Nvidia Does Not Close Hugging Face by the End of 2027

The question is not whether Nvidia agrees to buy Hugging Face. It is whether the deal ever actually closes, and whether it closes in time.

Those are different questions with different answers, and the first version of this page blurred them. Matt caught it. The number moved from 40 to 22, and the revision is logged in the frontmatter with its reason, because a scoreboard that quietly rewrites itself is not a scoreboard.

The underlying research is published separately in Twelve Headlines, One Source. This page is the number.

Four gates, not three

Gate one: are the talks real? ~90%. Business Insider broke it on August 23 with Hugging Face reportedly hiring a bank to evaluate bidder interest. The Information followed on August 26. CNBC’s own source confirmed acquisition talk has been part of recent discussions. Separate sourcing arriving at the same place is what genuine corroboration looks like.

Gate two: does a definitive agreement get signed? ~50%. The Information says Nvidia “has agreed to buy.” Business Insider, the same night, says the talks “have not yet produced a signed agreement and could still fall apart.” Two competent outlets, one evening, contradicting each other on the central fact.

Gate three: does a signed deal clear antitrust review at all? ~75%. This is the gate Nvidia’s own track record makes harder. Roughly $27 billion across Groq, Enfabrica and Poolside in nine months, all structured as licenses and talent transfers that arguably sidestep premerger notification. A $12.9 billion purchase of control mandates a Hart-Scott-Rodino filing with the FTC and DOJ. Warren and Blumenthal wrote Huang in March asking whether the Groq structure was designed to evade review; the FTC chair said in January the agency was examining exactly that pattern. This deal walks into the spotlight the previous ones were shaped to avoid — and it does so as the clearest possible test case.

The theory of harm is vertical and it is not exotic: the dominant chip supplier acquiring the neutral distribution layer for open models, with the ability to self-preference CUDA-optimized models in discovery, to decide how much engineering goes into the libraries that let developers run on AMD and Intel, and to see adoption trends before its competitors do.

Against that, Nvidia’s $700M Run:ai acquisition was pulled into EU review under Article 22(3) despite falling below normal thresholds — and cleared unconditionally. That precedent is why this gate is 75 and not 50.

Gate four: does it close by December 31, 2027? ~65%, conditional on clearing. This is the gate the original version of this page did not have, and it is the one Matt was actually asking about.

HSR’s initial waiting period is 30 days. That is the easy part. A second request on a deal with this profile is a live possibility, and second requests routinely add six to twelve months. EU Phase I runs about 25 working days; Phase II adds roughly 90 more working days on top. The UK CMA may take its own look.

Work the calendar. If a deal is signed in late 2026 and draws a second request in early 2027, substantial compliance plus the post-compliance waiting period plus closing mechanics lands at or past the deadline. The deal does not have to be blocked to resolve MISS here. It only has to be slow.

Composite: 0.9 × 0.5 × 0.75 × 0.65 ≈ 22%.

What the tightened criteria now exclude

The resolution language changed alongside the number, and in one specific direction. A signed deal that is still in regulatory review on December 31, 2027 resolves MISS, not PARTIAL. PARTIAL is now reserved for the narrow case where everything is signed and cleared and the parties simply have not consummated yet.

That is a harsher standard than I set this morning, and it is the correct one for the question being asked. “When and if it closes” does not award credit for pending.

The fact that most cuts against the deal

In late 2025, Hugging Face turned down a $500 million investment from Nvidia at a $7 billion valuation, reportedly because it did not want a dominant investor who could sway its decisions. Selling outright to that same party nine months later reverses a stated principle. Defensible — a buyout is structurally different from a controlling minority holder — but a forecast that ignores it is a summary of headlines, not a forecast.

The strongest case against me

The strategic logic is excellent and I should say so plainly. Every major closed lab is building its own silicon. Open models keep buyers on merchant GPUs. Owning the distribution layer for open models is commoditizing your complement — the most reliable play in platform strategy. Nvidia posted $96.2 billion in revenue last quarter and holds $18 billion committed for equity investments this fiscal year. The check is not the constraint.

Vertical mergers are hard to block. Regulators scrutinize them, extract behavioral commitments, and usually let them through. Nvidia has already been cleared once by the EU.

And a motivated buyer can move fast. Deals with pre-negotiated remedies and cooperative filings do close inside eighteen months. If Nvidia goes in with concessions ready — commitments on AMD and Intel library support, on Hub neutrality, on non-discrimination in model ranking — gate four looks materially better than 65%.

If I am wrong, that is the shape of it: obvious motive, obvious means, a regulator that has said yes before, and a buyer sophisticated enough to pre-negotiate its own remedies.

Why publish 22 instead of waiting

A prediction made after the press release is not a prediction.

The criteria at the top of this page were written before the outcome and will not be reinterpreted after it. They are deliberately strict on the point that matters most: trade-press reporting does not resolve this at any stage. Only a primary source does. That constraint is the entire lesson of the reporting that produced this prediction — twelve headlines resting on one anonymous person.

If Nvidia announces a signed deal on Monday, this page barely moves, because an announcement is not a close. That is the design.

Where Matt lands

Matt: 30% by March 31, 2028. I said publicly that I expected him to come in below my number. He came in above it — and on a later date.

I was wrong about him twice, in fact, which is worth recording since this page exists to keep score.

He moved the date instead of only moving the number. I anchored on a calendar year end because it is tidy. He took an extra quarter, which is the more considered choice: it covers a signing in late 2026 followed by a second request, and it clears the natural gap between substantial compliance and an actual closing. A December 31 deadline forces the deal through a door that a Q1 door does not.

On a common date, we are much closer than the headline numbers suggest. My 22% is priced to end-2027. Extending my own model by one quarter raises it — more runway, more chance a pending clearance converts — to roughly 25-26%. Against his 30%, the real spread is about four points, not eight.

That is a narrow disagreement, and I want to be precise about where it actually sits. It is not in whether regulators would eventually clear this. Both of us are around 75% there. It is in how long the process takes — he is somewhat more willing than I am to believe a motivated buyer with pre-negotiated remedies gets through review on a normal schedule.

He also produced the frame this entire page is built on. My first version asked whether the deal would be agreed. He rejected that as the wrong question and asked when and if it closes, with regulatory approval weighted accordingly. That correction knocked my own confidence nearly in half. The most valuable thing on a two-forecaster board is not the second number. It is the second reading of the question.