Comment by kkielhofner

I'm building an early stage startup where the application requirements are constant GPU usage, 30TB of VERY fast NVMe storage, 512GB of RAM, LOTS of bandwidth transfer, and roughly 130TB of local storage (dataset for training). There is no combination of abstracted API-driven cloud services that can be pieced together to do what we need to do.

I didn't want to bother dealing with the advanced calculator but the AWS EC2 instance to handle this bespoke application would be $130k/yr with one year upfront and 12 month reservation:

https://calculator.aws/#/estimate?id=b3de412ba6a7748e9acd782...

This doesn't include bandwidth (which AWS has an insane markup on) or nearly enough storage.

Needless to say you can lease the hardware from $VENDOR and host it for years with an all you can eat 10gig port. Oh and the cost is absolutely fixed - no worries about getting that shocking AWS bill we all know of.

With a lease and hosting it's still an operating expense and with Section 179 it's actually preferable from a tax standpoint.

With datacenter remote hands and vendor support you're still not dealing with hardware - you don't have to even step foot in the hosting facility. Ever. If a hardware component fails the vendor dispatches someone and it gets handled (same day). That said hardware failures are (in my experience) exceedingly rare until you get to some significant scale where it's a numbers game.

Because ec2 is OS up anyway from a dev ops standpoint it's pretty much identical.

Let's say you keep it for three years and upgrade. The old hardware is still worth something so you can repurpose, sell, trade, whatever.

Rough math but let's call this approach 1/4 the cost (but likely much less).

At AWS markup you can hire a dedicated full time person to manage just one server and still come out ahead. Want five of them around the world? Now you're saving a TON of money vs different AWS regions.

Even with this when I tell people we deployed our own hardware they look at me like I'm crazy. We have an entire generation in tech from jr devs to the C suite to investors that are terrified of hardware and will pay anything to avoid it.

Because of this they don't even know what hardware, bandwidth, etc actually costs. It's assumed the cost is whatever $BIGCLOUD offers it at. I know this is an extreme case but it probably happens more than most would think.