee!
Having worked only with the cloud I really wonder if these companies don't use other software with subscriptions. Even though AWS is "expensive" its a just another line item compared to most companies overall SaaS spend. Most businesses don't need that much compute or data transfer in the grand scheme of things.
Or better; write your software such that you can scale to tens of thousands of concurrent users on a single machine. This can really put the savings into perspective.
Well the article starts out with a suggestion that we should all get a data center... It's quite a jump to assume that everyone reading this article needs to train their own LLMs.
the “build your own datacenter” story is fun (and comma’s setup is undeniably cool), but for most companies it’s a seductive trap: you’ll spend your rarest resource (engineer attention) on watts, humidity, failed disks, supply chains, and “why is this rack hot,” instead of on the product. comma can justify it because their workload is huge and steady, they’re willing to run non-redundant storage, and they’ve built custom GPU boxes and infra around a very specific ML pipeline. ([comma.ai blog][1])
## 1) capex is a tax on flexibility
a datacenter turns “compute” into a big up-front bet: hardware choices, networking choices, facility choices, and a depreciation schedule that does not care about your roadmap. cloud flips that: you pay for what you use, you can experiment cheaply, and you can stop spending the minute a strategy changes. the best feature of renting is that quitting is easy.
## 2) scaling isn’t a vibe, it’s a deadline
real businesses don’t scale smoothly. they spike. they get surprise customers. they do one insane training run. they run a migration. owning means you either overbuild “just in case” (idle metal), or you underbuild and miss the moment. renting means you can burst, use spot/preemptible for the ugly parts, and keep steady stuff on reserved/committed discounts.
## 3) reliability is more than “it’s up most days”
comma explicitly says they keep things simple and don’t need redundancy for ~99% uptime at their scale. ([comma.ai blog][1]) that’s a perfectly valid trade—if your business can tolerate it. many can’t. cloud providers sell multi-zone, multi-region, managed backups, managed databases, and boring compliance checklists because “five nines” isn’t achieved by a couple heroic engineers and a PID loop.
## 4) the hidden cost isn’t power, it’s people
comma spent ~$540k on power in 2025 and runs up to ~450kW, plus all the cooling and facility work. ([comma.ai blog][1]) but the larger, sneakier bill is: on-call load, hiring niche operators, hardware failures, spare parts, procurement, security, audits, vendor management, and the opportunity cost of your best engineers becoming part-time building managers. cloud is expensive, yes—because it bundles labor, expertise, and economies of scale you don’t have.
## 5) “vendor lock-in” is real, but self-lock-in is worse
cloud lock-in is usually optional: you choose proprietary managed services because they’re convenient. if you’re disciplined, you can keep escape hatches: containers, kubernetes, terraform, postgres, object storage abstractions, multi-region backups, and a tested migration plan. owning your datacenter is also lock-in—except the vendor is past you, and the contract is “we can never stop maintaining this.”
## the practical rule
*if you have massive, predictable, always-on utilization, and you want to become good at running infrastructure as a core competency, owning can win.* that’s basically comma’s case. ([comma.ai blog][1])
*otherwise, rent.* buy speed, buy optionality, and keep your team focused on the thing only your company can do.
if you want, tell me your rough workload shape (steady vs spiky, cpu vs gpu, latency needs, compliance), and i’ll give you a blunt “rent / colo / own” recommendation in 5 lines.
And now go do that in another region. Bam, savings gone. /s
What I mean is that I'm assuming the math here works because the primary purpose of the hardware is training models. You don't need 6 or 7 nines for that is what I'm imagining. But when you have customers across geography that use your app hosted on those servers pretty much 24/7 then you can't afford much downtime.
This is an industry we're[0] in. Owning is at one end of the spectrum, with cloud at the other, and a broadly couple of options in-between:
1 - Cloud – This is minimising cap-ex, hiring, and risk, while largely maximising operational costs (its expensive) and cost variability (usage based).
2 - Managed Private Cloud - What we do. Still minimal-to-no cap-ex, hiring, risk, and medium-sized operational cost (around 50% cheaper than AWS et al). We rent or colocate bare metal, manage it for you, handle software deployments, deploy only open-source, etc. Only really makes sense above €$5k/month spend.
3 - Rented Bare Metal – Let someone else handle the hardware financing for you. Still minimal cap-ex, but with greater hiring/skilling and risk. Around 90% cheaper than AWS et al (plus time).
4 - Buy and colocate the hardware yourself – Certainly the cheapest option if you have the skills, scale, cap-ex, and if you plan to run the servers for at least 3-5 years.
A good provider for option 3 is someone like Hetzner. Their internal ROI on server hardware seems to be around the 3 year mark. After which I assume it is either still running with a client, or goes into their server auction system.
Options 3 & 4 generally become more appealing either at scale, or when infrastructure is part of the core business. Option 1 is great for startups who want to spend very little initially, but then grow very quickly. Option 2 is pretty good for SMEs with baseline load, regular-sized business growth, and maybe an overworked DevOps team!
[0] https://lithus.eu, adam@
I think the issue with this formulation is what drives the cost at cloud providers isn't necessarily that their hardware is too expensive (which it is), but that they push you towards overcomplicated and inefficient architectures that cost too much to run.
A core at this are all the 'managed' services - if you have a server box, its in your financial interest to squeeze as much per out of it as possible. If you're using something like ECS or serverless, AWS gains nothing by optimizing the servers to make your code run faster - their hard work results in less billed infrastructure hours.
This 'microservices' push usually means that instead of having an on-server session where you can serve stuff from a temporary cache, all the data that persists between requests needs to be stored in a db somewhere, all the auth logic needs to re-check your credentials, and something needs to direct the traffic and load balance these endpoint, and all this stuff costs money.
I think if you have 4 Java boxes as servers with a redundant DB with read replicas on EC2, your infra is so efficient and cheap that even paying 4x for it rather than going for colocation is well worth it because of the QoL and QoS.
These crazy AWS bills usually come from using every service under the sun.
The complexity is what gets you. One of AWS's favorite situations is
1) Senior engineer starts on AWS
2) Senior engineer leaves because our industry does not value longevity or loyalty at all whatsoever (not saying it should, just observing that it doesn't)
3) New engineer comes in and panics
4) Ends up using a "managed service" to relieve the panic
5) New engineer leaves
6) Second new engineer comes in and not only panics but outright needs help
7) Paired with some "certified AWS partner" who claims to help "reduce cost" but who actually gets a kickback from the extra spend they induce (usually 10% if I'm not mistaken)
Calling it it ransomware is obviously hyperbolic but there are definitely some parallels one could draw
On top of it all, AWS pricing is about to massively go up due to the RAM price increase. There's no way it can't since AWS is over half of Amazon's profit while only around 15% of its revenue.
29 replies →
Just this week a friend of mine was spinning up some AWS managed service, complaining about the complexity, and how any reconfiguration took 45 minutes to reload. It's a service you can just install with apt, the default configuration is fine. Not only is many service no longer cheaper in the cloud, the management overhead also exceed that of on-prem.
9 replies →
> If you're using something like ECS or serverless, AWS gains nothing by optimizing the servers to make your code run faster - their hard work results in less billed infrastructure hours.
If ECS is faster, then you're more satisfied with AWS and less likely to migrate. You're also open to additional services that might bring up the spend (e.g. ECS Container Insights or X-Ray)
Source: Former Amazon employee
1 reply →
I don’t understand why most cloud backend designs seem to strive for maximizing the number of services used.
My biggest gripe with this is async tasks where the app does numerous hijinks to avoid a 10 minute lambda processing timeout. Rather than structure the process to process many independent and small batches, or simply using a modest container to do the job in a single shot - a myriad of intermediate steps are introduced to write data to dynamo/s3/kinesis + sqs/and coordination.
A dynamically provisioned, serverless container with 24 cores and 64 GB of memory can happily process GBs of data transformations.
Fully agree to this. I find the cost of cloud providers is mostly driven by architecture. If you're cost conscious, cloud architectures need to be up-front designed with this in mind.
Microservices is a killer with cost. For each microservices pod - you're often running a bunch of side cars - datadog, auth, ingress - you pay massive workload separation overhead with orchestration, management, monitoring and ofc complexity
I am just flabbergasted that this is how we operate as a norm in our industry.
It's about fitting your utilization to the model that best serves you.
If you can keep 4 "Java boxes" fed with work 80%+ of the time, then sure EC2 is a good fit.
We do a lot of batch processing and save money over having EC2 boxes always on. Sure we could probably pinch some more pennies if we managed the EC2 box uptime and figured out mechanisms for load balancing the batches... But that's engineering time we just don't really care to spend when ECS nets us most of the savings advantage and is simple to reason about and use.
Agreed. There is a wide price difference between running a managed AWS or Azure MySQL service and running MySQL on a VM that you spin up in AWS or Azure.
> your infra is so efficient and cheap that even paying 4x for it rather than going for colocation is well worth it because of the QoL and QoS.
You don’t need colocation to save 4x though. Bandwidth pricing is 10x. EC2 is 2-4x especially outside US. EBS for its iops is just bad.
Great comment. I agree it's a spectrum and those of us who are comfortable on (4) like yourself and probably us at Carolina Cloud [0] as well, (4) seems like a no brainer. But there's a long tail of semi-technical users who are more comfortable in 2-3 or even 1, which is what ultimately traps them into the ransomware-adjacent situation that is a lot of the modern public cloud. I would push back on "usage-based". Yes it is technically usage-based but the base fee also goes way up and there are also sometimes retainers on these services (ie minimum spend). So of course "usage-based" is not wrong but what it usually means is "more expensive and potentially far more expensive".
[0] https://carolinacloud.io, derek@
The problem is that clouds have easily become 3 or 5 times the price of managed services, 10x the price of option 3, and 20x the price of option 4. To say nothing of the fact that almost all businesses can run fine on "pc under desk" type situations.
So in practice cloud has become the more expensive option the second your spend goes over the price of 1 engineer.
Hetzner is definitely an interesting option. I’m a bit scared of managing the services on my own (like Postgres, Site2Site VPN, …) but the price difference makes it so appealing. From our financial models, Hetzner can win over AWS when you spend over 10~15K per month on infrastructure and you’re hiring really well. It’s still a risk, but a risk that definitely can be worthy.
> I’m a bit scared of managing the services on my own
I see it from the other direction, when if something fails, I have complete access to everything, meaning that I have a chance of fixing it. That's down to hardware even. Things get abstracted away, hidden behind APIs and data lives beyond my reach, when I run stuff in the cloud.
Security and regular mistakes are much the same in the cloud, but I then have to layer whatever complications the cloud provide comes with on top. If cost has to be much much lower if I'm going to trust a cloud provider over running something in my own data center.
You sum it up very neatly. We've heard this from quite a few companies, and that's kind of why we started our ours.
We figured, "Okay, if we can do this well, reliably, and de-risk it; then we can offer that as a service and just split the difference on the cost savings"
(plus we include engineering time proportional to cluster size, and also do the migration on our own dime as part of the de-risking)
I've just shifted my SWE infrastructure from AWS to Hetzner (literally in the last month). My current analysis looks like it will be about 15-20% of the cost - £240 vs 40-50 euros.
Expect a significant exit expense, though, especially if you are shifting large volumes of S3 data. That's been our biggest expense. I've moved this to Wasabi at about 8 euros a month (vs about $70-80 a month on S3), but I've paid transit fees of about $180 - and it was more expensive because I used DataSync.
Retrospectively, I should have just DIYed the transfer, but maybe others can benefit from my error...
4 replies →
> I’m a bit scared of managing the services on my own (like Postgres, Site2Site VPN, …)
Out of interest, how old are you? This was quite normal expectation of a technical department even 15 years ago.
14 replies →
I’m wondering if it makes sense to distribute your architecture so that workers who do most of the heavy lifting are in hetzner, while the other stuff is in costly AWS. On the other hand this means you don’t have easy access to S3, etc.
2 replies →
No amount of money will make me maintain my own dbs. We tried it at first and it was a nightmare.
15 replies →
I don't know. I rent a bare metal server for $500 a month, which is way overkill. It takes almost no time to manage -- maybe a few hours a year -- and can handle almost anything I throw at it. Maybe my needs are too simple though?
Just curious, what is the spec you pay $6000/year for? Where/what is the line between rent vs buy?
3 replies →
Dead on. Recently, 3 and 4 have been compelling. Cloud costs have rocketed up. I started my casual transition to co-lo 2 years ago and just in december finished everything. I have more capacity at about 30% of the cost. If you go option 3, you even get the benefit of 6+ month retro pricing for RAM/storage. I'm running all DDR4, but I have so much of it I don't know what to do with it.
The flip side is that compliance is a little more involved. Rather than, say, carve out a whole swathe of SOC-2 ops, I have to coordinate some controls. It's not a lot, and it's still a lot lighter than I used to do 10+ years ago. Just something to consider.
you're missing 5, what they are doing.
There is a world of difference between renting some cabinets in an Equinix datacenter and operating your own.
Fair point!
5 - Datacenter (DC) - Like 4, except also take control of the space/power/HVAC/transit/security side of the equation. Makes sense either at scale, or if you have specific needs. Specific needs could be: specific location, reliability (higher or lower than a DC), resilience (conflict planning).
There are actually some really interesting use cases here. For example, reliability: If your company is in a physical office, how strong is the need to run your internal systems in a data centre? If you run your servers in your office, then there's no connectivity reliability concerns. If the power goes out, then the power is out to your staff's computers anyway (still get a UPS though).
Or perhaps you don't need as high reliability if you're doing only batch workloads? Do you need to pay the premium for redundant network connections and power supplies?
If you want your company to still function in the event of some kind of military conflict, do you really want to rely on fibre optic lines between your office and the data center? Do you want to keep all your infrastructure in such a high-value target?
I think this is one of the more interesting areas to think about, at least for me!
7 replies →
> 4 - Buy and colocate the hardware yourself – Certainly the cheapest option if you have the skills, scale, cap-ex, and if you plan to run the servers for at least 3-5 years.
Is it still the cheapest after you take into account that skills, scale, cap-ex and long term lock-in also have opportunity costs?
That is why the the second "if" is there.
You can get locked into cloud too.
The lock in is not really long term as it is an easy option to migrate off.
What is the upper limit of Hertzner? Say you have an AWS bill in the $100s of millions, could Hertzner realistically take on that scale?
An interesting question, so time for some 100% speculation.
It sounds like they probably have revenue in the €500mm range today. And given that the bare metal cost of AWS-equivalent bills tends to be a 90% reduction, we'll say a €10mm+ bare metal cost.
So I would say a cautious and qualified "yes". But I know even for smaller deployments of tens or hundreds of servers, they'll ask you what the purpose is. If you say something like "blockchain," they're going to say, "Actually, we prefer not to have your business."
I get the strong impression that while they naturally do want business, they also aren't going to take a huge amount of risk on board themselves. Their specialism is optimising on cost, which naturally has to involve avoiding or mitigating risk. I'm sure there'd be business terms to discuss, put it that way.
4 replies →
Who are you thinking of?
Netflix might be spending as much as $120m (but probably a little less), and I thought they were probably Amazon's biggest customer. Does someone (single-buyer) spend more than that with AWS?
Hertzner's revenue is somewhere around $400m, so probably a little scary taking on an additional 30% revenue from a single customer, and Netflix's shareholders would probably be worried about risk relying on a vendor that is much smaller than them.
Sometimes if the companies are friendly to the idea, they could form a joint venture or maybe Netflix could just acquire Hertzner (and compete with Amazon?), but I think it unlikely Hertzner could take on Netflix-sized for nontechnical reasons.
However increasing pop capacity by 30% within 6mo is pretty realistic, so I think they'd probably be able to physically service Netflix without changing too much if management could get comfortable with the idea
8 replies →
This space of #2 like Lithus is not something I'm very familiar with, so thank you for the comment that piqued my interest!
If you're willing to share, I'm curious who else you would describe as being in this space.
My last decade and a half or so of experience has all been in cloud services, and prior to that it was #3 or #4. What was striking to me when I went to the Lithus website was that I couldn't figure out any details without hitting a "Schedule a Call" button. This makes it difficult for me to map my experiences in using cloud services onto what Lithus offers. Can I use terraform? How does the kubernetes offering work? How does the ML/AI data pipelines work? To me, it would be nice if I could try it out in a very limited way as self-service, or at least read some technical documentation. Without that, I'm left wondering how it works. I'm sure this is a conscious decision to not do this, and for good reasons, but I thought I'd share my impressions!
Hello! I think this is a fair question, and improving the communication on the website is something that is steadily climbing up our priority list.
We're not really that kind of product company; we're more of a services company. What we do is deploy Kubernetes clusters onto bare metal servers. That's the core technical offering. However, everything beyond that is somewhat per-client. Some clients need a lot of compute. Some clients need a custom object storage cluster. Some clients need a lot of high-speed internal networking. Which is why we prefer to have a call to figure out specifically what your needs are. But I can also see how this isn't necessarily satisfying if you're used to just grabbing the API docs and having a look around.
What we will do is take your company's software stack and migrate it off AWS/Azure/Google and deploy it onto our new infrastructure. We will then become (or work with) your DevOps team to supporting you. This can be anything from containerising workloads to diagnosing performance issues to deploying a new multi-region Postgres cluster. Whatever you need done on your hardware that we feel we can reasonably support. We are the ones on-call should NATS fall over at 4am.
Your team also has full access to the Kubernetes cluster to deploy to as you wish.
I think the pricing page is the most concrete thing on our website, and it is entirely accurate. If you were to phone us and say, "I want that exact hardware," we would do it for you. But the real value we also offer is in the DevOps support we provide, actually doing the migration up-front (at our own cost), and being there working with your team every week.
1 reply →
> Option 1 is great for startups
Unfortunately, (successful) startups can quickly get trapped into this option. If they're growing fast, everyone on the board will ask why you'd move to another option at the first place. The cloud becomes a very deep local minimum that's hard to get out off.
Can someone explain 2 to me. How is a managed private cloud different from full cloud? Like you are still using AWS or Azure but you are keeping all your operation in a bundled, portable way, so you can leave that provider easily at any time, rather than becoming very dependent on them? Is it like staying provider-agnostic but still cloud based?
To put it plainly: We deploy a Kubernetes cluster on Hetzner dedicated servers and become your DevOps team (or a part thereof).
It works because bare metal is about 10% the cost of cloud, and our value-add is in 1) creating a resilient platform on top of that, 2) supporting it, 3) being on-call, and 4) being or supporting your DevOps team.
This starts with us providing a Kubernetes cluster which we manage, but we also take responsibility for the services run on it. If you want Postgres, Redis, Clickhouse, NATS, etc, we'll deploy it and be SLA-on-call for any issues.
If you don't want to deal with Kubernetes then you don't have to. Just have your software engineers hand us the software and we'll handle deployment.
Everything is deployed on open source tooling, you have access to all the configuration for the services we deploy. You have server root access. If you want to leave you can do.
Our customers have full root access, and our engineers (myself included) are in a Slack channel with you engineers.
And, FWIW, it doesn't have to be Hetzner. We can colocate or use other providers, but Hetzner offer excellent bang-per-buck.
Edit: And all this is included in the cluster price, which comes out cheaper than the same hardware on the major cloud providers
4 replies →
Instead of using the Cloud's own Kubernetes service, for example, you just buy the compute and run your own Kubernetes cluster. At a certain scale that is going to be cheaper if you have to know how. And since you are no longer tied to which services are provided and you just need access to compute and storage. you can also shop around for better prices than Amazon or Azure since you can really go to any provider of a VPS.
Getting rid of bureaucratic internal IT department is a game changer for productivity. That alone is worth 10x infra costs, especially for big companies where work can grind to a halt dealing with obstructionists through service now. Good leaders understand this.
Sadly true. Or, the so-called internal IT Dept. can be a shambolic mess of PHB's, Brunchlords, Catberts, metric maximizers, and micromanagers, presiding over the hollowed-out and burned out remains of the actual workforce that you'd need to reliably do the job.
#2.5ish
We rent hardware and also some VPS, as well as use AWS for cheap things such as S3 fronted with Cloudflare, and SES for priority emails.
We have other services we pay for, such as AI content detection, disposable email detection, a small postal email server, and more.
We're only a small business, so having predictable monthly costs is vital.
Our servers are far from maxed out, and we process ~4 million dynamic page and API requests per day.
I am using something inbetween 2 and 3, a hosted Web-site and database service with excellent customer support. On shared hardware it is 22 €/month. A managed server on dedicated hardware starts at about 50 €/month.
5. On-premise and engineers touch the wires every few days.
Where do AWS reserved instances come into your hierarchy? What if there existed a “perpetual” reserved instance? Is cap-ex vs. op-ex really the key distinction?
Been using Hetzner Cloud for Kubernetes and generally like it, but it has its limitations. The network is highly unpredictable. You at best get 2Gbit/s, but at worst a few hundreds of Mbit/s.
https://docs.hetzner.com/cloud/technical-details/faq/#what-k...
Is that for the virtual private network? I heard some people say that you actually get higher bandwidth if you're using the public network instead of the private network within Hetzner, which is a little bit crazy.
1 reply →
We looked at option 4. And colocation is not cheap. It was cheaper for us to lease VMs from Hetzner than to buy boxes and colocate at Equinix.
this is what we did in the 90ies into mid 2000:
> Buy and colocate the hardware yourself – Certainly the cheapest option if you have the skills
back then this type of "skill" was abundant. You could easily get sysadmin contractors who would take a drive down to the data-center (probably rented facilities in a real-estate that belonged to a bank or insurance) to exchange some disks that died for some reason. such a person was full stack in a sense that they covered backups, networking, firewalls, and knew how to source hardware.
the argument was that this was too expensive and the cloud was better. so hundreds of thousands of SME's embraced the cloud - most of them never needed Google-type of scale, but got sucked into the "recurring revenue" grift that is SaaS.
If you opposed this mentality you were basically saying "we as a company will never scale this much" which was at best "toxic" and at worst "career-ending".
The thing is these ancient skills still exist. And most orgs simply do not need AWS type of scale. European orgs would do well to revisit these basic ideas. And Hetzner or Lithus would be a much more natural (and honest) fit for these companies.
I wonder how much companies pay yearly in order to avoid having an employee pick up a drive from a local store, drive to the data center, pull the disk drive, screw out the failing hard drive and put in the new one, add it in the raid, verify the repair process has started, and then return to the office.
3 replies →
> ancient skills https://youtu.be/ZtYU87QNjPw?&t=10
It baffles me that my career trajectory somehow managed to insulate me from ever having to deal with the cloud, while such esoteric skills as swapping a hot swap disk or racking and cabling a new blade chassis are apparently on the order of finding a COBOL developer now. Really?
I can promise you that large financial institutions still have datacenters. Many, many, many datacenters!
1 reply →
[dead]
if someone on the DevOps team knows Nix, option 3 becomes a lot cheaper time-wise! yeah, Nix flakes still need maintenance, especially on the `nixos-unstable` branch, but you get the quickest disaster recovery route possible!
plus, infra flexibility removes random constraints that e.g. Cloudflare Workers have
There are a bunch of ways to manage bare metal servers apart from Nix. People have been doing it for years. Kickstart, theforeman, maas, etc, [0]. Many to choose from according to your needs and layers you want them to manage.
Reality is these days you just boot a basic image that runs containers
[0] Longer list here: https://github.com/alexellis/awesome-baremetal
Indeed! We've yet to go down this route, but it's something we're thinking on. A friend and I have been talking about how to bring Nix-like constructs to Kubernetes as well, which has been interesting. (https://github.com/clotodex/kix, very much in the "this is fun to think about" phase)
This is what we do, I gave a talk about our setup earlier this week at CfgMgmtCamp: https://www.youtube.com/watch?v=DBxkVVrN0mA&t=8457s
Option 4 as well, that's how we do it at work and it's been great. However, it can't really be "someone on the team knows Nix", anyone working on Ops will need Nix skills in order to be effective.
1 reply →
I'm a NixOS fan, but been using Talos Linux on Hetzner nodes (using Cluster-API) to form a Kubernetes Cluster. Works great too!
Everything comes circle. Back in my day, we just called it a "data center". Or on-premise. You know, before the cloud even existed. A 1990s VP of IT would look at this post and say, what's new? Better computing for sure. Better virtualization and administration software, definitely. Cooling and power and racks? More of the same.
The argument made 2 decades ago was that you shouldn't own the infrastructure (capital expense) and instead just account for the cost as operational expense (opex). The rationale was you exchange ownership for rent. Make your headache someone else's headache.
The ping pong between centralized vs decentralized, owned vs rented, will just keep going. It's never an either or, but when companies make it all-or-nothing then you have to really examine the specifics.
There's a very interesting insight from your message.
The Cloud providers made a lot of sense to finance departments since aside from the promised savings, you would take that cloud expense now and lower your tax rate.
After the passing of the One Beautiful Bill ("OBB"), the law allows you to accelerate CapEx to instead expense it[1], similar to the benefit given by cloud service providers.
This puts way more wind on the sails of the on-prem movement, for sure
[1] https://www.iqxbusiness.com/big-beautiful-bill-impact-on-cap...
> you shouldn't own the infrastructure (capital expense) and instead just account for the cost as operational expense (opex)
That was part of the reason.
The real reason was the internal infrastructure team in many orgs got nowhere. There was a huge queue and many teams instead had to find infinite workarounds including standing up their own. The "cloud" provided a standardized way to at least deal with this mess e.g. single source of billing.
> A 1990s VP of IT would look at this post and say, what's new?
Speed. The US lives in luxury but outside of that it often takes a LONG time to get proper servers. You don't just go online. There are many places where you have to talk to a vendor with no list price and the drama continues. Being out of capacity can mean weeks to months before you get anywhere.
Yep! The biggest win for me when AWS came out was that I could self-serve what I needed and put it on a credit card, rather than filing a ticket and waiting some number of days / weeks / months to get a new VM approved and deployed.
I agree - my reference to the 1990s VP of IT was looking at the post, which is about on-premise data centers... not the cloud. I don't think there's a speed advantage for on-premise data centers now vs the 1990s, but if there is let me know. Otherwise, indeed, it's a 1990s-era blast from the past.
Agreed. Also, a realistic assessment should not downplay the very real overhead and headache of managing your on-premise data center. It comes at a cost in engineering/firefighting hours, it's not painless. There's a reason this eternal ping pong keeps going on!
Yeah, I think the major improvement of cloud services was the rationalization of them into services with a cost instead of "ask that person for a whatsit" and "hopefully the associate goomba will approve."
All teams will henceforth expose their data and functionality through service interfaces
https://gist.github.com/chitchcock/1281611
>San Diego has a mild climate and we opted for pure outside air cooling. This gives us less control of the temperature and humidity, but uses only a couple dozen kW. We have dual 48” intake fans and dual 48” exhaust fans to keep the air cool. To ensure low humidity (<45%) we use recirculating fans to mix hot exhaust air with the intake air. One server is connected to several sensors and runs a PID loop to control the fans to optimize the temperature and humidity.
Oh man, this is bad advice. Airborn humidity and contaminants will KILL your servers on a very short horizon in most places - even San Diego. I highly suggest enthalpy wheel coolers (kyotocooling is one vendor - switch datacenters runs very similar units on their massive datacenters in the Nevada desert) as they remove the heat from the indoor air using outdoor air (+can boost slightly with an integrated refrigeration unit to hit target intake temps) without switching the air from one side to the other. This has huge benefits for air control quality and outdoor air tolerance and a single 500KW heat rejection unit uses only 25KW of input power (when it needs to boost the AC unit's output). You can combine this with evaporative cooling on the exterior intakes to lower the temps even further at the expense of some water consumption (typically far cheaper than the extra electricity to boost the cooling through an hvac cycle).
Not knocking the achievement just speaking from experience that taking outdoor air (even filtered + mixed) into a datacenter is a recipe for hardware failure and the mean time to failure for that is highly dependant on your outdoor air conditions. I've run 3MW facilities with passive air cooling and taking outdoor air directly into servers requires a LOT more conditioning and consideration than is outlined in this article.
Yes, it's easy to destroy the servers with a lot of dust and/or high humidity. But with filtering and ensuring humidity never exceeds 45% we've had pretty good results.
I remember visiting a small data center (about half the size of the Comma one) where shoe covers were required. Apparently they were worried about people’s shoes bringing in dust and other contamination.
It's not a static number as it's also based on ambient air temperature in the form of dew point - 45% RH at low temps can be far more dangerous than 65% RH at warm ambient.
Likewise the impact on server longevity is not a finite boundary but rather "exposure over time" gradient that, if exceeding the "low risk" boundary (>-12'C/10'f dew point or >15'C/59'f dry bulb temp) results in lower MTBF than design. This is defined (and server equipment manufacturers conform and build to) ASHRAE TC 9.9. This mean - if you're running your servers above high risk curve for humidity and temperature, you're shortening the life considerably compared to low risk curve.
Generally, 15% RH is considered suboptimal and can be dangerous near freezing temperatures - in San Diego in January there were several 90%+RH scenarios that would have been dangerous for servers even when mixed down with warm exhaust air - furthermore, the outdoor air at 76'f during that period means you have limited capacity to mix in warm exhaust air (which btw came from that same 99%RH input air) without getting into higher-than-ideal intake temps.
Any dew points above 62.5'f are considered high risk for servers - as are any intake temps exceeding 32'C/90'f. You want to be on the midpoint between those and 16'C/65'f temps & -12'C/10'f dew point to have no impact on server longevity or MTBF rates.
As a recent example:
Lastly, air contaminants - in the form of dust (that can be filtered out) and chemicals (which can't without extensive scrubbing) are probably the most detrimental to server equipment if not properly managed, and require very intentional and frequent filter changes (typically high MERV pleated filters changed on a time or pressure drop signal) to prevent server degradation and equipment risks.
The last consideration is fire suppression - permitted datacenters usually require compliance with separate fire code, such that direct outdoor air exchange without active shutdown and dry suppression is not permitted - this is to prevent a scenario where your equipment catches on fire and a constant supply of fresh oxygen-rich outdoor air turns that into an inferno. Smoke detection systems don't operate well with outdoor-mixed air or any level of airborn particulates.
So - for those reasons - among a few others - open air datacenters are not recommended unless you're doing them at google or meta scale, and in those scenarios you typically have much more extensive systems and purpose-designed hardware in order to operate for the design life of the equipment without issues.
I didn't even know this is something you had to worry about. This is why I use the cloud, all the unknown unknowns.
LOL’ed IRL at “ In a future blog post I hope I can tell you about how we produce our own power and you should too.” Producing own power as a pre-requisite for running on-prem is a non-starter for many.
I would suggest to use both on-premise hardware and cloud computing. Which is probably what comma is doing.
For critical infrastructure, I would rather pay a competent cloud provider than being responsible for reliability issues. Maintaining one server room in the headquarters is something, but two servers rooms in different locations, with resilient power and network is a bit too much effort IMHO.
For running many slurm jobs on good servers, cloud computing is very expensive and you sometimes save money in a matter of months. And who cares if the server room is a total loss after a while, worst case you write some more YAML and Terraform and deploy a temporary replacement in the cloud.
Another thing between is colocation, where you put hardware you own in a managed data center. It’s a bit old fashioned, but it may make sense in some cases.
I can also mention that research HPCs may be worth considering. In research, we have some of the world fastest computers at a fraction of the cost of cloud computing. It’s great as long as you don’t mind not being root and having to use slurm.
I don’t know in USA, but in Norway you can run your private company slurm AI workloads on research HPCs, though you will pay quite a bit more than universities and research institutions. But you can also have research projects together with universities or research institutions, and everyone will be happy if your business benefits a lot from the collaboration.
> but two servers rooms in different locations, with resilient power and network is a bit too much effort IMHO
I worked in a company with two server farms (a main and a a backup one essentially) in Italy located in two different regions and we had a total of 5 employees taking care of them.
We didn't hear about them, we didn't know their names, but we had almost 100% uptime and terrific performance.
There was one single person out of 40 developers who's main responsibility were deploys, and that's it.
It costed my company 800k euros per year to run both the server farms (hardware, salaries, energy), and it spared the company around 7-8M in cloud costs.
Now I work for clients that spend multiple millions in cloud for a fraction of the output and traffic, and I think employ around 15+ dev ops engineers.
it depends on complexity of your infra.
Running full scale kubernets, with multiple databases and services and expected 99.99% uptime likely can't be handled by one person.
1 reply →
> I would rather pay a competent cloud provider than being responsible for reliability issues.
Why do so many developers and sysadmins think they're not competent for hosting services. It is a lot easier than you think, and its also fun to solve technical issues you may have.
The point was about redundancy / geo spread / HA. It’s significantly more difficult to operate two physical sites than one. You can only be in one place at a time.
If you want true reliability, you need redundant physical locations, power, networking. That’s extremely easy to achieve on cloud providers.
4 replies →
Also I'd add this question, why do so many developers and sysadmins think, that cloud companies always hire competent/non-lazy/non-pissed employees?
> Why do so many developers and sysadmins think they're not competent for hosting services. It is a lot easier than you think, and its also fun to solve technical issues you may have.
It is a different skillset. SRE is also an under-valued/paid (unless one is in FAANGO).
3 replies →
Maybe you find it fun. I don’t, I prefer building software not running and setting up servers.
It’s also nontrivial once you go past some level of complexity and volume. I have made my career at building software and part of that requires understanding the limitations and specifics of the underlying hardware but at the end of the day I simply want to provision and run a container, I don’t want to think about the security and networking setup it’s not worth my time.
Because when I’m running a busy site and I can’t figure out what went wrong, I freak out. I don’t know whether the problem will take 2 hours or 2 days to diagnose.
1 reply →
> Why do so many developers and sysadmins think they're not competent for hosting services.
Because those services solve the problem for them. It is the same thing with GitHub.
However, as predicted half a decade ago with GitHub becoming unreliable [0] and as price increases begin to happen, you can see that self-hosting begins to make more sense and you have complete control of the infrastructure and it has never been more easier to self host and bring control over costs.
> its also fun to solve technical issues you may have.
What you have just seen with coding agents is going to have the same effect on "developers" that will have a decline in skills the moment they become over-reliant on coding agents and won't be able to write a single line of code at all to fix a problem they don't fully understand.
[0] https://news.ycombinator.com/item?id=22867803
At a previous job, the company had its critical IT infrastructure on their own data center. It was not in the IT industry, but the company was large and rich enough to justify two small data centers. It notably had batteries, diesel generators, 24/7 teams, and some advanced security (for valid reasons).
I agree that solving technical issues is very fun, and hosting services is usually easy, but having resilient infrastructure is costly and I simply don't like to be woken up at night to fix stuff while the company is bleeding money and customers.
> Maintaining one server room in the headquarters is something, but two servers rooms in different locations, with resilient power and network is a bit too much effort IMHO.
Speaking as someone who does this, it is very straightforward. You can rent space from people like Equinix or Global Switch for very reasonable prices. They then take care of power, cooling, cabling plant etc.
Yes, we still use the azure for user-facing services and the website. They don't need GPUs and don't need expensive resources, so it's not as worth it to bring those in-house.
We also rely on github. It has historically been good a service, but getting worth it.
I don't get why most everyone insists on comparing cloud to on-premises and not to dedicated. Why would anyone run own DC infra when there's Hetzner and many others?
Unfortunately we experienced an issue where our Slurm pool was contaminated by a misconstrained Postgres Daemon. Normally the contaminated slurm pool would drain into a docker container, but due to Rust it overloaded and the daemon ate its own head. Eventually we returned it to a restful state so all's well that ends well.
(hardware engineer trying to understand wtaf software people are saying when they speak)
This is cool. Yet, there are levels of insanity and those depend on your inability to estimate things.
When I'm launching a project it's easier for me to rent $250 worth of compute from AWS. When the project consumes $30k a month, it's easier for me to rent a colocation.
My point is that a good engineer should know how to calculate all the ups and downs here to propose a sound plan to the management. That's the winning thing.
It goes further than this first order, though. If you're trying to build a business that attracts the types of talent who wants to know the stack up and down, starting with an AWS instance might give you a better shot at funding (and thus a better overall shot), but it's not clear that it gives you a shot a building the business you're aiming for. For the things that "don't make your beer better", sure, but we're talking about training ML models at an ML shop. Here it makes sense for this reason.
That last part is exactly it and I while I know the intro sentence nails it I don’t think compute resonates with people (everyone uses compute). If you are 24/7 running work at scale it absolutely makes sense past the initial first couple years to build out your own DC like this.
We’re past the point in history where most engineers get to make even a recommendation about which platform to use to management.
In 99.999999% of cases management has already decided and is just informing you, because they know better.
I work in a multi-billions dollars company and do not face what you describe
Perhaps an exception (yet so far, I've never encounter the situation you describe)
Feels like I’ve lived through a full infrastructure fashion cycle already. I started my career when cloud was the obvious answer and on-prem was “legacy.”
Now on-prem is cool again.
Makes me wonder whether we’re already setting up the next cycle 10 years from now, when everyone rediscovers why cloud was attractive in the first place and starts saying “on-prem is a bad idea” again.
> Makes me wonder whether we’re already setting up the next cycle 10 years from now, when everyone rediscovers why cloud was attractive in the first place and starts saying “on-prem is a bad idea” again.
My entire career I’ve encountered people passionately pushing for on-prem and railing against anything cloud. I can’t remember a time when Hacker News comments leaned pro-cloud because it’s always been about self-hosting.
The few times the on-prem people won out in my career never went exactly as they imagined. Buying a couple servers and setting them up at the colo is easy enough, but the slow and steady drag of maintaining your own infrastructure starts to work its way into every development cycle after that. In my experience, every team has significantly underestimated how all the little things add up to a drag on available time for other work.
The best case for on-prem that I saw was when a company was basically in maintenance mode. Engineers had a lot of extra time to optimize, update. maintain, and cost reduce without subtracting from feature development or bug fixes.
The worst cases for on-prem I’ve seen have been funded startups. In this situation it’s imperative that everyone focus on feature development and rapid iteration. Letting some of the engineers get sidetracked with setting up and maintaining their own hosting to save a dollar amount that barely hires 1-2 more engineers but sets the schedule back by many months was a huge mistake.
In my experience, most engineers become less enchanted with rolling their own on premises hosting as they get older. Their work becomes more about getting the job done quickly and to budget, not hyper-optimizing the hosting situation at the expense of inviting more complexity and miscellaneous tasks into their workload.
If this were cyclical, I'd be inclined to agree, but this seems to be more of a wave. I also think the push back is more than just one against rented compute. It is tied to a societal ennui that comes from the feeling that we no longer own anything, be it music, housing, movies, land, tools, phones, or cars. Everything is moving to either being rented or on credit. There's a push back against this self-made feudal revival, and that scales all the way from individuals through to corporations; in this case, against the idea that a mega-corporation gets to decide how and when you get to use your compute, and at what variable price.
Just one cycle?
This is cyclical and I see the main axis of contention as centralized vs de-centralized computing.
Mainframes (network) gave way to mini and microcomputers (PCs). PCs gave way to server farms and web-based applications. Private servers and data centers gave way to the Cloud. Edge computing is again a push towards a more decentralized model.
Like all good engineering problems, where data and applications are hosted involve tradeoffs. Priorities change. Technologies change. But oftentimes, what works in one generation doesn't in the next. Part of it is the slow march of progress. But I think some of it is just not wanting to use your parent's technology stack and wanting to build your own.
The cloud vs. on-prem tradeoff is one of flexibility, capacity, maintenance, and capex vs opex.
It's a similar story in application development. At one point, we're navigating text forms on a mainframe, the next it's a GUI local application, followed by Electron or Web applications with remote data. We'll cycle back to local-first data (likely on-phone local models).
When you start to hear about the network being the computer again, you'll know we've started to swing back the other way again.
Mainframe -> Desktop -> Server Room -> Data Center -> Cloud (rented data center) -> Space (Skynet)
Sometimes, I feel like this is indicative of the incredible waste present in IT and development. Granted the cost of this kind of infrastructure upheaval is orders of magnitude cheaper than something like manufacturing - but still, it feels ridiculous that established companies can swap back and forth on a whim.
The problem was always the platform. For me, I saw very early on that kubernetes was exactly what I wanted after reading about how Google "treats the datacenter like one large computer." And I've been very happily running my own side projects on my own home cluster for 10 ish years (my kube-system namespace is 9y old). But selling any of my employers on this was a very hard proposition until enough people had shown it working at that scale.
On premises isn't only about saving money (that's not always clear). The article neglects the most important benefits which are freedom (control) and privacy. It's basically the same considerations that apply to owning vs renting a house.
The entire second section is about different benefits of having your own data centers. Cost is listed as the last one, not the primary one.
The lowest grade I got in my business degree was in the "IT management" course. That's because the ONLY acceptable answer to any business IT problem is to move everything to the cloud. Renting is ALWAYS better than owning because you transfer cost and risk to a 3rd party.
That's pretty much the dogma of the 2010s.
It doesn't matter that my org runs a line-of-business datacentre that is a fraction of the cost of public cloud. It doesn't matter that my "big" ERP and admin servers take up half a rack in that datacentre. MBA dogma says that I need to fire every graybeard sysadmin, raze our datacentre facility to the ground, and move to AWS.
Fun fact, salaries and hardware purchases typically track inflation, because switching cost for hardware is nil and hiring isn't that expensive. Whereas software is usually 5-10% increases every year because they know that vendor lock-in and switching costs for software are expensive.
Right, but is that a like for like comparison?
AWS has redundant data centres across the world and within each region. A file in S3 will never be lost, even if you store it for a thousand years.
What happens if your city has a tornado and your data centre gets hit? Is your company now dead?
And how much do you spend on all these sysadmins? 200k each? If you’re saving 20k/month by paying 100k/month in salaries, you aren’t saving anything.
I was an on-prem maxi (if thats a thing) for a long time. I've run clusters that costed more than $5M, but these days I am a changed man. I start with PaaS like Vercel and work my way down to on-prem depending on how important and cost conscious that workload is.
Pains I faced running BIG clusters on-prem.
1. Supply chain Management -- everything from power supplies all the way to GPUs and storage has to be procured, shipped, disassembled and installed. You need labor pool and dedicated management.
2. Inventory Management -- You also need to manage inventory on hand for parts that WILL fail. You can expect 20% of your cluster to have some degree of issues on an ongoing basis
3. Networking and security -- You are on your own defending your network or have to pay a ton of money to vendors to come in and help you. Even with the simplest of storage clusters, we've had to deal with pretty sophisticated attacks.
When I ran massive clusters, I had a large team dealing with these. Obviously, with PaaS, you dont need anyone.
> I was an on-prem maxi (if thats a thing) for a long time. I've run clusters that costed more than $5M, but these days I am a changed man.
I have had a similar transformation. I still host non-critical services on-prem. They are exceptionally cheap to run. Everything else, I host it on Hetzner.
In addition to those sorts of non-first-hardware-purchase costs, the person writing the check needs to think long and hard about how bad an outage would be, and how much money it makes sense to budget simply to "avoiding outages." And the more important it is not to have any downtime, the more it's gonna cost to build up some sort of substitute for cross-datacenter cloud functionality. (You are also likely not going to be as good at either managing and configuring those networks, or hiring people to do so, as AWS, either.)
At scale (like comma.ai), it's probably cheaper. But until then it's a long term cost optimization with really high upfront capital expenditure and risk. Which means it doesn't make much sense for the majority of startup companies until they become late stage and their hosting cost actually becomes a big cost burden.
There are in between solutions. Renting bare metal instead of renting virtual machines can be quite nice. I've done that via Hetzner some years ago. You pay just about the same but you get a lot more performance for the same money. This is great if you actually need that performance.
People obsess about hardware but there's also the software side to consider. For smaller companies, operations/devops people are usually more expensive than the resources they manage. The cost to optimize is that cost. The hosting cost usually is a rounding error on the staffing cost. And on top of that the amount of responsibilities increases as soon as you own the hardware. You need to service it, monitor it, replace it when it fails, make sure those fans don't get jammed by dust puppies, deal with outages when they happen, etc. All the stuff that you pay cloud providers to do for you now becomes your problem. And it has a non zero cost.
The right mindset for hosting cost is to think of it in FTEs (full time employee cost for a year). If it's below 1 (most startups until they are well into scale up territory), you are doing great. Most of the optimizations you are going to get are going to cost you in actual FTEs spent doing that work. 1 FTE pays for quite a bit of hosting. Think 10K per month in AWS cost. A good ops person/developer is more expensive than that. My company runs at about 1K per month (GCP and misc managed services). It would be the wrong thing to optimize for us. It's not worth spending any amount of time on for me. I literally have more valuable things to do.
This flips when you start getting into the multiple FTEs per month in cost for just the hosting. At that point you probably have additional cost measured in 5-10 FTE in staffing anyway to babysit all of that. So now you can talk about trading off some hosting FTEs for modest amount of extra staffing FTEs and make net gains.
> At scale (like comma.ai), it's probably cheaper. But until then it's a long term cost optimization with really high upfront capital expenditure and risk. Which means it doesn't make much sense for the majority of startup companies until they become late stage and their hosting cost actually becomes a big cost burden.
You rent a dataspace, which is OPEX not CAPEX, and you just lease the servers, which turns big CAPEX into monthly OPEX bill
Running your own DC is "we have two dozen racks of servers" endeavour, but even just renting DC space and buying servers is much cheaper than getting same level of performance from the cloud.
> This flips when you start getting into the multiple FTEs per month in cost for just the hosting. At that point you probably have additional cost measured in 5-10 FTE in staffing anyway to babysit all of that. So now you can talk about trading off some hosting FTEs for modest amount of extra staffing FTEs and make net gains.
YOU NEED THOSE PEOPLE TO MANAGE CLOUD TOO. That's what always get ignore in calculations, people go "oh, but we really need like 2-3 ops people to cover datacenter and have shifts on the on-call", but you need same thing for cloud too, it is just dumped on programmers/devops guys in the team rather than having separate staff.
We have few racks and the part related to hardware is small part of total workload, most of it is same as we would (and do for few cloud customers) in cloud, writing manifests for automation.
> YOU NEED THOSE PEOPLE TO MANAGE CLOUD TOO.
Finally, some sense! "Cloud" was meant to make ops jobs disappear, but they just increased our salary by turning us into "DevOps Engineers" and the company's hosting bill increased fivefold in the process. You will never convince even 1% of devs to learn the ops side properly, therefore you'll still end up hiring ops people and we will cost you more now. On top of that, everyone that started as a "DevOps Engineer" knows less about ops than those that started as ops and transitioned into being "DevOps Engineers" (or some flavour of it like SREs or Platform Engineers).
If you're a programmer scared into thinking AI is going to take away your job, re-read my comment.
I'm not disagreeing... but it depends on how you shift the complexity/work and how you lean into or don't lean into the services a given cloud provider offer or not.
Just database management is a pretty specialized skill, separate from development or optimizing the structures of said data... For a lot of SaaS providers, if you aren't at a point where you can afford a dedicated DBA/Ops staff just for data, that's one reason you might lean into cloud operations or hybrid ops just for dbms management, security and backups. This is a low hanging fruit in terms of cloud offerings evem... but can shift a lot of burden in terms of operational overhead.
Again, depending on your business and data models.
Honestly, the way I've seen a lot of cloud done, they need _more_ people to manage that than a sensible private cloud setup.
To be fair, I think people are vastly over estimating the work they would have and the power they would need. Yes, if you have to massively scale up, then it'll take some work, but most of it is one-time work. You do it, and when it runs, you only have a fraction of work over the next months to maintain it. And with fraction, I mean below 5%. And keep in mind that >99% of startups who think of "yeah we need this and that cloud, because we need to scale" will never scale. Instead they are happily locking themselves into a cloud service. And if they actually scale at some point, this service will be massively more expensive.
One decent server would be enough to run 99.5% of startups backends.
We have two on site servers that we use. For various reasons (power cuts, internet outages, cleaners unplugging them) I’d say we have to intervene with them physically about once a month. It’s a total pain in the ass, especially when you don’t have _an_ it person sitting in the office to mind it. I’m in the Uk and our office is in Spain…
But it is significantly cheaper and faster
1 reply →
Startups don't know how much hardware they need when they release to customers. The extreme flexibility of cloud makes a lot of sense for them.
2 replies →
Your calculation assumes that an FTE is needed to maintain a few beefy servers.
Once they are up and running that employee is spending at most a few hours a month on them. Maybe even a few hours every six months.
OTOH you are specifically ignoring that you'll require mostly the same time from a cloud trained person if you're all-in on AWS.
I expect the marginal cost of one employee over the other is zero.
> Once they are up and running
You should also calculate the cost of getting it up and running. With Google Cloud (I don't actually use AWS), I mainly worry about building docker containers in CI and deploying them to vms and triggering rolling restarts as those get replaced with new ones. I don't worry about booting them. I don't worry about provisioning operating systems or configuration to them. Or security updates. They come up with a lot of pre-provisioned monitoring and other stuff. No effort required on my side.
And for production setups. You need people on stand by to fix the server in case of hardware issues; also outside office hours. Also, where does the hardware live? What's your process when it fails? Who drives to wherever the thing is and fixes it? What do you pay them to be available for that? What's the lead time for spare components? Do you actually keep those in supply? Where? Do you pay for security for wherever all that happens? What about cleaning, AC, or a special server room in your building. All that stuff is cost. Some of it is upfront cost. Some of it is recurring cost.
The article is a about a company that owns its own data center. The cost they are citing (5 million) is substantial and probably a bit more complete. That's one end of the spectrum.
4 replies →
Maintainer is a real work
If you do it only a few hours every 6 months, you are not maintaining your infrastructure, you are letting it die (until the need arises and everything must be done and this is a massive project)
On the software side... depending on your business model, you can factor in a lot of the cost structures into your structure. Especially for say B2B arrangements.
Cloud integrations, for example, allow you to simply use a different database instance altogether per customer, while you can share services that utilize a given db connection. But actually setting up and managing that type of database infrastructure yourself may be much more resource intensive from a head count perspective.
I mention this, because having completely separate databases is an abstraction that cloud operations have already solved... while you can choose other options, such as more complex data models to otherwise isolate or share resources how does this complexity affect your services down-stream and the overall data complexities across one or all clients.
Harder still, if your data/service is centered around b2b clients of yours that have direct consumer interactions... then what if the industry is health or finance where there are even more legal concerns. Figuring a minimal (off the top) cost of each client of yours and scaling to the number of users under them isn't too hard to consider if you're using a mix of cloud services in concert with your own systems/services.
So yeah.. there's definitely considerations in either direction.
> it doesn't make much sense for the majority of startup companies until they become late stage
Here's what TFA says about this:
> Cloud companies generally make onboarding very easy, and offboarding very difficult. If you are not vigilant you will sleepwalk into a situation of high cloud costs and no way out.
and I think they're right. Be careful how you start because you may be stuck in the initial situation for a long time.
> But until then it's a long term cost optimization with really high upfront capital expenditure and risk.
The upfront capex does not need to be that high, unless you're running your own AI models. Other than leasing new ones, as a sibling comment stated, you can buy used. You can get a solid Dell 2U with a full service contract (3 years) for ~$5-10K depending on CPU / memory / storage configuration. Or if you don't mind going older - because honestly, most webapps aren't doing anything compute-heavy - you can drop that to < $1K/node. Replacement parts for those are cheap, so buy an extra of everything.
And if each of your clients is in the Healthcare industry and dealing with end-user medical data? Or financial data? Are you prepared for appropriate data isolation/sharding and controls? Do you have a strategy for scaling database operations per client or across all clients?
It really depends on the business model as to how well you might support your own infrastructure vs. relying on a new backend instance per client in a cloud infrastructure that has already solved many of the issues at play.
3 replies →
>At scale (like comma.ai), it's probably cheaper. But until then it's a long term cost optimization with really high upfront capital expenditure and risk.
The issue with comma.ai is that the company is HEAVILY burdened with Geohotz ideals, despite him no longer even being on the board. I used to be very much into his streams and he rants about it plenty. A large reason of why they run their own datacenter is that they ideologically refuse to give money to AWS or Google (but I guess Microsoft passes their non woke test).
Which is quite hilarious to me because they live in a very "woke" state and complain about power costs in the blog post. They could easily move to Wyoming or Montana and with low humidity and colder air in the winter run their servers more optimally.
Our preference for training in our own datacenter has nothing to do with wokeness. Did you read the blog post? The reasons are clearly explained.
The climate in Wyoming and Montana are actually worse in terms of climate. San Diego's climate extremes are less extreme than those places. Though moving out of CA is a good idea for power cost reasons, also addressed in the blog.
And not just any FTEs, probably few senior / staff level engineers who would cost a lot more.
You should keep in mind that for a lot of things you can use a servicing contract, rather than hiring full-time employees.
It's typically going to cost significantly less; it can make a lot of sense for small companies, especially.
The reason companies don’t go with on premises even if cloud is way more expensive is because of the risk involved in on premises.
You can see it quite clearly here that there’s so many steps to take. Now a good company would concentrate risk on their differentiating factor or the specific part they have competitive advantage in.
It’s never about “is the expected cost in on premises less than cloud”, it’s about the risk adjusted costs.
Once you’ve spread risk not only on your main product but also on your infrastructure, it becomes hard.
I would be vary of a smallish company building their own Jira in house in a similar way.
I'm starting to wonder though whether companies even have the in-house competence to compare the options and price this risk correctly.
>Now a good company would concentrate risk on their differentiating factor or the specific part they have competitive advantage in.
Yes, but one differentiating factor is always price and you don't want to lose all your margins to some infrastructure provider.
Software companies have higher margins so these decisions are lower stakes. Unless on premises helps the bottom line of the main product that the company provides, these decisions don't really matter in my opinion.
Think of a ~5000 employee startup. Two scenarios:
1. if they win the market, they capture something like ~60% margin
2. if that doesn't happen, they just lose, VC fund runs out and then they leave
In this dynamic, costs associated with infrastructure don't change the bottomline of profitability. The risk involved with rolling out their on infrastructure can hurt their main product's existence itself.
1 reply →
Precious real-world engineering skills also play a role.
But most importantly, the attractive power that companies doing on-premise infrastructure have towards the best talent.
Yes, the idea is that you focus on the things that differentiate you from the competition. If you’re a factory that makes nails, a better data centre won’t make you any more money. It won’t help you sell more nails. So you should leave the data centres to the experts, and focus on work which improves your actual product.
If you don’t, you’ll be stuck trying to figure out data centres. Hiring tons of infrastructure experts, trying to manage power consumption. And for what? You won’t sell any more nails.
If you’re a company like Google, having better data centres does relate to your products, so it makes sense to focus on them and build your own.
It’s also opex vs capex, which is a battle opex wins most of the time.
Opex is faster. Login, click, SSH, get a tea.
Capex needs work. A couple of years, at least.
If you are willing to put in the work. Your mundane computer is always better than the shiny one you don't own.
4 replies →
Well, capex has a multi-year depreciation schedule and has to cover interest rates. So the simplified "opex wins most of the time" is right.
But we are talking about a cost difference of tens of times, maybe a few hundred. The cloud is not like "most of the time".
It depends. Grant funding (e.g. in academia) makes capex easier to manage than opex (because when the grant runs out you still have device).
I think it wins because opex is seen as stable recurring cost and capex is seen as the money you put in your primary differentiation for long term gains.
8 replies →
Note that they're running R630/R730s for storage. Those are 12-year old servers, and yet they say each one can do 20 Gbps (2.5 GBps) of random reads. In comparison, the same generation of hardware at AWS ({c,m,r}4) instance maxes out at 50% of that for EBS throughput on m4, and 70% on r4 - and that assumes carefully tuned block sizes.
Old hardware is _plenty_ powerful for a lot of tasks today.
I’m on a project at work replacing our R430s and R730s. They’ve been absolute tanks with very few hardware failures. That said, my company chooses to have OEM support for replacing failed components and keeping firmware/bios/idrac updated. You can absolutely run these if you’re OK with 3rd party replacements or parting out spare machines. Some industries are more tolerant to this than others.
I ran 3x R620s 24/7/365 in my homelab for ~6 years (well, other than when I moved, or shut one down for a clean-and-inspect, or lost power in excess of what my UPS could handle... thanks, Texas). The only things that failed during that time were a couple of sticks of RAM, and a PSU.
The own vs rent calculus for compute is starting to mirror the market value vs replacement cost divergence we see in physical assets. Cloud is convenient because it lowers OpEx initially, but you lose control over the long-term CapEx efficiency. Once you reach a certain scale, paying the premium for AWS flexibility stops making sense compared to the raw horsepower of owned metal.
Using "big" cloud providers is often a mistake. You want to use rented assets to bootstrap and then start deploying on instances that are more and more under your control. With big cloud providers, it is easy to just succumb to their service offerings rather than do the right thing. Do your PoC on Hetzner and DigitalOcean then scale with purpose.
Datacenters need cool dry air? <45%
No, low isn't good perse. I worked in a datacenter which in winters had less than 40%, ram was failing all over the place. Low humidity causes static electricity.
The datacenter is in San Diego - a quick Google confirms that external humidity pretty much never drops below 50% there.
Things would be different in a colder climate where humidity goes --> 0% in the winter
Low humidity causes static electricity.
RAM that is plugged in and operating isn't subject to external ESD, unless you count lightning strikes. Where are you getting this?
Low is good if you are also adding more humidity back in. If you want to maintain 45-50% (guessing), then you would want <45% environmental humidity so that you can raise it to the level you want. You're right about avoiding static, but you'd still want to try to keep it somewhat consistent.
It is much cheaper to use external air for cooling if you can.
Yeah but the article makes it sound as if lower is better, which it is definitely not. And yeah you need to control humidity, that might mean sometimes lowering, and sometimes increase it by whatever solution you have.
Also this is where cutting corner indeed results in lower cost, which was the reason for the OP to begin with. It just means you won't get as good a datacenter as people who are actually tuning this whole day and have decades of experience.
The distinction between rent/own is kind of a false dichotomy. You never truly own your platform - you just "rent" it in a more distributed way that shields you from a single stress point. The tradeoff is that you have to manage more resources to take care of it, but you have much greater flexibility.
I have a feeling AI is going to be similar in the future. Sure, you can "rent" access to LLM's and have agents doing all your code. And in the future, it'll likely be as good as most engineers today. But the tradeoff is that you are effectively renting your labor from a single source instead of having a distributed workforce. I don't know what the long-term ramifications are here, if any, but I thought it was an interesting parallel.
I think this is how IBM is making tons of money on mainframes. A lot of what people are doing with cloud can be done on premises with the right levels of virtualization.
https://intellectia.ai/news/stock/ibm-mainframe-business-ach...
60% YoY growth is pretty excellent for an "outdated" technology.
For ML it makes sense, because you’re using so much compute that renting it is just burning money.
For most businesses, it’s a false economy. Hardware is cheap, but having proper redundancy and multiple sites isn’t. Having a 24/7 team available to respond to issues isn’t.
What happens if their data centre loses power? What if it burns down?
It would be interesting to hear their contingency plan for any kind of disaster (most commonly a fire) that hits their data center.
Yep, does anyone remember the OVH fire[1][2]?
[1] https://www.techradar.com/news/remember-the-ovhcloud-data-ce...
[2] https://blocksandfiles.com/wp-content/uploads/2023/03/ovhclo...
I fully lost three small VPS there, and their response was poor: they didn't even refund time lost, they didn't compensate for time lost (e.g. a couple of months of free VPS), I got better updates from the news than from them (news were saying "almost total loss", while them were trying to convince me that I had the incredible bad luck that my three VPS were in the very small zone affected by the fire). The only way I had to recover what I lost was backups in local machines.
When someone point out how safe are cloud providers, as if they have multiple levels of redundancy and are fully protected against even an alien invasion, I remember the OVH fire.
1 reply →
contingency plan: Don't build your data center out of wood.
2 replies →
They use the datasenter for model training, not to serve online users. Presumably even if it will be offline for a week or even a month it will not be a total disaster as long as they have, for example, offsite tape backups.
Flooding due to burst frozen pipe, false sprinkler trigger, or many others.
Something very similar happened at work. Water valve monitoring wasn’t up yet. Fire didn’t respond because reasons. Huge amount of water flooded over a 3 day weekend. Total loss.
Theres only one solution to this problem and its 2 data centres in some way or form
What's the line from Contact?
why build one when you can have two at twice the price?
But, if you're building a datacenter for $5M, spending $10-15M for redundant datacenters (even with extra networking costs), would still be cheaper than their estimated $25M cloud costs.
1 reply →
the plan is to not set it on fire. If your office burns down you are already screwed
Ah Slurm, so good to see it still being used. As soon as I touched it in ~2010 I realized this was finally the solid queue management system we needed. Things like Sun Grid Engine or PBS were always such awful and burdensome PoS.
IIRC, Slurm came out of LLNL, and it finally made both usage and management of a cluster of nodes really easy and fun.
Compare Slurm to something like AWS Batch or Google Batch and just laugh at what the cloud has created...
I’m impressed that San Diego electrical power manages to be even more expensive than in the UK. That takes some doing.
Not nearly on the article's level, but I've been operating what I call a fog machine (itsy bitsy personal cloud) for about 15 years. It's just a bunch of local and off-site NAS boxes. It has kinda worked out great. Mostly Synology, but probably won't be when their scheduled retirement comes up. The networking is dead simple, the power use is distributed, and the size of it all is still a monster for me - back in the day, I had to use it for a very large audio project to keep backups of something like 750,000 albums and other audio recordings along with their metadata and assets.
This quote is gold:
The cloud requires expertise in company-specific APIs and billing systems. A data center requires knowledge of Watts, bits, and FLOPs. I know which one I rather think about.
> Having your own data center is cool
This company sounds more like a hobby interest than a business focused on solving genuine problems.
It kinda' does, doesn't it?
Re: the "hobby" part is where I agree with you the most. Where you say it's not solving genuine problems is where I differ the most.
It really feels to me like Comma is staffed by people who recognize that they never stopped enjoying playing with Lego -- their bricks just grew up, and they realized they can:
1) solve real-world problems
2) not be jerks about it
3) get paid to do it
Not everything has to be about optimizing for #3.
I'm a happy paying customer of Comma.ai (Comma four, baby!) -- their product is awesome, extremely consumer-friendly, and I hope they can grow in their success!
To me it sounds more like a return to vertical integration.
This is becoming increasingly common as far as I can tell.
There are benefits either direction, and I think that each company needs to evaluate the pros and cons themselves. Emotional pros/cons are something companies need to evaluate as employee morale can make or break a company. If the company is super technical in culture and they gain something intangible that is boosting the bottom line, having a datacenter as a "cool" factor is probably worth it.
I'd argue that it is in the long-term interest of any genuinely innovative company to attract intellectually curious talent with some coolness factor.
> Self-reliance is great, but there are other benefits to running your own compute. It inspires good engineering.
It's easy to inspire people when you have great engineers in the first place. That's a given at a place like comma.ai, but there are many companies out there where administering a datacenter is far beyond their core competencies.
I feel like skilled engineers have a hard time understanding the trade-offs from cloud companies. The same way that comma.ai employees likely don't have an in-house canteen, it can make sense to focus on what you are good at and outsource the rest.
> I feel like skilled engineers have a hard time understanding the trade-offs from cloud companies.
They spend too much time on yet another cloud native support group call, learning for ThatOneCloudProvider certificates, figuring out that single implementation caveats, standardizing security procedures between cloud teams, and so on.
Yet people in the article just throw a 1000 lines of code KV store mkv [0] on a huge raw storage server and call it a day. And it's a legit choice, they did actual study beforehand and concluded: we don't need redundancy in most cases. At all. I respect that.
[0] https://github.com/geohot/minikeyvalue
We actually do have an in-house chef lol.
If it were me, instead of writing all these bespoke services to replicate cloud functionality, I'd just buy oxide.computer systems.
You can also buy the hardware and hire an IT vendor to rack and help manage it as smart hands so you never need to visit the datacenter. With modern beefy hardware, even large web services only need a few racks so most orgs don't even to manage a large footprint.
Sure you have to schedule your own hardware repairs or updates but it also means you don't need to wrangle with the ridiculous cost-engineering, reserved instances, cloud product support issues or API deprecations, proprietary configuration languages, etc.
Bare metal is better for a lot of non-cost reasons too, as the article notes it's just easier/better to reason about the lower level primitives and you get more reliable and repeatable performance.
That’s called managed servers or managed services.
I have run bare metal and manage services you just have to be clear on what you have coverage for when disaster strikes or be willing to proactively replace hard drives before they die.
This is not a valid reason for running your own datacenter, or running your own server.
This is not a valid reason for running your own datacenter, or running your own server.
This is not a valid reason for owning a datacenter, or running your own server.
This is one of only two valid reasons for owning a datacenter, and one of several valid reasons for running your own server.
The only two valid reasons to build/operate a datacenter: 1) what you're doing is so costly that building your own factory is the only profitable way for your business to produce its widgets, 2) you can't find a datacenter with the location or capacity you need and there is no other way to serve your business needs.
There's many valid reasons to run your own servers (colo), although most people will not run into them in a business setting.
This also depends so much on your scaling needs. If you need 3 mid-sized ECS/EC2 instances, a load balancer, and a database with backups, renting those from AWS isn’t going to be significantly more expensive for a decent-sized company than hiring someone to manage a cluster for you and dealing with all the overhead of keeping it maintained and secure.
If you’re at the scale of hundreds of instances, that math changes significantly.
And a lot of it depends on what type of business you have and what percent of your budget hosting accounts for.
I also thinks it’s risk model too. Every time I see these kind of posts I think it misses the point there is a balance not only on cost like you describe but risk as well. You are paying to offload some of the risk from yourself.
> You are paying to offload some of the risk from yourself.
The opposite is also true: one is risking being banned by exascalers.
The issue is that they have already paid off their datacenter 5x over compared to cloud. For offline, batch training, I don't ses how any amount of risk could offset the savings.
1 reply →
> The cloud requires expertise in company-specific APIs and billing systems.
This is one reason I hate dealing with AWS. It feels like a waste of time in some ways. Like learning a fly-by-night javascript library - maybe I'm better off spending that time writing the functionality on my own, to increase my knowledge and familiarity?
Naive comment from a hobbyist with nothing close to $5M: I'm curious about the degree to which you build a "home lab" equivalent. I mean if "scaling" turned out to be just adding another Raspberry Pi to the rack (where is Mr. Geerling when you need him?) I could grow my mini-cloud month by month as spending money allowed.
(And it would be fun too.)
The degree is whatever you want to deal with. I had a rack at my last house (need to redesign the space for it at new house) with 3x Dell R620s in a Proxmox cluster, running K8s, serving Ceph from NVMe drives over Infiniband (for the mesh traffic), and 2x Supermicros running independent ZFS pools.
It was fun to build - especially Infiniband - but my next iteration is going to be a single beefy server, maybe with storage attached externally. What I had had outstanding uptime, but ultimately it was massively overkill, noisy, hot, and sucked power down.
You sure can. Pi are pretty underpowered you can get machines with more cores and memory and pcie lanes and networking out there and virtualize them
I paid 150€ for a Mini PC with an Intel N100, 16 GB of DDR5 memory, and a 500 GB SSD.
While I have no intention to scale up low spec hardware like this, it at least seems to beat the Azure VMs we use at work with "4 CPUs", which corresponds to two physical cores on an AMD EPYC CPU.
And that super slow machine I understand costs more than $100 per month, and that's without charges for disk space slower than the SSD, or network traffic.
Renting at Azure seems to be a terrible decision, particularly for desktop use.
Working at a non-tech regional bigco, where ofc cloud is the default, I see everyday how AWS costs get out of hand, it's a constant struggle just to keep costs flat. In our case, the reality is that NONE of our services require scalability, and the main upside of high uptime is nice primarily for my blood pressure.. we only really need uptime during business hours, nobody cares what happens at night when everybody is sleeping.
On the other hand, there's significant vendor lockin, complexity, etc. And I'm not really sure we actually end up with less people over time, headcount always expands over time, and there's always cool new projects like monitoring, observability, AI, etc.
My feeling is, if we rented 20-30 chunky machines and ran Linux on them, with k8s, we'd be 80% there. For specific things I'd still use AWS, like infinite S3 storage, or RDS instances for super-important data.
If I were to do a startup, I would almost certainly not base it off AWS (or other cloud), I'd do what I write above: run chunky servers on OVH (initially just 1-2), and use specific AWS services like S3 and RDS.
A bit unrelated to the above, but I'd also try to keep away from expensive SaaS like Jira, Slack, etc. I'd use the best self-hosted open source version, and be done with it. I'd try Gitea for git hosting, Mattermost for team chat, etc.
And actually, given the geo-political situation as an EU citizen, maybe I wouldn't even put my data on AWS at all and self-host that as well...
Same thing. I was previously spending 5-8K on DigitalOcean, supposedly a "budget" cloud. Then the company was sold, and I started a new company on entirely self-hosted hardware. Cloudflare tunnel + CC + microk8s made it trivial! And I spend close to nothing other than internet that I already am spending on. I do have solar power too.
I can see how this would work fine if the primary purpose is for training rather than serving large volumes of customer traffic in multiple regions
It would probably even make sense for some companies to still use cloud for their API but do the training on prem as that may be the expensive part.
what is the underling filesystem for your kv store, it doesn't appear to use raw devices.
> Cloud companies generally make onboarding very easy, and offboarding very difficult. If you are not vigilant you will sleepwalk into a situation of high cloud costs and no way out. If you want to control your own destiny, you must run your own compute.
Cost and lock-in are obvious factors, but "sovereignty" has also become a key factor in the sales cycle, at least in Europe.
Handing health data, Juvoly is happy to run AI work loads on premise.
On top of that, now when the US cloud act is again a weapon against EU, most European companies know better and are migrating in droves to colo, on-prem and EU clouds. Bye bye US hyperscalers!
The #1 reason I would advocate for using AWS today is the compliance package they bring to the party. No other cloud provider has anything remotely like Artifact. I can pull Amazon's PCI-DSS compliance documentation using an API call. If you have a heavily regulated business (or work with customers who do), AWS is hard to beat.
If you don't have any kind of serious compliance requirement, using Amazon is probably not ideal. I would say that Azure AD is ok too if you have to do Microsoft stuff, but I'd never host an actual VM on that cloud.
Compliance and "Microsoft stuff" covers a lot of real world businesses. Going on prem should only be done if it's actually going to make your life easier. If you have to replicate all of Azure AD or Route53, it might be better to just use the cloud offerings.
> The #1 reason I would advocate for using AWS today is the compliance package they bring to the party.
I was going to post the same comment.
Most of the people agreeing to foot the AWS bill do it because they see how much the compliance is worth to them.
I used to colocate a 2U server that I purchased with a local data center. It was a great learning experience for me. Im curious why a company wouldn't colocate their own hardware? Proximity isnt an issue when you can have the datacenter perform physical tasks. Bravo to the comma team regardless. It'll be a great learning experience and make each person on their team better.
Ps... bx cable instead of conduit for electrical looks cringe.
The main reason not to colocate is if you're somewhere with high real estate costs... E.g Hetzner managed servers competes on price w/co-location for me because I'm in London.
I colocate in London, a single server / firewall comes to around £5k a year. I also colocate two other servers in some northern UK location in some industrial estate for £2k as my backups. I've never enjoyed the cloud and dedicated server's have their own caveats too.
Budget hosts such as Hetzner/OVH have been known to suddenly pull the plug for no reason.
My kit is old, second hand old (Cisco UCS 220 M5, 2xDell somethings) and last night I just discovered I can throw in two NVIDIA T4's and turn it in to a personal LLM.
I'm quite excited having my own colocated server with basic LLM abilities. My own hardware with my own data and my own cables. Just need my own IP's now.
3 replies →
This was one of the coolest job ads that I've ever read :). Congrats for what you have done with your infrastructure, team and product!
Agreed!
Gives a whole new level to the idea of "full stack developer"
SSD's don't last longer than HDDs. Also they're much more expensive due to AI now. They should move to cutdown on power costs.
This is hackernews, do the math for the love of god.
There are good business and technical reasons to choose a public cloud.
There are good business and technical reasons to choose a private cloud.
There are good business and technical reasons to do something in-between or hybrid.
The endless "public cloud is a ripoff" or "private clouds are impossible" is just a circular discussion past each other. Saying to only use one or another is textbook cargo-culting.
The company I work for used to have a hybrid where 95% was on-prem, but became closer to 90% in the cloud when it became more expensive to do on-prem because of VMware licensing. There are alternatives to VMware, but not officially supported with our hardware configuration, so the switch requires changing all the hardware, which still drives it higher than the cloud. Almost everything we have is cloud agnostic, and for anything that requires resilience, it sits in two different providers.
Now the company is looking at doing further cost savings as the buildings rented for running on-prem are sitting mostly unused, but also the prices of buildings have gone up in recent years, notably too, so we're likely to be saving money moving into the cloud. This is likely to make the cloud transition permanent.
> Cloud companies generally make onboarding very easy, and offboarding very difficult.
I reckon most on-prem deployments have significantly worse offboarding than the cloud providers. As a cloud provider you can win business by having something for offboarding, but internally you'd never get buy-in to spend on a backup plan if you decide to move to the cloud.
> As a cloud provider you can win business by having something for offboarding, but internally you'd never get buy-in to spend on a backup plan if you decide to move to the cloud.
Its the other way around. How do you think all businesses moved to the cloud in the first place?
15-years ago or so a spreadsheet was floating around where you could enter server costs, compute power, etc and it would tell you when you would break-even by buying instead of going with AWS. I think it was leaked from Amazon because it was always three-years to break-even even as hardware changed over time.
Azure provides their own "Total Cost of Ownership" calculator for this purpose [0]. Notably, this makes you estimate peripheral costs such as cost of having a server administrator, electricity, etc.
[0] - https://azure-int.microsoft.com/en-us/pricing/tco/calculator...
I plugged in our own numbers (60 servers we own in a data centre we rent) and Microsoft thinks this costs us an order of magnitude more than it does.
Their "assumption" for hardware purchase prices seems way off compared to what we buy from Dell or HP.
It's interesting that the "IT labour" cost they estimate is $140k for DIY, and $120k for Azure.
Their saving is 5 times more than what we spend...
2 replies →
If you buy, maybe. Leasing or renting tends to be cheaper from day one. Tack on migration costs and ca. 6 months is a more realistic target. If the spreadsheet always said 3 years, it sounds like an intentional "leak".
Did the AWS part include the egress costs to extract your data from AWS, if you ever want to leave them?
AWS says they will waive all egress costs when exiting https://aws.amazon.com/blogs/aws/free-data-transfer-out-to-i...
1 reply →
Well, somebody should recreate it. I smell a potential startup idea somewhere. There's a ton of "cloud cost optimizers" software but most involve tweaking AWS knobs and taking a cut of the savings. A startup that could offload non critical service from AWS to colo and traditional bare metal hosting like Hetzner has a strong future.
One thing to keep in mind is that the curve for GPU depreciation (in the last 5 years at least) is a little steeper than 3 years. Current estimates is that the capital depreciation cost would plunge dramatically around the third year. For a top tier H100 depreciation kicks in around the 3rd year but they mentioned for the less capable ones like the A100 the depreciation is even worse.
https://www.silicondata.com/use-cases/h100-gpu-depreciation/
Now this is not factoring cost of labour. Labor at SF wages is dreadfully expensive, now if your data center is right across the border in Tijuana on the other hand..
I love articles like this and companies with this kind of openness. Mad respect to them for this article and for sharing software solutions!
There's the HN I know and love
> Maintaining a data center is much more about solving real-world challenges. The cloud requires expertise in company-specific APIs and billing systems. A data center requires knowledge of Watts, bits, and FLOPs. I know which one I rather think about.
I find this to be applicable on a smaller scale too! I'd rather setup and debug a beefy Linux VPS via SSH than fiddle with various propietary cloud APIs/interfaces. Doesn't go as low-level as Watts, bits and FLOPs but I still consider knowledge about Linux more valuable than knowing which Azure knobs to turn.
I'm thinking about doing a research project at my university looking into distributed "data centers" hosted by communities instead of centralized cloud providers.
The trick is in how to create mostly self-maintaining deployable/swappable data centers at low cost...
Goes for small business and individuals as well. Sure, there are times that cloud makes sense, but you can and should do a lot on your own hardware.
Realistically, it's the speed with which you can expand and contract. The cloud gives unbounded flexibility - not on the per-request scale or whatever, but on the per-project scale. To try things out with a bunch of EC2s or GCEs is cheap. You have it for a while and then you let it go. I say this as someone with terabytes of RAM in servers, and a cabinet I have in the Bay Area.
Cloud, in terms of "other company's infrastructure" always implies losing the competence to select, source and operate hardware. Treating hardware as commodity will eventually treat your very own business as commodity: Someone can just copy your software/IP and ruin your business. Every durable business needs some kind of intellectual property and human skills that are not replaceable easily. This sounds binary, but isn't. You can build long-lasting partnerships. German Mittelstand did that over decades.
I just read about Railway doing something similar, sadly their prices are still high compared to other bare metal providers and even VPS such as Hetzner with Dokploy, very similar feature set yet for the same 5 dollars you get way more CPU, storage and RAM.
https://blog.railway.com/p/launch-week-02-welcome
Their pricing page is so confusing: CPU: $0.00000772 per vCPU / sec
This seems to imply $40 / month for 2 vCPU which seems very high?
Or maybe they mean "used" CPU versus idle?
Billing per used or not idle cpu cycle would be quite interesting. Number of cores would just effectively be your cost cap. Efficiency would be even more important. And if the provider over subscribes cores you just pay less. Actually that's probably why they don't do it...
1 reply →
Hetzner bare metal ran much of crypto for many years before they cracked down on it.
Don't even have to go this far. Colocating in a couple regions will give you most of the logistical thrills at a fraction of the cost!
Heavy ML workloads make this more worthwhile since you get to design it to squeeze value out of every facet. For a basic web server and database it’s definitely overkill and something like a colocation makes much more sense
I love this article. Great write up. Gave me the same feeling when I would read about Stackoverflows handful of servers that ran all of the sites.
> San Diego power cost is over 40c/kWh, ~3x the global average. It’s a ripoff, and overpriced simply due to political dysfunction.
Mind anyone elaborate? Always thought this is was a direct cause of the free market. Not sure if by dysfunction the op means lack of intervention.
Electricity cost in California is generally more expensive than most other US states, except Hawaii. Not sure why.
Perhaps Comma needed the datacenter to be in San Diego for latency or other reasons, but if they need it mostly for compute, it would have been cheaper to operate their datacenter elsewhere... but if we keep going down that path, maybe it actually becomes cheaper to rent a cloud after all.
>Mind anyone elaborate? Always thought this is was a direct cause of the free market. Not sure if by dysfunction the op means lack of intervention.
The majority of Californians have no say and cannot choose their utilities provider. This is the polar opposite of the "free market".
Did you say “free market”? There is one provider. There is a lot of regulation, mostly incompetent. It’s a mess.
The cloud is a psyop, a scam. Except at the tiniest free-tier / near free-tier use cases, or true scale to zero setups.
I've helped a startup with 2.5M revenue reduce their cloud spend from close to 2M/yr to below 1M/yr. They could have reached 250k/yr renting bare-metal servers. Probably 100k/yr in colos by spending 250k once on hardware. They had the staff to do it but the CEO was too scared.
Cloud evangelism (is it advocacy now?) messed up the minds of swaths of software engineers. Suddenly costs didn't matter and scaling was the answer to poor designs. Sizing your resource requirements became a lost art, and getting into reaction mode became law.
Welcome to "move fast and get out of business", all enabled by cloud architecture blogs that recommend tight integration with vendor lock-in mechanisms.
Use the cloud to move fast, but stick to cloud-agnostic tooling so that it doesn't suck you in forever.
I've seen how much cloud vendors are willing to spend to get business. That's when you realize just how massive their margins are.
> The cloud is a psyop, a scam.
You're just young.
> Suddenly costs didn't matter and scaling was the answer to poor designs.
It did.
Did you know that cloud cost less than what the internal IT team at a company would charge you?
Let's say you worked on product A for a company and needed additional VM. Besides paperwork, the cost to you (for your cost center) would be more than using the company credit card for the cloud.
> Sizing your resource requirements became a lost art
In what way? We used to size for 2-4x since getting additional resources (for the in-house team) would be weeks to months. Same old - just cloud edition.
> You're just young.
And I feel great!
> Did you know that cloud cost less than what the internal IT team at a company would charge you?
Yes. Internal IT teams ran old-school are inefficient. And that's what the vendor tells you while they create shadow IT inside your company. Skip ITSM and ITIL... do it the SRE way.
Until the cloud economist (real role) comes in and finds a way to extract more rent out of their customer base (like GCP's upcoming doubling rates on CDN Interconnect). And until internal IT kills shadow IT and regains management of cloud deployments. Cybersecurity and stuff...
Back to square one. ITIL with cloud deployments. Some use cases will be way cheaper... but for your 100s of PBs of enterprise data, that's another story. And data gravity will kill many initiatives just based on bit movement costs.
> Besides paperwork, the cost to you (for your cost center) would be more than using the company credit card for the cloud.
To some extent. One is hard dollars the other is funny money. But I thought paying for cloud with the company credit card was a 2016 thing. Now it's paid through your internal IT cost center, with internal IT markup.
I've seen petabytes of data move to the cloud and then we couldn't perform some queries on it anymore as that store wouldn't support it, and we'd need to spend 7 figures to move to another cloud database to query it. And that's hard dollars.
Yes, during early cloud days it was lean and aimed at startups. Now it's aimed at enterprise, and for some reason lots of startups still think it's optimized for them. It's not and it hasn't been for a long time.
1 reply →
Even at the personal blog level, I'd argue it's worth it to run your own server (even if it's just an old PC in a closet). Gets you on the path to running a home lab.
Absolutely. I don't have a blog but run my own email, several game servers, Matrix instance, Nextcloud and other internal services on a retired gaming PC. The total cost of my cloud subscriptions is $0, and no one is snooping on me. It's a great setup when combined with Linux machines and GrapheneOS phones, completely private and free of Big Tech.
I cancelled my digital ocean server of almost a decade late last year and replaced it with a raspberry pi 3 that was doing nothing. We can do it, we should do it.
Microsoft made the TCO argument and won. Self-hosting is only an option if you can afford expensive SysOps/DevOps/WhateverWeAreCalledTheseDays to manage it.
So.... you're saying they must be understaffed and paying poverty range wages to afford the San Diego climate and still cut a profit? ;)
Look the bottom of that page:
An error occurred: API rate limit already exceeded for installation ID 73591946.
Error from https://giscus.app/
Fellow says one thing and uses another.
IT dinosaur here, who has run and engineered the entire spectrum over the course of my career.
Everything is a trade-off. Every tool has its purpose. There is no "right way" to build your infrastructure, only a right way for you.
In my subjective experience, the trade-offs are generally along these lines:
* Platform as a Service (Vercel, AWS Lambda, Azure Functions, basically anything where you give it your code and it "just works"): great for startups, orgs with minimal talent, and those with deep pockets for inevitable overruns. Maximum convenience means maximum cost. Excellent for weird customer one-offs you can bill for (and slap a 50% margin on top). Trade-off is that everything is abstracted away, making troubleshooting underlying infrastructure issues nigh impossible; also that people forget these things exist until the customer has long since stopped paying for them or a nasty bill arrives.
* Infrastructure as a Service (AWS, GCP, Azure, Vultr, etc; commonly called the "Public Cloud"): great for orgs with modest technical talent but limited budgets or infrastructure that's highly variable (scales up and down frequently). Also excellent for everything customer-facing, like load balancers, frontends, websites, you name it. If you can invoice someone else for it, putting it in here makes a lot of sense. Trade-off is that this isn't yours, it'll never be yours, you'll be renting it forever from someone else who charges you a pretty penny and can cut you off or raise prices anytime they like.
* Managed Service/Hosting Providers (e.g., ye olde Rackspace): you don't own the hardware, but you're also not paying the premium for infrastructure orchestrators. As close to bare metal as you can get without paying for actual servers. Excellent for short-term "testing" of PoCs before committing CapEx, or for modest infrastructure needs that aren't likely to change substantially enough to warrant a shift either on-prem or off to the cloud. You'll need more talent though, and you're ultimately still renting the illusion of sovereignty from someone else in perpetuity.
* Bare Metal, be it colocation or on-premises: you own it, you decide what to do with it, and nobody can stop you. The flip side is you have to bootstrap everything yourself, which can be a PITA depending on what you actually want - or what your stakeholders demand you offer. Running VMs? Easy-peasy. Bare metal K8s clusters? I mean, it can be done, but I'd personally rather chew glass than go without a managed control plane somewhere. CapEx is insane right now (thanks, AI!), but TCO is still measured in two to three years before you're saving more than you'd have spent on comparable infrastructure elsewhere, even with savings plans. Talent needs are highly variable - a generalist or two can get you 80% to basic AWS functionality with something like Nutanix or VCF (even with fancy stuff like DBaaS), but anything cutting edge is going to need more headcount than a comparable IaaS build. God help you if you opt for a Microsoft stack, as any on-prem savings are likely to evaporate at your next True-Up.
In my experience, companies have bought into the public cloud/IaaS because they thought it'd save them money versus the talent needed for on-prem; to be fair, back when every enterprise absolutely needed a network team and a DB team and a systems team and a datacenter team, this was technically correct. Nowadays, most organizational needs can be handled with a modest team of generalists or a highly competent generalist and one or two specialists for specific needs (e.g., a K8s engineer and a network engineer); modern software and operating systems make managing even huge orgs a comparable breeze, especially if you're running containers or appliances instead of bespoke VMs.
As more orgs like Comma or Basecamp look critically at their infrastructure needs versus their spend, or they seriously reflect on the limited sovereignty they have by outsourcing everything to US Tech companies, I expect workloads and infrastructure to become substantially more diversified than the current AWS/GCP/Azure trifecta.
> We use SSDs for reliability and speed.
Hey, how do SSDs fail lately? Do they ... vanish off the bus still? Or do they go into read only mode?
TLDR:
> In comma’s case I estimate we’ve spent ~5M on our data center, and we would have spent 25M+ had we done the same things in the cloud.
IMO, that's the biggie. It's enough to justify paying someone to run their datacenter. I wish there was a bit more detail to justify those assumptions, though.
That being said, if their needs grow by orders of magnitude, I'd anticipate that they would want to move their servers somewhere with cheaper electricity.
Interesting that they go for no redundancy
What redundancy are we talking about? AWS has proven to the world on multiple occasions that redundancy across geo locations is useless, because if us-east-1 is down, their whole cloud is done, causing a big chunk of the world to be down.
Half sarcasm of course, but it goes to show that the world is not going to fall apart in many cases when it comes to software. Sure, it's not ideal in lots of cases, but we'll survive without redundancy.
Is there a client to sell on your own unused private cloud?
Well, their comment section is fore sure not running on premises, but on the cloud:
"An error occurred: API rate limit already exceeded for installation ID 73591946."
This is a great solution for a very specific type of team but I think most companies with consistent GPU workloads will still just rent dedicated servers and call it a day.
I agree, and cloud compute is poised to become even more commoditized in the coming years (gazillion new data centers + AI plateauing + efficiency gains, the writing is on the wall). There’s no way this makes sense for most companies.
> AI plateauing
Ummm is that plateauing with us in the room?
The advantage of renting vs. owning is that you can always get the latest gen, and that brings you newer capabilities (i.e. fp8, fp4, etc) and cheaper prices for current_gen-1. But betting on something plateauing when all the signs point towards the exact opposite is not one of the bets i'd make.
1 reply →
It's the opposite. The more consistent your workload the more practical and cost-effective it is to go on-prem.
Cloud excels for bursty or unpredictable workloads where quickly scaling up and down can save you money.
Other benefits: easy access to reliable infrastructure and latest hardware which you can swap as you please. There are cases where it makes sense to navigate away from the big players (like dropbox going from aws to on-prem), but again you make this move when you want to optimize costs and are not worried about the trade-offs.
Not long ago Railway moved from GCP to their own infrastructure since it was very expensive for them. [0] Some go for a Oxide rack [1] for a full stack solution (both hardware and software) for intense GPU workloads, instead of building it themselves.
It's very expensive and only makes sense if you really need infrastructure sovereignty. It makes more sense if you're profitable in the tens of millions after raising hundreds of millions.
It also makes sense for governments (including those in the EU) which should think about this and have the compute in house and disconnected from the internet if they are serious about infrastructure sovereignty, rather than depending on US-based providers such as AWS.
[0] https://blog.railway.com/p/data-center-build-part-one
[1] https://oxide.computer/
I was under impression that Oxide rack does not currently ship with GPU's - at least with buildin. . Has this changed recently ?
Oxide racks don't yet have a GPU solution. But it is a good options for general compute and even with GPU required, general compute hasn't gone away.
The observation about incentives is underappreciated here. When your compute is fixed, engineers optimize code. When compute is a budget line, engineers optimize slide decks. That's not really a cloud vs on-prem argument, it's a psychology-of-engineering argument.
mark my words. cloud will fall out of fashion, but it will come back in fashion under another name in some amount of years. its cyclical.
In case anyone from comma.ai reads this: "CTO @ comma.ai" the link at the end is broken, it’s relative instead of absolute.
no because it's on premise you see? you don't need to access the world wide web, just their server
/s
I've just shifted to Hetzner, no regret
One thing I don't really understand here is why they're incurring the costs of having this physically in San Diego, rather than further afield with a full-time server tech essentially living on-prem, especially if their power numbers are correct. Is everyone being able to physically show up on site immediately that much better than a 24/7 pair of remote hands + occasional trips for more team members if needed?
Coolness factor of having a datacenter right in your office.
... and you can be one good earthquake away from insolvency.
Clouds suck. But so does “on premises”. Or co-location.
In the future, what you will need to remain competitive is computing at the edge. Only one company is truly poised to deliver on that at massive scale.
I like Hotz’s style: simply and straightforwardly attempting the difficult and complex. I always get the impression: “You don’t need to be too fancy or clever. You don’t need permission or credentials. You just need to go out and do the thing. What are you waiting for?”
This was written by Harald Schäfer, the CTO of comma.ai. I'm not so sure if G. Hotz is still involved in comma.ai.
Don't think he is, but it does seem like he inspired a hacker mentality in the shop during his tenure.
Ah I missed that.
if i understood correctly, you dont kubernetes, rights? Did you consider it?
And finally we reach the point where you're not shot for explaining if you invest in ownership after everything is over you have something left that has intrinsic value regardless of what you were doing with it.
Otherwise, well just like that gym membership, you get out what you put into it...
Am I the only one that is simply scared of running your own cloud? What happens if your administrator credentials get leaked? At least with Azure I can phone microsoft and initiate a recovery. Because of backups and soft deletion policies quite a lot is possible. I guess you can build in these failsafe scenarios locally too? But what if a fire happens like in South Korea? Sure most companies run more immediate risks such as going bankrupt, but at least Cloud relieves me from the stuff of nightmares.
Except now I have nightmares that the USA will enforce the patriot act and force Microsoft to hand over all their data in European data centers and then we have to migrate everything to a local cloud provider. Argh...
Do you have a computer at home? Are you scared of its credentials leaking? A server is just another computer with a good internet connection.
You can equip your server with a mouse, keyboard and screen and then it doesn't even need credentials. The credential is your physical access to the mouse and keyboard.
I mean people are nowadays are really scared of using microwave oven too. What happens if I heat my coffee 1 min too long. Could be near death experience. Thats why I always drive down to Starbucks for coffee!
1 reply →
Then literally own the cloud, like run the hardware on-prem yourself.
Looks insanely daunting imo
> In a future blog post I hope I can tell you about how we produce our own power and you should too.
Rackmounted fusion reactors, I hope. Would solve my homelab wattage issues too.
Having worked only with the cloud I really wonder if these companies don't use other software with subscriptions. Even though AWS is "expensive" its a just another line item compared to most companies overall SaaS spend. Most businesses don't need that much compute or data transfer in the grand scheme of things.
Stopped reading at "Our main storage arrays have no redundancy". This isn't a data center, it's a volatile AI memory bank.
You should have kept reading:
> Redundancy is not needed since no specific data is critical.
> we have a redundant mkv storage array to store all of our trained models and training metrics.
That's just called understanding your failure domains, and RTO/RPO needs.
This turns out to be a more and more important primitive for companies who are building their own models [1].
[1] https://si.inc/posts/the-heap/
Or better; write your software such that you can scale to tens of thousands of concurrent users on a single machine. This can really put the savings into perspective.
If you were to read TFA, it is about ML training workloads, not web servers
Well the article starts out with a suggestion that we should all get a data center... It's quite a jump to assume that everyone reading this article needs to train their own LLMs.
Chatgpt:
# don’t own the cloud, rent instead
the “build your own datacenter” story is fun (and comma’s setup is undeniably cool), but for most companies it’s a seductive trap: you’ll spend your rarest resource (engineer attention) on watts, humidity, failed disks, supply chains, and “why is this rack hot,” instead of on the product. comma can justify it because their workload is huge and steady, they’re willing to run non-redundant storage, and they’ve built custom GPU boxes and infra around a very specific ML pipeline. ([comma.ai blog][1])
## 1) capex is a tax on flexibility
a datacenter turns “compute” into a big up-front bet: hardware choices, networking choices, facility choices, and a depreciation schedule that does not care about your roadmap. cloud flips that: you pay for what you use, you can experiment cheaply, and you can stop spending the minute a strategy changes. the best feature of renting is that quitting is easy.
## 2) scaling isn’t a vibe, it’s a deadline
real businesses don’t scale smoothly. they spike. they get surprise customers. they do one insane training run. they run a migration. owning means you either overbuild “just in case” (idle metal), or you underbuild and miss the moment. renting means you can burst, use spot/preemptible for the ugly parts, and keep steady stuff on reserved/committed discounts.
## 3) reliability is more than “it’s up most days”
comma explicitly says they keep things simple and don’t need redundancy for ~99% uptime at their scale. ([comma.ai blog][1]) that’s a perfectly valid trade—if your business can tolerate it. many can’t. cloud providers sell multi-zone, multi-region, managed backups, managed databases, and boring compliance checklists because “five nines” isn’t achieved by a couple heroic engineers and a PID loop.
## 4) the hidden cost isn’t power, it’s people
comma spent ~$540k on power in 2025 and runs up to ~450kW, plus all the cooling and facility work. ([comma.ai blog][1]) but the larger, sneakier bill is: on-call load, hiring niche operators, hardware failures, spare parts, procurement, security, audits, vendor management, and the opportunity cost of your best engineers becoming part-time building managers. cloud is expensive, yes—because it bundles labor, expertise, and economies of scale you don’t have.
## 5) “vendor lock-in” is real, but self-lock-in is worse
cloud lock-in is usually optional: you choose proprietary managed services because they’re convenient. if you’re disciplined, you can keep escape hatches: containers, kubernetes, terraform, postgres, object storage abstractions, multi-region backups, and a tested migration plan. owning your datacenter is also lock-in—except the vendor is past you, and the contract is “we can never stop maintaining this.”
## the practical rule
*if you have massive, predictable, always-on utilization, and you want to become good at running infrastructure as a core competency, owning can win.* that’s basically comma’s case. ([comma.ai blog][1]) *otherwise, rent.* buy speed, buy optionality, and keep your team focused on the thing only your company can do.
if you want, tell me your rough workload shape (steady vs spiky, cpu vs gpu, latency needs, compliance), and i’ll give you a blunt “rent / colo / own” recommendation in 5 lines.
[1]: https://blog.comma.ai/datacenter/ "Owning a $5M data center - comma.ai blog"
capex vs opex the Opera.
[dead]
[dead]
[dead]
And now go do that in another region. Bam, savings gone. /s
What I mean is that I'm assuming the math here works because the primary purpose of the hardware is training models. You don't need 6 or 7 nines for that is what I'm imagining. But when you have customers across geography that use your app hosted on those servers pretty much 24/7 then you can't afford much downtime.
[dead]
[dead]