NVIDIA 同意以 12.93 billion dollars 收購 Hugging Face，美國司法部展開反壟斷審查

NVIDIA will acquire the AI model hub Hugging Face for 12.93 billion dollars, a transaction that has triggered a US Justice Department antitrust review over fears the chip giant would gain control of what regulators call a "fundamental distribution point" for open AI models. NVIDIA has pledged to keep Hugging Face open and multi-vendor, but trust in that promise is thin.

The deal that reframes the whole open-model economy
For most of the last five years, the AI story was told as a competition between model builders: OpenAI, Google, Anthropic, Meta, DeepSeek, Mistral, and a long tail of open-weight labs. Underneath that contest, a quieter layer quietly became the industry's town square. Hugging Face is where researchers publish weights, where startups pull pretrained checkpoints, where benchmarks are hosted, and where the phrase "download the model" almost always begins. It is not a chip company, not a lab, and not a cloud. It is the shelf, the catalog, and the loading dock all at once.

So when NVIDIA — the company that already sits on the supply side of nearly every GPU that runs those models — says it will spend 12.93 billion dollars to own that shelf, the reaction is less "congratulations" and more "wait, how much of the stack does one company need?" This is vertical integration of a kind the software industry has not seen since the great platform monopolies of the early 2000s, except here the integration runs from silicon all the way up to the app store for intelligence.

Why the DOJ is paying attention
The Justice Department's concern is not merely size. It is structural. A merger between the dominant supplier of training and inference hardware and the dominant distribution hub for model weights creates a position where the same entity controls both the picks and the shovels and the only general store in town. Regulators have flagged the risk that NVIDIA could, even implicitly, steer traffic, prioritize certain models, or make life harder for competitors who also happen to be its customers.

NVIDIA has responded by pledging to keep Hugging Face open and multi-vendor, and the deal is expected to close in the first half of 2027. But pledges from a company with this much leverage tend to be read through a skeptical lens. The history of platform acquisitions is full of "we will keep it neutral" promises that aged poorly once incentives shifted.

The open-model ecosystem holds its breath
For the developers, startups, and researchers who treat Hugging Face as neutral infrastructure, the question is practical before it is philosophical. Will model downloads stay free? Will competitors' checkpoints get equal visibility? Will the leaderboards, the datasets, and the Spaces demos remain a level playing field? None of these are guaranteed by a pledge, and all of them matter to the thousands of small teams that built their workflows on the assumption that the catalog was nobody's product.

There is also a geopolitical edge. A meaningful share of the world's open-weight models now originate in China, and a US-owned distribution hub introduces new friction points for cross-border model sharing. That is precisely the kind of thing that accelerates the splintering of the open AI web into regional silos — the opposite of the borderless model zoo the community once took for granted.

The bigger picture: the gold rush consolidates
Step back and the pattern is hard to miss. The labs are burning capital on compute, the clouds are selling the compute, and the chipmaker is booking the revenue either way. Now the chipmaker is also buying the marketplace. If the model hub, the cloud, and the silicon all tilt toward one balance sheet, the "open" in open models starts to look like a marketing term rather than a property right. That is the frustrating part. The exciting part is that we are watching, in real time, the governance questions around AI concentration get decided by acquisition rather than by legislation — and the DOJ inquiry may be the only thing standing between today's open ecosystem and tomorrow's toll road.

Key data points
Deal value: 12.93 billion dollars, expected to close in the first half of 2027.
Regulator: US Department of Justice reviewing control of a "fundamental distribution point" for AI models.
NVIDIA's commitment: keep Hugging Face open and multi-vendor; no exclusivity.
Market structure: NVIDIA already dominates AI training and inference silicon; Hugging Face dominates open-weight model distribution.
Precedent risk: platform acquisitions historically erode neutrality after incentives shift.
Sources: aibriefing.dev, TechCrunch reporting via techbeat.co, The Decoder, threerouter