Nvidia is buying Hugging Face for something around $13bn, and the deal will probably clear.
It should clear on the current tests, and that is the thing worth arguing about. Nvidia does not sell a model hub. There is no horizontal overlap, no market share to add up, no obvious foreclosure. Assessed the way acquisitions are normally assessed, this is unremarkable.
The tests are the problem.
The harm is real and it is not exclusion
Nobody expects Nvidia to remove competitors' hardware from the platform. It would be commercially foolish and legally provocative, and it is not necessary.
A distribution layer governs behaviour through defaults. Which format is first-class. Which runtime the quickstart assumes. Which benchmark table appears without being asked for. Which conversions happen automatically and which require somebody to volunteer. Each of those is a small convenience, none of them excludes anybody, and together they determine what a large majority of developers actually ship.
An owner who controls the defaults does not need a policy. The ecosystem tilts without a single act that a regulator could point at, and by the time the tilt is measurable it is also the status quo.
We solved this elsewhere and forgot we did
Other industries reached the same problem and answered it structurally rather than through conduct rules.
Clearing houses sit between counterparties and are governed so that no participant controls them. Standards bodies write specifications their members implement, under constitutions that exist precisely to stop a member capturing the specification. Domain registries hold a namespace everyone needs and are separated from the registrars who sell into it. Payment networks have spent decades under governance arrangements about exactly this.
None of that came from a natural preference for committees. It came from the recognition that some layers are worth more to everybody when they are worth less to anybody in particular, and that ownership by a participant destroys the property that made the layer valuable.
The AI stack has arrived at the same structure and has not built any of the same institutions.
The counter-argument, taken seriously
The strongest defence is that neutrality was never load-bearing here and the value is in the engineering, which will now be better resourced. Model-to-hardware optimisation is genuinely difficult, it costs real money, and an owner with an interest in solving it may solve it faster than a neutral party without the incentive.
That may well be right in the near term, and this column is not predicting degradation. Nvidia has every reason to run the platform well.
But "the current owner has good intentions" is the argument that precedes every structural problem in this category, and it is unfalsifiable until it is too late to act on. The question is not whether this owner behaves. It is whether a layer this many people depend on should sit somewhere that requires anyone to behave.
What should have happened, and still could
Not a block. The remedy that fits the harm is governance rather than prohibition.
Commitments with a duration and a monitor: parity of first-class support across hardware vendors, default publication of benchmarks across accelerators, and no preferential placement in defaults or documentation — enforceable, reviewable, and not dependent on continued goodwill. Those are the sorts of conditions that get attached to vertical deals in other sectors as a matter of routine.
They would cost the acquirer very little if the engineering case is the real one. That is precisely why they are worth asking for, and why the response to being asked would be informative.
The part that should worry people beyond this deal
This paper has written that enterprises diversified their model suppliers and left the layer underneath untouched, and that the API economy is consolidating into priced infrastructure. The pattern is consistent: the visible market stays competitive and the substrate consolidates.
Competition law is very good at the visible market. It has almost nothing to say about a substrate, because a substrate is not a market until somebody has already captured it.
That is the gap this deal sits in, and it will not be the last one to.
Nvidia's announced agreement to acquire the Hugging Face platform in a deal reported at around $12.9 to $13bn, subject to approvals, is as reported in industry coverage on 9 September 2026, and is described in this publication's accompanying news report. This is an opinion column. No commitments regarding the platform's future operation have been made public, and none is assumed here either way. The comparisons with clearing houses, standards bodies and domain registries, and the argument for governance separation, are the author's own.
Topics opinioncompetitionacquisitionsopen sourceinfrastructure
