Miami Mike had three separate conversations about on premise AI inference last week, and then Wednesday night The Information reported that Nvidia agreed to buy Hugging Face for $12.9 billion. Those turned out to be the same conversation. One is a chip company paying a serious premium for the place open models live. The other is a room full of people trying to work out why the AI line item went up again.
One caveat before anything else. This is reported, not signed. CNBC confirmed acquisition talks with a source close to the process, and Business Insider reported the same night that no agreement had been executed yet. Neither company has said anything. Treat the number as directional and the direction as real.
Why Nvidia would pay $12.9 billion for a model repository
Hugging Face is where open weight models get published, versioned, and downloaded. It is the GitHub of the model layer. The Information put the price at $12.9 billion against annualized revenue reported around $150 million, which is roughly 80 times revenue. Nobody pays that multiple for the revenue.
They pay it for what happens after the download. A model you pull off Hugging Face has to run somewhere, on hardware you own or rent, and that hardware is overwhelmingly Nvidia. Fortune made this point cleanly: a healthy open ecosystem gives buyers an alternative to the closed labs while keeping the compute underneath tied to CUDA. Every closed lab that ships its own silicon is a customer Nvidia partially loses. Every enterprise that downloads Llama or Nemotron or Qwen and stands it up internally is a customer it keeps.
There is a second reason and it is less flattering. Nvidia has guaranteed a large volume of cloud capacity for customers as part of its credit support arrangements. If those customers do not lease it out, Nvidia is holding empty racks. Owning the platform where millions of developers deploy models gives it somewhere to put that capacity.
Why the price of AI keeps rising while token prices collapse
Finance teams get stuck on this one, and it is why the Hugging Face deal matters even to a company that will never buy a GPU.
Per token prices have been in freefall. Blended enterprise pricing dropped roughly 67 percent year over year between early 2025 and early 2026. Over a longer window the collapse is closer to two orders of magnitude. And yet the FinOps Foundation’s 2026 State of FinOps report, drawn from more than a thousand practitioners, found that 73 percent of organizations blew past their original AI cost projections. The share of FinOps teams managing AI spend went from 31 percent in 2024 to 98 percent this year.
Volume ate the price cut. Agents do not ask one question and stop. They plan, call a tool, read the result, revise, call another tool, and every one of those hops resends the accumulated context because the model remembers nothing. EY put the cost of a single agentic customer service interaction at about $1.20 in 2026, up from roughly four cents in 2023. Same category of task. Thirty times the tokens.
I wrote the long version of this in the piece on AI inference costs, and the conclusion holds up better than I expected. Your unit price is fine. Your consumption curve is the problem.
Nvidia is betting that curve keeps bending upward and that a meaningful slice of it lands on infrastructure people own rather than rent. Buying the model repository is how you sit at the front of that.
Open weight adoption is not where the headlines say it is
Now the inconvenient number, because the tidy version of this story falls apart on contact with survey data.
Open models got dramatically better in 2025 and 2026. On coding benchmarks the best open weight releases now land within a couple of points of frontier systems. Adoption should have followed. It did not. Menlo Ventures put open source models at 11 percent of enterprise usage, down from 19 percent the year before. Capability went up. Share went down.
Some of that is a measurement artifact, since a survey of API traffic undercounts a model running quietly on a company’s own hardware. But not all of it. Most enterprises pulled back from open weights for reasons that have nothing to do with benchmarks: license terms that are not actually Apache, no clear provenance on training data, questions about where a Chinese model’s telemetry lands, and the plain fact that running a model in production requires people you may not have.
So if the deal closes, Nvidia is buying the most important asset in a category whose measured enterprise share is shrinking. That is either very smart or very early. My read is that it is a bet on the next five years rather than the last two, and that the thing standing between open weights and broad adoption is packaging, not quality. Nvidia is unusually good at packaging.
When on premise AI inference is the right call
Nothing about this deal changes the math on owning your own inference. It just makes the option more visible, and visibility is how bad purchases happen.
The full breakdown of what these platforms cost and what the vendors leave out of the quote is in my enterprise AI factory writeup. The compressed version is that on premise AI inference wins on a narrow band of workloads and loses badly everywhere else, and the deciding variable is utilization, not price per token. An owned box running at 30 percent is more expensive than the API you were complaining about.
Signals that on premise AI inference is worth evaluating:
- Your token volume is steady week to week rather than swinging 3x
- You can already say which workload consumes which share of your spend
- Your heaviest use cases are extraction, classification, summarization, or retrieval, where an open weight model is genuinely close enough
- Data residency is a real constraint in your industry, not a preference someone voiced in a meeting
- Monthly cloud AI spend is approaching 60 to 70 percent of what the equivalent owned hardware would cost amortized over the same period
If your AI program is under a year old, none of this applies to you yet. Optimize what you have first. Fixing model routing is free and a real multi model AI strategy will take more out of a quarterly bill than any hardware purchase does in a year. It also produces the usage data you need to size anything.
What I am watching
Two things. The first is whether Hugging Face stays neutral. Its value to the ecosystem is that it hosts everything from everyone, and the moment the model catalog starts steering toward the hardware the parent company sells, developers notice. Nvidia has been reasonably disciplined about this in other acquisitions.
The second is regulatory. A chip supplier with this much of the stack buying the distribution point for open models is the kind of deal that draws attention, and the review timeline may end up mattering more than the price.
If you want the practical takeaway before any of that resolves: go measure your token volume by workload for thirty days. That single exercise decides whether the on premise conversation is worth having at all, and it costs you nothing but attention.
[…] by the hour will eventually cost more than hardware you own, which is the whole argument in my on premise AI inference piece, and the deciding variable there is utilization. Brev actually acknowledges this better than […]
[…] Data that never leaves the building is nobody’s disclosure problem. I have written up when on premise AI inference actually pencils out, and the deciding variable is utilization, not the legal […]
[…] disclosure problem. Some agencies decide that is worth the trouble. I went through when on premise AI inference actually pencils out, and the deciding variable is utilization rather than the legal picture. The […]