Last week Nvidia agreed to pay $12.93 billion for Hugging Face, a company running roughly $150 million in annualized revenue. That is about 86 times revenue, for a company that reportedly turned down a $500 million Nvidia investment at a $7 billion valuation last year because it did not want a single dominant shareholder. Nobody pays 86 times revenue for revenue unless they are buying something else: distribution, a way to make open models easier for enterprises to buy, and relief from its own dependence on a small number of enormous customers.
The most common question I get from clients is which models they should use. Up until now I have steered them to the major frontier players: OpenAI, Anthropic, and Google. They are safe, they are supported, there is a company accountable for them, and if the budget matters, an older version of a frontier model beats moving to open weights.
Very little corporate work actually requires the most advanced model available. My rough split is 90/10. About 90 percent of what companies want AI to do is routine work over their own documents and data, and an open model or last year's frontier model handles it. The other 10 percent is genuinely hard, and that is where the newest models earn their price and where your big bets belong. Most organizations are still struggling to execute the 90 percent, which makes a good deal of the argument below academic for them.
So the reason I have kept clients on the frontier for that 90 percent was never capability. It was procurement. Enterprise support for open weights already exists: Hugging Face has sold enterprise contracts with support and SLAs for years, and Bedrock serves managed open-weight models today. What has been missing is scale and credibility. A risk committee approves vendors, not files, and it wants an indemnity, a patch commitment, and a name that its third-party risk process already recognizes.
A risk committee approves vendors, not files.
Jurisdiction sat on top of that. The most capable open weights have been coming out of Chinese labs, and by August 2025 Chinese models had passed U.S. models in total downloads on Hugging Face. NIST evaluations cited in this year's congressional review found DeepSeek's open models more susceptible to cyber risk than comparable American ones. Some of the strongest open-weight models were also among the hardest for a U.S. regulated firm to approve.
What Nvidia will likely bring is an enterprise counterparty with enough scale, support, and procurement credibility to make open models routine. A Nasdaq-listed company with a balance sheet, a support organization, and a name on the contract. I expect Nvidia to start hosting open models itself: you pick a model, run it on Nvidia's machines, and pay Nvidia directly, with support, a record of where the model came from, and an invoice from a company your procurement team already knows. Buyers can assemble something like that today, but not from a single vendor most of them already have on the books. If that lands, open weights become a real option for the 90 percent instead of a theoretical one.
Why Nvidia Needs More Customers
The risk that matters most to Nvidia is its own customers. Most of its demand now comes from a small number of frontier labs and the hyperscalers financing them, a concentration Nvidia discloses in its own filings. Buyers that large negotiate hard, they buy from a second supplier to keep the first one honest, and when they get big enough they design their own chips. Google did. Amazon did. OpenAI and Anthropic have every reason to follow.
Ten thousand enterprises running open models on Nvidia infrastructure is a smaller sale each time and a far steadier business than a handful of companies placing the majority of orders. Buying the place where 18 million developers choose their models is an investment in having many customers instead of a few.
Nvidia tried selling cloud services directly and pulled back from it in 2025, because doing so put it in competition with the hyperscalers who are its largest customers. Owning Hugging Face lets it sell to developers without that problem.
How Good the Open Models Already Are
In 2023, Claude 2 solved less than 2 percent of the original SWE-bench, and that was the best result anyone had. Today DeepSeek V4 Flash reports 79 percent on SWE-bench Verified. The test itself, the software wrapped around the model, and the way answers are scored have all changed since 2023, so the two numbers are not directly comparable. The direction of travel is not ambiguous.
Epoch AI puts the current gap between the best open models and the frontier at an average of four months, roughly eight points on their capabilities index. Apply that to what companies actually run: document classification, extraction from scanned PDFs, retrieval over an internal policy corpus. That is the 90 percent. On narrow work like that, those eight points may be worth very little, and the outcome is decided instead by speed, cost, where the data sits, and whether the whole process can be audited. GLM 5.2 scores 51 on the Artificial Analysis index and Nvidia's own Nemotron 3 Ultra scores 48. The point is even the "mediocre" models are pretty good.
What This Means for OpenAI, Anthropic, and Google
If a model four months behind is good enough for most work, and a Nasdaq-listed American company distributes it on infrastructure a bank already has an MSA with, charging premium prices per token gets very hard. OpenAI, Anthropic, and Google have understood this, which is why all three moved toward agents, coding products, and subscriptions. What survives is the part nobody can copy out of a weights file: systems that act inside a company's real environment, memory built up over months of use, and a contract that puts the vendor on the hook for the result.
The Antitrust Question
Whether regulators allow this is an open question, and the deal has been agreed rather than closed. The FTC sued to block Nvidia's $40 billion purchase of Arm in 2021, arguing that Nvidia would control a neutral technology its rivals depend on, see their competitively sensitive information, and have reason to disadvantage them. Those objections apply here with very little translation. The 1948 Paramount decrees, which forced the film studios to sell the theaters they owned, are the older version of the same question, and courts have tended to rule on what an owner is capable of doing rather than on what it promises.
Jensen Huang says Hugging Face will stay open and that Nvidia compute will never be required there. Justin Boitano, Nvidia's vice president for enterprise computing, goes further and calls the platform a force for deconcentration. Both may be sincere. A promise held up by goodwill instead of by competition does not last as long.
What Will Change the Calculus
Four months is an average, and averages hide what decides whether a production system works: whether it stays right across many steps in a row. A model correct 95 percent of the time per step and one correct 98 percent look identical in a demo. Run twenty steps in a row, assuming each one is independent, and the first finishes the job about a third of the time while the second finishes about two thirds. As companies string agents together with more sophisticated orchestration, those eight index points stop being a rounding error, and the frontier models could well pull ahead again on exactly that work.
That is the 10 percent.
Anyone selling you a pure open-weight strategy is describing your current workload. Anyone selling you a pure frontier strategy is charging you frontier rates to classify invoices.
What decides the outcome in the meantime is which models are easy to use, and whoever owns the place where developers choose decides what is one click away and what takes six weeks of procurement. Nvidia spent $12.9 billion so that the question of which model to run keeps having thousands of viable answers, and so that more of those answers run on Nvidia.
Whether the frontier regains its advantage depends on how many organizations can actually make these systems work in practice. That is an organizational problem, not a model problem.
I welcome you to join me in this conversation.
An organizational problem, not a model problem.