DOJ: US antitrust is looking into Nvidia's acquisition of Hugging Face ($12.9B)

DOJ: US antitrust authority inquires into Nvidia's acquisition of Hugging Face ($12.9B)

🔎 The DOJ takes aim at open AI's control tower

On October 28, 2026, CNBC and Reuters published the same story within minutes of each other: the US Department of Justice (DOJ) is looking into Nvidia's acquisition of Hugging Face, a deal valued at $12.9 billion.

The question posed by the investigators boils down to a single sentence: would Nvidia take control of a "fundamental distribution point for AI"? That's the phrasing quoted by Reuters, and it says almost everything.

For a decade, the antitrust debate focused on the models themselves: GPT-5.5 at OpenAI, Gemini 3.1 Pro at Google, Claude Opus 4.7 at Anthropic. Today it is shifting to a layer that has stayed under the radar — distribution.

Because Hugging Face is not a lab. It is the tap through which most of open-weight AI flows: more than a million models, millions of developers. And that tap is about to fall into the hands of the supplier of 80 to 95% of the AI accelerators on the market.

My take: this is probably the most interesting antitrust case of 2026. Not because Nvidia would be "too big," but because it forces regulators to treat open-weight distribution as a market in its own right — something that has never been done so explicitly.

Key takeaways

• The DOJ is investigating Nvidia's acquisition of Hugging Face ($12.9B), as revealed by CNBC and Reuters on October 28, 2026.
• Central question: does control of the Hub — the world's leading platform for open-weight models — constitute a "fundamental AI distribution point"?
• The review goes beyond Hugging Face: a $17B investment in Groq, the acquisition of Run:ai (~$700M, 2024), a partnership with SchedMD (Slurm), and Nvidia's sales practices.
• France is examining the case in parallel through the Autorité de la concurrence (the French competition authority). Hugging Face was co-founded by three French nationals.
• Nvidia has committed to keeping Hugging Face "open and neutral." Regulators are showing clear skepticism.
• The deal is targeted to close in the first half of 2027. Nothing is blocked at this stage.

Recommended tools

What the DOJ Is Actually Looking For

Direct answer: the DOJ wants to determine whether the acquisition of Hugging Face would give Nvidia lock-in power over AI model distribution, on top of its overwhelming position in chips.

According to Reuters, citing sources close to the matter, investigators are examining one specific question: would Nvidia take control of a "fundamental AI distribution point"? The phrasing is anything but trivial. It suggests the DOJ views the Hub as essential infrastructure, not just a simple download website.

CNBC confirms the investigation and broadens its scope: it is part of a review of Nvidia's acquisitions, investments, and business practices. The Decoder adds an explicit dimension: sales practices — conditional discounts, exclusivity deals, and pressure exerted on customers who might consider buying elsewhere.

This is not a new discovery, either: as early as August 2024, Bloomberg revealed subpoenas issued to Nvidia in an antitrust investigation focused precisely on its business methods.

The Mechanics of a US Investigation

In the United States, any acquisition above a certain threshold must be notified under the Hart-Scott-Rodino Act. The regulator has 30 days to respond; beyond that, it can issue a "second request," a demand for information that extends the review by several months, sometimes more than a year.

A closing planned for the first half of 2027 — well beyond the standard timeline — betrays an in-depth review. The DOJ is not blocking anything for now: it is documenting. This is exactly how the Run:ai and SchedMD cases started in late 2024.

The Theory at Stake: Vertical Lock-In

A "vertical" acquisition — a supplier buying a distribution channel — is legally harder to attack than a merger between direct competitors. The classic question: will the new combined entity be able to close off or degrade its rivals' access (AMD, Google TPU, Groq, Cerebras) to that channel?

Concretely, the DOJ is seeking to determine whether Hugging Face could favor Nvidia's runtimes (NIM, TensorRT-LLM), penalize models optimized for other architectures, or give Nvidia privileged access to Hub data: which models are taking off, where, and for what use cases.

Why Hugging Face Has Become a "Fundamental Distribution Point"

Direct answer: because the Hub has become the default reflex for anyone downloading, comparing, or deploying an open weight model.

The numbers speak for themselves. The Hub crossed the one-million-model milestone in September 2024 and claims millions of active developers. And since the DeepSeek shock of January 2025 — when R1 proved that an open model could rival the closed heavyweights — open weight is no longer a niche movement: it's a pillar of the industry.

Proof in point, the current leaderboards:

The gap between the best closed model and the best open models is now measured in points, not categories. And almost all of these open models go through the Hub when they launch.

When Hugging Face releases Carbon-X, its open 14B model dedicated to genomics — which we broke down here — it's on the Hub that the community discovers it, evaluates it, and forks it. The Hub doesn't just host: it shapes discovery, through its trends, its leaderboards, and its verified organizations.

A Longstanding Nvidia-Hugging Face Relationship

A useful reminder for understanding the DOJ's concern: Nvidia and Hugging Face are no strangers to each other. Nvidia had already participated in Hugging Face's $235M Series D in 2023 (valuation: $4.5B). The two companies then collaborated on optimized inference, then on robotics with LeRobot — with Isaac GR00T 1.7, Isaac Teleop, and Cosmos 3 released as open source.

The $12.9B acquisition — the largest in Nvidia's history, which we covered at the announcement — thus transforms a partnership into ownership. It's that leap that changes the nature of the regulatory problem. Far ahead of Mellanox ($6.9B, 2019), previously the record.

The Combination That Raises Concerns: 90% of the Chips + The Faucet

Let's recall Nvidia's position. For the quarter alone ending in late October 2025, the company posted $57B in revenue, including $51.2B in data center. In October 2025, it became the first listed company to exceed $5 trillion in market capitalization. Analyst firms (Jon Peddie Research, TechInsights) estimate its share of AI accelerators at between 80 and 95%.

Add Hugging Face, and Nvidia would hold the factory (the chips), the fuel (CUDA), and the distribution channel (the Hub) all at once. It's this chain-of-control reading that motivates the investigation — and which Nvidia obviously contests.

Groq, Run:ai, SchedMD: an investigation far broader than Hugging Face

Direct answer: no, the DOJ is not looking at Hugging Face alone. The case is part of a comprehensive mapping of Nvidia's positions across the entire AI stack.

First front: Groq. Nvidia's $17B investment in Groq, the inference specialist, is among the items under examination. Inference is where the money flows once models are trained; a stake of this size in one of the sector's most high-profile challengers does not go unnoticed.

Second front: Run:ai. Nvidia had acquired this GPU cluster orchestration software for roughly $700M (announced in 2024). As early as December 2024, Reuters revealed that the DOJ had opened an in-depth review of the deal — a rare precedent for a transaction of this size, which ultimately closed in early 2025 after that scrutiny.

Third front: SchedMD. The maker of Slurm, the scheduler used by the majority of supercomputers, was also in the DOJ's sights in late 2024, again according to Reuters. The rationale: understanding how Nvidia secures control of the software that drives its own chips.

And beyond that: our analysis of the $40B plan Nvidia is devoting to AI in 2026 shows it — the company is investing massively in its own customer ecosystem, all the way up to the $100B promised to OpenAI in September 2025. Economists call these "circular deals": Nvidia funds its customers, who buy from Nvidia.

My take: Hugging Face is the most symbolic case, but not necessarily the most legally dangerous one for Nvidia. Cross-investments raise a harder question to dodge: what is a competitor for, if its main supplier both funds and distributes it?

"Hugging Face will stay open": Nvidia's commitments face skepticism

Direct answer: Nvidia has committed to keeping Hugging Face open and neutral. The track record of behavioral remedies justifies regulators' skepticism.

The commitments, as reported by the press: keeping the Hub free of charge, neutrality toward other chip manufacturers, no privileged access to the platform's data, unchanged licenses. On paper, it's clean.

The problem is that this type of commitment carries heavy historical baggage. Google got the green light for DoubleClick in 2007 without conditions; the complaints came afterward. Microsoft promised ten years of commitments to acquire Activision, including Call of Duty on Nintendo — a title that never materialized. The FTC lost its case against Meta/Within in 2022–2023. Behavioral remedies age poorly, because they assume permanent oversight and an acquirer acting in good faith.

The real risk, moreover, is not that Hugging Face closes its doors. It's that "open" gets redefined bit by bit: gated models by default, download quotas, prominent placement of Nvidia's NIMs, community licenses of shifting shape. None of these changes would violate the commitment to "stay open," and all of them would alter the market's balance.

My take: only a structural commitment would be credible — an irrevocable license on the Hub, or independent governance guaranteeing neutrality. A "good faith" commitment is worth exactly as much as the good faith of whoever signs it. The regulators, from Washington to Paris, seem to have understood this.

France too: the Autorité de la concurrence is examining the case

Direct answer: yes, Paris is looking at the same case at the same time as Washington — and that's no coincidence.

A quick biographical note: Hugging Face was co-founded in 2016 by three Frenchmen — Clément Delangue, Julien Chaumond, and Thomas Wolf. The company claims French roots and a massive European community. The Autorité de la concurrence therefore has both the jurisdiction and the motivation to take an interest.

Above all, it has experience with the Nvidia dossier: its investigators raided Nvidia's French offices in 2023, as part of a strand of the investigation concerning CUDA-related practices, which was revealed in 2024. The authority has known the firm for three years.

The stakes go beyond symbolism. For Europe, open AI is a lever of technological sovereignty: it's what allows players without billions of dollars' worth of GPUs to deploy competitive models, from DeepSeek models to self-hosted deployments. Control of a central distribution channel by an American player — one that also owns the chips — is a matter of industrial policy as much as of competition law.

If the deal is notified at the European level, the Commission could also take it up. Transatlantic regulators coordinate regularly; a case can very well move forward on both sides of the Atlantic in parallel.

What this concretely changes for developers and businesses

Direct answer: in the short term, nothing changes. In the medium term, three risks deserve your attention.

Risk #1: the privileged runtime. If Nvidia's NIM microservices become the "official" path on the Hub, deployments on AMD, TPU, or LPU could become second-class citizens — fewer optimizations, less documentation, less visibility.

Risk #2: telemetry. The Hub knows which models are downloaded, by whom, at what rate, for which use cases. In the hands of the chip leader, this data becomes a competitive advantage: Nvidia would know before anyone else where demand is heading.

Risk #3: bundling. Hugging Face credits backed by Nvidia compute credits, packaged enterprise offerings, "exclusive" integrations. Every practical bundle tightens the ecosystem a little more around a single vendor.

Three habits to preserve your independence:

• Diversify your access channels: OpenRouter to switch between closed and open models without changing your code, regional mirrors for the weights.
• Test self-hosting: Ollama for local use, vLLM for production. A quantized model like GLM-5.1 runs on a simple VPS — Hostinger offers VPS starting at around €5/month (October 2026, check hostinger.fr).
• Read the licenses before reading about them in the press. That's where the drifts play out.

Timeline: the milestones before the expected closing in the first half of 2027

Direct answer: three scenarios, one deadline, and a few milestones to watch.

Scenario 1 (the most likely): approval with strengthened commitments. The DOJ accepts the deal in exchange for verifiable guarantees — an irrevocable license, neutral governance, independent reporting.

Scenario 2: the DOJ takes legal action. Rare for a vertical acquisition, but the characterization of a "fundamental distribution point" is unprecedented. If the DOJ believes it has a doctrinal case to make, it may go to trial.

Scenario 3: Nvidia restructures. Divestiture of certain Hub assets, extended commitments, or a simple delay. Major acquirers often prefer to negotiate rather than face a trial with an uncertain outcome.

The milestones: the HSR notification (filed), the possible second request, the DOJ's position, the March 2027 GTC — the traditional showcase for Nvidia integrations — then the targeted closing in the first half of 2027. On the French side, the Autorité de la concurrence's review will follow its own timeline.

What to watch most closely: any license changes on the Hub, any announcement of an "exclusive" integration, and any changes to the terms of use. That's where commitments are kept — or unraveled.

❌ Common mistakes

Mistake 1: believing the investigation already blocks the acquisition

No. A DOJ investigation is not a veto. As long as the department has not taken legal action, the deal can still close in the first half of 2027. Solution: track the formal steps (second request, complaint, consent decree), not the rumors.

Mistake 2: confusing open weight and open source

Open weights are not automatically open source in the OSI sense: restrictive licenses, gated models, restricted commercial uses. Distribution control remains a lever even when the weights can be freely downloaded. Solution: read the licenses, not the slogans.

Mistake 3: thinking that "free" provides immunity from antitrust

Free of charge does not preclude market power — it's actually the classic pattern of platform markets. The DOJ isn't interested in the Hub's pricing, but in control of the channel. Solution: analyze the incentives, not the price tags.

Mistake 4: reducing the case to the United States

France is also examining the case, and Brussels may follow suit if the deal is notified at the European level. A European decision can impose commitments that are tougher than Washington's. Solution: keep an eye on both sides of the Atlantic.

❓ Frequently Asked Questions

Is Nvidia's acquisition of Hugging Face cancelled?

No. The DOJ is conducting an investigation; it has not (yet) taken action. The deal close remains targeted for the first half of 2027, subject to regulatory approvals. The scenario of a judicial block exists, but it remains a minority outcome for a vertical acquisition of this type.

Why is the DOJ interested in a free platform?

Because being free does not rule out market power. The Hub is the dominant distribution channel for open-weight models; its control by Nvidia, which also supplies 80 to 95% of AI accelerators, creates a risk of vertical lock-in across the entire stack.

Will Hugging Face remain free and open?

That is Nvidia's public commitment: free access maintained, neutrality toward other vendors, licenses unchanged. But behavioral commitments age poorly, and "open" can be redefined bit by bit. Watch the licenses and terms of use rather than the press releases — that's the only indicator that counts.

What is the link with open-weight models like DeepSeek V4 Pro or Kimi K2.6?

These models — DeepSeek V4 Pro (88), Kimi K2.6 (84), or GLM-5.1 (83) on the overall leaderboard — are distributed via the Hub. If Nvidia controls that channel, it influences the visibility, and therefore the commercial success, of the entire open-weight ecosystem.

When will the acquisition be finalized?

Closing is targeted for the first half of 2027, subject to American and European regulatory approvals. The DOJ investigation and the review by the competition authority may push that timeline back. Watch GTC in March 2027 — the first symbolic milestone.

✅ Conclusion

Nvidia's acquisition of Hugging Face turns open weight distribution into a leading antitrust issue — and the DOJ investigation, reported by CNBC, Reuters, and The Decoder, makes it the real-world test of this new doctrine.

To place this case within Nvidia's overall strategy, start with our analysis of the $13 billion acquisition and its plan for $40 billion in investments in 2026. And keep one date in mind: the first half of 2027.