Back of the envelope calculation (could be off, correct me if I am)
If you are a large enough company that spends million+ on inference a month, it makes sense to buy a GB300 rack ($6M on top range from what I could find) which has 20.7 TB. Since the model is mixed trained (MXFP4), you would need less than 10% of the rack's memory to serve the full model. Aggregate HBM bandwidth: 576 TB/s. You can run over 6000 parallel agentic workflows (each with ~100k context on average) at ~30 tok/s.
Assuming the annual amortization+electricity at $1.5M/year and about 50% average annual utilization, you get less than 60 cents (USD) per million output token, for a frontier model with plenty of capacity to share, all your data never leaving premises and well over an order of magnitude cheaper!
As long as a company believes that the openweight models will continue to get more capable and 'AI is here to stay', this model provides the first solid footing for a decision to just buy a rack.
We're running Kimi 2.8 on a $107k server and getting around 50k tokens per second or better on most things.
We've already saved money compared to last years token cost on Claude/Gemini
> We're running Kimi 2.8 on a $107k server
Equipped with what? Is it CPU based inference, a mix...?
Not GP, but my educated guess is that they are running a system with between 4 and 6 MI325 or MI355x or similar AMD GPUs. With the 50k tps as the total figure for all parallel requests. Those cards have a lot of memory for their price, allowing you to push to really high batch sizes while still having a large context size for each request
What is your config?
how many requests per second can the server take?
Are you developing software? Is most of it used on a coding agent? (Like Claude Code or ChatGPT Codex?) If so, what coding agent do you use? If you're not developing software what do you use it for (roughly)?
And hire 2 or 3 dev ops to keep it running ?
That another 400 to 700k.
It becomes your problem and not someone else’s. However, I don’t trust hosted LLMs for anything that needs to be private.
Where do I sign up to get 200k/yr to keep one rack running? Sounds like an incredibly chill job
Apparently it’s going to take the 3 of us to do this, mate. Going to get so much reading done.
What you get is not what you cost.
40% overhead is quite typical, so you'd be looking at $120k/year. In the USA I'd consider that a competitive salary for an admin capable or keeping a $6M rack of specialized hardware running 24/7.
Yeah but you don't need two such people, or even one, dedicated to this single rack.
A company of the size that this is worthwhile for, probably has dedicated devops on staff already and can add this rack to the inventory with no additional staff.
Who is going to upgrade the models ?
Who is going to fix it when the api does something weird ?
Who is going to proactively make sure it’s not overheating?
Chat GPT has enterprise contracts for a reason.
The same person managing the company's email accounts and whatnot.
I didn't say it's fire-and-forget. I'm saying all that is maybe a day of work every 3 months.
It doesn’t really work like that.
The companies which have the will and the budget to host their own LLMs typically require a ton of other, much smaller, models as well. They have internal security requirements, guardrails, audit, critical workflows start depending on your onprem setup, downtime is now something that’s not even allowed. There’s going to be a zoo of tooling, lots of bespoke work with internal clients who have no clue about docker, but now want their vibe-coded app to access a model they downloaded yesterday, and this model better be served and monitored 24/7 because now C-levels use it.
No one is going to budget several millions to then look at an email admin who hears ‘cuda’ for the first time in their life and ask them to just support the entire thing somehow.
And god forbid it’s an AMD setup.
This obviously isn’t relevant for a 10-person startup and their second-hand xeon with a single H100.
From my experience only the companies which are REALLY interested in privacy and data security bother with hosting their own models
My company is not even the same ballpark, but even we already have people for it. And doesn't include the fact that you can get a colo location and just a MSP or a contractor to do it for you
All of the answers to your questions above are in gp's comment already:
> A company of the size that this is worthwhile for, probably has dedicated devops on staff already
> However, I don’t trust hosted LLMs for anything that needs to be private.
Why not? Do you trust AWS with things that need to be private?
More than I trust frontier labs. AWS doesn't need to recoup 9 digits USD of capex
So you can just use Bedrock?
As mentioned, it is a "large enough company" already; they have full time sysadmins running things.
Adding another rack beside the VMWare cluster, managing any storage/networking issues, etc. will be incremental costs; they already have a pager (probably not a pager not anymore just an app on their phone) like rotation schedule etc.
There will be cloud/SaaS vendors who have lower cost of labor/capital due to automation and financing terms.
Having these models in the open caps the inference margin.
> I don’t trust hosted LLMs for anything that needs to be private
I'd update this to
'I don’t LLMs for anything that needs to be private'
What's to prevent the LLM from sliding a heavily obfuscated binary blob into the application that does nefarious things? If you aren't creating the LLM itself from scratch, I don't feel it can be trusted.
Why do you feel that creating the LLM from scratch is sufficient to trust it? Are you suggesting that you personally would read all 15 trillion tokens (plus every single agentic trade used in RL, along with its relative advantage in the batch) and personally guarantee that gradient descent would train a model which would not exfiltrate your corporate data?
Or that perhaps you have a perfect alignment algorithm which you are unwilling to share with the broader research community (evil)?
> What's to prevent the LLM from
A NN per se is a file... The executable that runs it can "act"...
> And hire 2 or 3 dev ops to keep it running
Not a devops but I'd say one full time is already too many.
Yes but zero is not enough and where do you get a fraction of a competent dev op from?
From the other dev-op work you’re doing.
Now you're thinking like management!
Understood but sharing your existing devops resources with this will soon become a bottleneck especially when any major downtime will keep several engineers (and long-running agents) blocked from any meaningful work until availability improves.
Not my experience, from an SMB that maintains its own hardware and services. You have a certain contingent of competent engineers who distribute their work across projects, and it generally works out fine. Or course you plan with some redundancy and fall-back plans in your systems.
The top poster mentioned LLM spend of millions/month to justify the estimated capex of $6m to self-host Kimi on own infra.
Add to this number another $1.5m/yr in opex, so not sure I’d call such an enterprise wealthy enough to spend those kinds of sums on LLMs an “SMB”.
Just like that new jobs created by AI! Localized model maintainer/technician.
You’ll slap some training on existing technologists/infra/sysadmin folks and perhaps have a support contract for the edge cases (hardware troubleshooting and advanced replacement).
(managed an entire data center building with thousands of servers a lifetime ago with ~2-3 other people, it’s only gotten easier over the last two decades imho)
Why would a singe system require 3 full-time dev ops?
Just let it manage itself, what could go wrong! :)
You also need a place to put it. With liquid cooling and extremely dense power capability. Your typical colo or closet server room isn't going to cut it.
I think the licensing that would likely apply to a company that's able to afford ~$6M rack and the associated infrastructure muddies this somewhat
I think internal use is allowed at any scale in the license?
> 4. The requirements set forth in Sections 2 and 3 do not apply to: (a) internal use of the Software, defined as any use that does not make the Software, its outputs, or its underlying capabilities available to third parties; [...]
It's very hard to make sure none of the outputs are ever made available to third parties. Source code can end up widely distributed (e.g. client-side js, open source). Prose will frequently get shared across organization boundaries (e.g. emails, websites, documents).
License for LLMs already have very little strength (it's not clear at all that they have any legal basis whatsoever) an AI lab is very, very unlikely to sue you for violating this kind of license clause for the reason you mentioned. It's clearly worded in a way to deter people from running a third party inference business out of it, they would have been much more explicit if they wanted to deter people from running it for internal use.
As far as I can tell, license fees are only applicable if you have more than 20m USD/year revenue from services provided using the model, or serve more than 100m users.