Nvidia Buys Hugging Face and Promises Not to Change It

For most of the last decade, the safest thing you could say about Hugging Face was that it belonged to nobody in particular. That is no longer true. On September 3, Nvidia confirmed it will acquire the open-model platform for $12.93 billion, and the entire question now is whether a neutral hub can stay neutral once the world's largest chip company owns it.

What Happened

The announcement arrived as a short post under Jensen Huang's byline on Nvidia's blog, dated September 3, 2026. The number in it was oddly precise: $12,930,300,000. Nvidia had entered the definitive agreement a day earlier, on September 2, and the news ended several weeks of reporting that had the two companies circling each other since late August. Huang framed it not as an acquisition of a competitor or a supplier, but as taking custody of something he described as a home for the open model developer community.

The scale Nvidia is buying is genuinely unusual for a company of Hugging Face's size. The platform hosts more than three million models, roughly 500,000 datasets, and about a million applications. More than 18 million developers, researchers and creators use it, and Nvidia says over 200,000 companies rely on it to discover, evaluate, customize and deploy AI. There is no comparable single location for open weights anywhere else in the industry.

The deal structure splits roughly $11.9 billion to Hugging Face shareholders and about $1 billion in equity earmarked to retain employees joining Nvidia. Hugging Face was founded in 2016 and had raised somewhere north of $395 million, with its last round in 2023 — a $235 million raise led by Salesforce Ventures that also drew in Google, Amazon, IBM and, notably, Nvidia itself. Recent reporting put the company's annualized revenue near $150 million, with its CEO telling TechCrunch in July that growth was pushing it close to profitability. The Financial Times has reported that Hugging Face turned down a roughly $500 million offer from Nvidia last year, which makes the eventual price look less like a windfall and more like the cost of waiting.

The load-bearing sentence in Huang's announcement is a commitment, not a claim: "NVIDIA compute will not be required to build on or deploy through Hugging Face." He went further, promising continued support for multi-cloud and multi-accelerator development, and said the platform will keep its own brand — emoji included. The acquisition is expected to close in the first half of next year.

Why It Matters

Nvidia has bought companies before, but almost always ones that made its silicon easier to run — the roughly $700 million purchase of Run:ai being the clearest recent example. Spending $12.93 billion on a software community and a website is a different category of decision entirely, and it is the first time the company has committed serious capital to something that is not, in any sense, a chip.

The strategic logic is not hard to reconstruct. Nvidia's advantage has never been the GPU alone; it is that the software ecosystem assumes the GPU. Open-weight models are the layer where hardware assumptions actually get baked in — which kernels get optimized first, which quantization formats ship by default, which accelerator gets a working example in the README. Whoever hosts the distribution point for open models has enormous influence over those defaults without ever having to write a rule. TechCrunch also noted a blunter commercial angle: Nvidia can bundle spare compute capacity with a Hugging Face offering and sell it to enterprises directly.

Nvidia's investment in the open ecosystem is already substantial enough to make the purchase feel less like a swerve. The company says it has released more than 500 models and over 250 open datasets on the platform, making it the largest single contributor there. It struck a reported $6 billion arrangement with the coding startup Poolside to develop open models, and told analysts on a recent earnings call that it has put more than $50 billion into frontier AI labs. Huang has also been campaigning on the policy side, co-authoring an open letter arguing that open weights broaden access and keep AI leadership distributed rather than concentrated.

The worry running underneath all of this has a specific name: vertical foreclosure through self-preferencing. Today the Hub maintains Optimum AMD and Optimum Intel as first-class libraries alongside its Nvidia tooling. Nobody expects Nvidia to delete them. The realistic failure mode is quieter — a maintenance budget that drifts, support for new architectures that lands three months late, documentation that surfaces one path more readily than another. None of that would violate a single sentence of Huang's promise.

Reaction

Clem Delangue's own framing has been notably unapologetic. In a post on X, the Hugging Face CEO thanked the community for proving that an open alternative to closed APIs was viable, then argued that scaling that alternative required resources the company did not have: "it needs more compute, more support, more collaboration." Huang, for his part, said Delangue came to him — that Hugging Face initiated the conversation about its next chapter rather than being cornered into one.

The developer response has been more divided, and the recurring phrase in commentary is that Hugging Face had become the Switzerland of AI software — a place every lab, every research group and every competing chip vendor used precisely because it was aligned with none of them. VentureBeat placed the deal alongside Stripe's acquisition of OpenRouter, reading the pair as a single trend: the independent connective tissue of the AI stack is being bought up by the companies it was meant to sit between.

Nvidia's public posture is that regulators will see the deal the same way it does. A company vice president and general manager for enterprise computing has argued that open-source AI is close to structurally a deconcentration platform, and said he expects the review to land overwhelmingly positive. That is an interesting argument to make in advance of filing, and it tells you which fight Nvidia thinks it is in.

What's Next

The most consequential thing between here and closing is antitrust review. Much of Nvidia's recent expansion has run through structures that avoided merger notification altogether — minority investments, licensing arrangements, compute deals, talent agreements. A $12.93 billion outright acquisition does not have that option. It will be reviewed in the United States and almost certainly in Europe, with the United Kingdom and possibly China as additional gates.

Nvidia has been here twice, with opposite outcomes. Its $40 billion attempt to buy Arm collapsed in 2022 after coordinated resistance from the FTC, the UK Competition and Markets Authority, the European Commission and Chinese regulators, on the argument that owning a neutral platform every competitor depended on was itself the harm. The much smaller Run:ai deal cleared European review and closed. Hugging Face sits uncomfortably between those precedents: far cheaper than Arm, but occupying a similarly load-bearing position in everyone else's workflow.

For developers, the useful markers over the next year are mundane and checkable. Do the Optimum libraries for AMD and Intel keep shipping support for new architectures on the same cadence? Does the inference-provider list on model pages stay genuinely multi-vendor? Do non-Nvidia deployment paths keep their place in the documentation? Those commits will say more than any announcement. The deal is expected to close in the first half of 2027.

Closing Thoughts

There is a version of this story where nothing bad happens, and it is not naive. Nvidia's incentive really does point toward open models being everywhere, because almost all of them run on its hardware anyway. A Hugging Face that stayed neutral and grew would serve Nvidia better than one that narrowed into a vendor storefront and drove developers to build somewhere else. The promise and the business case are, for now, pointing in the same direction.

What makes people uneasy is that this alignment is a matter of judgment rather than structure. Hugging Face became critical infrastructure almost by accident — it started in 2016 as a chatbot company, produced a library that happened to be the most convenient way to load a model, and then found that the entire field had quietly moved in. Being infrastructure without a durable business model is a precarious position, and $150 million in annualized revenue against the cost of hosting three million models is thinner than the platform's importance suggests. Somebody was going to end up owning this.

So the honest read is neither celebration nor alarm. The commitments Huang made are specific enough to be tested, which is more than most acquisition posts offer, and the ecosystem now has both the motive and the tooling to check them. The thing worth watching is not what Nvidia says in the next twelve months. It is whether the parts of Hugging Face that serve its competitors keep getting maintained with the same care as the parts that serve Nvidia.

한글 요약

엔비디아가 9월 3일 오픈 모델 플랫폼 허깅페이스를 129억 3,030만 달러에 인수한다고 공식 발표했습니다. 주주에게 약 119억 달러, 합류하는 직원 유지용 지분으로 약 10억 달러가 배정되며, 거래는 내년 상반기 마무리될 예정입니다. 허깅페이스는 모델 300만 개, 데이터셋 50만 개, 애플리케이션 100만 개를 호스팅하며 개발자 1,800만 명과 기업 20만 곳이 사용하는 사실상 유일한 오픈 웨이트 집결지입니다. 2016년 설립 이후 누적 3억 9,500만 달러를 조달했고, 최근 연환산 매출은 1억 5,000만 달러 수준으로 알려졌습니다. 지난해 엔비디아의 5억 달러 제안을 거절했다는 보도도 있었습니다.

젠슨 황 CEO는 발표문에서 "허깅페이스에서 빌드하거나 배포하는 데 엔비디아 컴퓨트가 요구되지 않는다"고 못박고, 멀티 클라우드·멀티 가속기 지원과 기존 브랜드 유지를 약속했습니다. 엔비디아 입장에서 오픈 모델은 하드웨어 선택이 실제로 결정되는 지점이고, 이 플랫폼은 그 기본값에 영향을 미치는 자리에 있습니다. 반대로 업계의 우려는 명시적 차별이 아니라 조용한 방치입니다. 현재 1급 라이브러리로 관리되는 Optimum AMD·Optimum Intel의 유지보수 속도가 서서히 느려지는 식의 변화는 황 CEO의 약속을 한 문장도 어기지 않으면서 일어날 수 있습니다.

가장 큰 변수는 반독점 심사입니다. 그동안 엔비디아의 확장은 소수 지분 투자나 라이선스 계약처럼 신고 의무를 피해 가는 구조가 많았지만, 129억 달러 규모의 직접 인수는 미국과 유럽, 영국의 심사를 정면으로 통과해야 합니다. 2022년 Arm 인수가 각국 규제당국의 공동 압박으로 무산된 전례와, EU 승인을 받아 마무리된 Run:ai 인수 전례가 동시에 존재합니다. 개발자 입장에서 확인할 지표는 단순합니다. 비엔비디아 가속기 라이브러리가 신규 아키텍처를 같은 속도로 지원하는지, 모델 페이지의 추론 제공자 목록이 다변화된 상태를 유지하는지입니다. 발표문보다 커밋 기록이 답을 줄 것입니다.