Comment by somenameforme
3 days ago
Another interesting potential market here will be 'LLM in a box'. All the hardware and other tooling in a prebuilt, but modular, package ready to go. Pay one up-front cost, get a system running [whatever open LLM] with a token rate of [x], optionally configured to be immediately ready for distributed usage. Basically the opposite of cloud stuff: no rent, no dependency, 100% guaranteed uptime, guaranteed security/privacy (at least subject to your own actions), and so on.
Palantir already offers a "turnkey AI datacenter", i.e. a rack with "NVIDIA Blackwell Ultra systems with eight NVIDIA Blackwell Ultra GPUs and NVIDIA Spectrum-X™ Ethernet networking for AI training and inference".
It is said that it comes with all hardware and software required to run inference or training with an open weights LLM.
The existence of this product, which competes with cloud-based offerings like those of OpenAI and Anthropic, is presumably the reason why the Palantir CEO criticized very harshly some time ago the business model of OpenAI/Anthropic.
While I doubt that the ethics of Palantir is any better than of OpenAI/Anthropic, in this particular case I have to agree with Alex Karp about "Sovereign AI", i.e. that only losers will make their business completely dependent on an external entity like OpenAI or Anthropic, who are certainly not trustworthy.
I'm not sure a data center run by ... Palantir of all organizations is what people have in mind when they worry about data sovereignty.
They are selling it, not running it.
It is just a dedicated computer system, which should be managed by its owner, like any other on-prem servers.
I doubt that it has a good price/performance ratio, but it is a solution for those who feel that they do not want to search, buy, assemble, install and configure every HW/SW component.
6 replies →
Fair but the idea of "running your LLM setup" at every "need" level and corresponding cost does make sense.
For a lot of people (and orgs I'd guess) who just go and buy ≈$20 per month plans (or more for teams), they might not even need a fraction of that cost or capability. A lot of them don't even need it for coding or graphics. Even the API access based pricing aren't great from these frontier US AI houses. The distribution of "LLM being" offered will also give rise to many open-router like offering but at the end point level - direct interfaces to the customers. Pick your vendor sort.
AI shouldn't become another "search means Google".
I mean you could opt for the exabox from tinygrad https://tinygrad.org/#tinybox
It comes in a full sized shipping container and costs around $10M but money has stopped being connected to reality now anyway with all the AI company valuations being floated around, so who cares about a few million here or there.
How is this different from buying a supermicro rack? Better support?
“100% guaranteed downtime when you least can afford it and the support tickets are your problem.”
We’ve a hybrid shop, including hosting our own ML infra, and we save a ton from cloud spend with local ML. Easily one million USD over past three years. But it’s not “free”, you are shifting a lot of labor into your plate.
And with that also gain institutional knowledge, skill up your workers and attract talent that wants to work on this stuff.
All boils down to short-term/long-term thinking.
This. People WANT to work on this stuff. And having skilled workers is a precious advantage.
1 reply →
> 100% guaranteed uptime
Disagree there but I think this is an interesting idea. We would need to find some more cost-efficient hardware to run it on than Nvidia GPUs.
It will come... all big hardware players (Intel, AMD, Broadcom) and dozens of startups (Tenstorrent, etc.) are working on it...
Exactly! As I've argued here on HN before, such an "LLM in a box" might end up being serviced/upgraded once or twice a year by a company very similar to the one servicing the coffee machine at the office. In contrast to databases, storage, etc. it doesn't matter much if the box breaks at some point – they'll just come by and replace it with a new one – and there's barely any software on the box to speak of, at least none that requires continuous development and feature upgrades, beyond rolling out security patches. This makes the business case drastically different from cloud and SaaS offerings, where most of the moat is in the software and the state maintenance (and the vendor lock-in of course). The LLM in a box is destined to become a commodity.
What makes that kinda complicated is that multi-user throughput of LLMs scale well but single-user performance often stays constant at low ends. If you could saturate e.g. 16 concurrent session-month of demand, you can just go buy 16 of 32GB GPUs and start charging monthly for inference. That could work if you had e.g. over thousand total employees with hundreds of devs eager to trying it out, but only if the company is also interested in a private inference experiment.
You're talking about multi-session vs. single-session throughput. A single user can easily leverage multiple sessions via e.g. subagent swarms, especially on a lower-end setup where any single session is going to be quite slow. Saturating utilization during off-hours is harder but potentially quite feasible by assigning lower priority, unattended tasks/inference loops.
so something like this? https://tinygrad.org/#tinybox
I see ads for this all the time.
[dead]