英伟达129.3亿美元买下Hugging Face的交易合同都签了，美国司法部先敲门了。
一张支票买断十年
一笔没交割就被盯上的收购
129.3亿美元，问询函已递到英伟达门口
今年9月，英伟达宣布以129.3亿美元收购Hugging Face，这是该公司史上最大的一笔整体收购。黄仁勋当时的表态很干脆：Hugging Face将继续作为面向整个AI生态系统的开放平台运营，允许开发者自由上传、下载模型与数据集，不强制要求使用英伟达硬件。交割时间写的是2027年上半年。
但合同刚签，问询函就到了。据路透社援引两位知情人士的消息，美国司法部已就此交易展开反垄断调查，关注点不在芯片卖得太好，而在分发渠道会不会被一家同时控制算力硬件的公司垄断。
这锅我们不背
监管看的不是芯片，是分发渠道
司法部官员询问的核心问题只有一个：这笔收购是否会让英伟达掌控AI领域一个基础性的分发渠道。监管层看到的不是单次交易，而是一条正在拉紧的纵向链条。
英伟达在算力硬件层是绝对的寡头，然后通过投资OpenAI、Anthropic、CoreWeave织了一张纵网，再收购SchedMD（容器调度软件）和Run:ai，最近又把技术授权嵌入Groq的业务。Hugging Face是这条链条的最后一环——谁控制了模型的分发入口，谁就有了定义兼容标准、排序默认路径、影响开发者选择的杠杆。
``mermaid
英伟达纵向整合图谱
英伟达的纵向野心
从Mellanox到SchedMD、Run:ai，再到Hugging Face
2020年收购Mellanox，那是围绕GPU主业的横向补充。这次买Hugging Face，逻辑完全不同。
但仔细看英伟达过去两年的动作，它其实同时落了三步棋。
第一步，算力硬件。数据中心GPU市占率超过80%，CUDA生态形成了事实上的锁定。
第二步，算力调度软件。2024年收购SchedMD，拿到Slurm作业调度系统；同年收购Run:ai，掌控Kubernetes环境下的GPU资源编排。这两家都是数据中心里跑大模型训练的必经之路。
第三步，模型分发入口。Hugging Face托管了超过300万个AI模型，月活开发者1800万，是全球最大的开源模型社区。
``mermaid
英伟达纵向整合三节点
这三步合在一起，形成的是一张纵向的控制网。即使不直接封杀竞争对手，也可能通过默认配置、搜索排名来影响整个行业的技术选择。
算力+调度+分发，三个关键节点同时落子
黄仁勋在宣布收购时强调，Hugging Face将继续作为开放平台运营。问题是，承诺归承诺，实际控制权的归属是另一回事。当平台的所有权变成一家公司的，所谓「开放」就变成了一种策略选择，而不是一种结构保障。
有开发者直言，这等于「把GitHub之后，Hugging Face也变成了卖GPU的工具」。合法开源，实际锁定。
开源社区的失语与担忧
Hugging Face CEO去年还说「不想被单一巨头影响独立方向」
2025年底，Hugging Face CEO Clément Delangue在接受TechCrunch采访时明确表示，公司已拒绝英伟达5亿美元的独家投资邀约。「我不想让单一占主导地位的巨头投资人影响平台的独立方向。」不到一年，这笔交易的金额从5亿飙升至129.3亿美元，Delangue也从拒绝者变成了卖方。
这种「变脸」并非偶然。Hugging Face自2020年融资以来走过六轮融资，估值从2023年的40亿美元飙升至2025年底的300亿美元以上。当一块蛋糕足够大时，创始人的独立性往往最先让位于资本的逻辑。
开发者担心的不是被收购，而是中立性消失
在X平台和Hacker News上，开发者们的焦虑集中在同一个问题上：当Hugging Face成为英伟达的子公司后，搜索排名、推荐算法、默认配置会不会偏向英伟达的生态？
一位名为@mlengr的开发者发帖称：「他们只是想关掉开源。」更深层的问题在于生态锁定。英伟达已经在芯片层面形成垄断，如果Hugging Face再被纳入麾下，开发者上传模型时的默认路径、框架兼容性、推理优化的推荐顺序，都可能向英伟达体系倾斜。
``mermaid
开发者生态锁定路径
值得注意的是，这种担忧并非空穴来风。OpenAI、Anthropic等头部AI实验室已经在自研芯片以降低对英伟达的依赖，Meta更是推出了自己的开源模型和硬件方案。这些动作的背后，是对单一供应商风险的警惕。
开源社区的失语，正是这场交易中最令人不安的信号。1800万开发者和300万个模型构成的生态，十年积累，如今被一张支票买断。
反垄断这道坎怎么过
129亿美元超过HSR申报门槛，FTC和司法部双重审查
129.3亿美元，远超美国《哈特-斯科特-罗迪诺法》（HSR Act）的申报门槛。按照2026年的标准，超过5.37亿美元的交易就必须向FTC和司法部同时申报，等待至少30天的静默期。
英伟达主动选择了这个流程，把交割时间线写在声明里——预计2027年上半年完成。这给了监管机构充足的时间打量这笔交易的全貌。
路透社的报道指出，司法部的问询已经扩展到英伟达过去数年的并购、对外投资、销售模式，以及是否对客户施加压力。Hugging Face只是最新一环，前面还有对SchedMD和Run:ai的收购，以及与Groq的技术授权交易。
英伟达的护身符是「保持开放」，但开源能有多开放
英伟达的回应很直接：通过Hugging Face进行开发或部署，并不要求使用英伟达的计算资源。
这句话的逻辑是对的，但漏了一个关键变量——默认路径。开源平台的中立性从来不取决于「技术上能不能用别人的产品」，而取决于「开发者在遇到问题时最先看到什么」。搜索排名、推荐算法、文档优先展示的顺序、预配置环境的默认选择——这些才是软性锁定的真正抓手。
一位在Hugging Face上托管过模型的开发者透露，他们团队曾经发现自家模型在相关搜索里的排名持续下降，最后不得不付费投广告才能回到前页。「当时不知道为什么，现在回头看，可能就是平台规则微调整。」
这不是阴谋论，这是平台经济的基本逻辑。免费的平台需要变现，当它的最大股东变成了一家卖GPU的公司时，「兼容并支持各类芯片」这句话的成本就会悄悄改变。
更狠的是，开源社区干了十年，攒出1800万开发者和300万个模型，然后被一张支票买走。社区创造价值，巨头收购价值，这个剧本一遍遍重演。
Hugging Face创始人Clément Delangue去年年底还说过一句漂亮话：「我们拒绝了一笔5亿美元的独家投资。」不到一年，这笔交易的金额翻了二十多倍，而且不再是投资，是全资收购。
开发者社区的反应可以用一个词概括：累了。
落点：如果Hugging Face真的改姓
不是会不会被吞，而是开源生态还有多少议价权
这笔交易最终会不会被拆，其实不是最值得关注的问题。更值得问的是：当开源平台的定价权彻底消失，下一次有类似的平台出现时，开发者手里还有什么筹码？
答案是几乎没有。开源社区的议价权来源于两个东西：用户的不可替代性和替代平台的稀缺性。但当这个渠道本身变成并购标的时，开发者群体已经没有任何议价工具了。你不能「换一家平台」，因为不存在第二家体量相当的分发枢纽；你也不能「用脚投票」，因为你的模型挂在那里才有流量。
监管层显然意识到了这个问题。路透社的分析认为，FTC和司法部关注的核心不是「英伟达会不会垄断芯片」，而是「英伟达会不会通过控制分发渠道来巩固芯片垄断」。这是一个纵向垄断的经典问题。
英伟达的退路可能在于两点：一是证明Hugging Face的开放性不会被实质性削弱；二是把这笔交易纳入更广泛的「促进AI创新」叙事中。
后者在某些层面已经生效。司法部发言人在回应中提到：「在本届政府领导下，将继续优先推动超级智能发展，以造福美国人民。」这句话没有明说，但翻译过来就是：我们不希望反垄断执法拖慢美国在AI领域的竞争力。
但这不代表审查会放松。合规的终点往往是条件交换，而不是无条件放行。
如果你是一家依赖Hugging Face分发模型的创业公司，现在能做的只有一件事：评估一旦平台规则改变对你的依赖程度，并在交割完成前完成关键数据的本地备份和替代渠道的搭建。这不是悲观，是风险管理的基本功。
开源不会死，但开源的入口正在成为新一轮基础设施战争的焦点。
参考文献
- [1] 消息称美国司法部就英伟达 129.3 亿美元收购 Hugging Face 一事展开问询 - IT之家. www.ithome.com/1/019/272.h…
- [2] 英伟达129亿美元收购Hugging Face遭美司法部问询 - 电子工程专辑. www.eet-china.com/news/202610…
- [3] 美国司法部正调查英伟达收购Hugging Face - 财闻网. www.cww.net.cn/article?id=…]
- [4] 美国司法部就英伟达收购Hugging Face展开问询 | 网易订阅. www.163.com/dy/article/…
- [5] Hugging Face官网声明（2026年9月3日）：英伟达宣布同意以129.3亿美元收购Hugging Face.
延伸入口
- 原文归档：tobemagic.github.io/ai-magician…
- 公众号：计算机魔术师
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
The contract for the purchase of Hugging Face was signed by the United States Department of Justice, which knocked first.
Ten years for a check.
I've been targeted for a few days.
129.30 billion United States dollars. Letters of enquiry have been sent to the gates of Inverda.
In September of this year, Yin Weida announced the acquisition of Hugging Face at $12.93 billion, the largest total acquisition ever made by the company. Hoang In-hoon made a very clear statement at that time: Hugging Face would continue to operate as an open platform for the entire AI ecosystem, allowing developers to freely upload and download models and data sets and not to impose the use of British hardware. The date of delivery was the first half of 2027.
However, as soon as the contract was signed, the letter of enquiry arrived. According to Reuters sources, the United States Department of Justice has launched an anti-monopoly investigation into the transaction, focusing on whether the chip is sold too well and whether the distribution channel is monopolized by a company that controls both computing hardware.
We don't have to carry this pot.
It's not a chip, it's a distribution channel.
Only one of the core questions asked by Ministry of Justice officials was: Whether or not this acquisition will allow Yvette to take control of a basic distribution channel in the AI area. It is not a single transaction, but a vertical chain that is being tightened.
Young Weida is an absolute oligopoly at the level of computing hardware, and then buys SchedMD (container control software) and Run:ai and recently embeds technical authorization into Groq's business by investing in Open AI, Anthropic, CoreWeve. Hugging Face is the last link of the chain - who controls the distribution portal of the model has defined compatibility criteria, sorted default paths, leverage to influence developers' choice.
``mermaid
Longitudinal integration map for Inverda
Young Waida's vertical ambition.
From Mellanox to SchedMD, Run:ai, to Hugging Face
The acquisition of Mellanox in 2020 was a horizontal supplement to the GPU industry. It's a different logic to buy Hugging Face this time.
But looking closely at the actions of Inverda over the past two years, it's actually a three-step game.
First step, computational hardware. The data centre GPU has a rate of over 80 per cent and the CUDA ecology has created a de facto lock-in.
Step two, calculator control software. In 2024, SchedMD was acquired and handed over to the Slurm Operations Control System; in the same year, Run:ai was acquired to control the GPU resource allocation in the Kubernetes environment. Both are essential for running LLM training in data centres.
Step three, model distribution portal. Hugging Face hosts over 3 million AI models, 18 million of which are the largest open source model community in the world.
``mermaid
Imwida vertically integrated three nodes
Together, these three steps form a vertical control network. Even if they do not shut down competitors directly, they may influence technology choices across the industry through default configurations and searches for rankings.
Calculus + dispatch + distribution, with three key nodes down simultaneously
Hoang In-hoon, in announcing the takeover, stressed that Hugging Face would continue to operate as an open platform. The problem is that the promise is the promise and the attribution of the actual control is another matter. When the ownership of the platform becomes a company, so-called "open" becomes a strategic choice, not a structural safeguard.
There are developers who speak out, and this means "Hugging Face became a GPU seller after GitHub." Legal open source, actual lock.
Disappointment and fear in open-source communities
Hugging Face CEO last year said, "Don't want to be influenced by a single giant."
In an interview with TechCrunch at the end of 2025, Hugging Face CEO Clément Delange made it clear that the company had rejected Britain’s $500 million exclusive investment offer. “I do not want a single dominant major investor to influence the platform’s independent direction. "In less than a year, the amount of the deal soared from 500 million to $12.93 billion, and Delange became a seller from the rejecter.
This "face change" is not an accident. Hugo Face has gone through six rounds of financing since 2020, with the valuation soaring from $4 billion in 2023 to over $30 billion at the end of 2025. When a piece of cake is large enough, the independence of the founders tends to give way first to the logic of capital.
The developers are not worried about being bought, but about being neutral.
On platform X and Hacker News, developers focused their anxiety on the same issue: would searching for rankings, recommended algorithms, default configurations favouring the ecology of Britain when Hughes Face became a subsidiary of Britain?
A developer named @mlengr wrote: "They just want to turn off the source. The deeper problem is ecological locking. Young Weeda already has a monopoly at the chip level, and the default path, framework compatibility, and the recommended order of reasoning optimization for the developers to upload the model could be tilted in favour of the British Weeders system if Hugging Face is once again included under the umbrella.
``mermaid
Developer ecological lock path
It is noteworthy that this concern is not empty. Open AI, Anthropic and other head AI laboratories are already developing their own chips to reduce their dependence on Weeda, and Meta has launched its own open source model and hardware programme. Behind these moves is an alert to single-supplier risks.
The silence of open-source communities is the most disturbing signal of the deal. The ecology of 18 million developers and 3 million models has accumulated over a decade and is now bought off by a cheque.
What about the antimonopoly?
$12.9 billion above the HSR threshold, double review by the FTC and the Ministry of Justice
US$ 129.30 billion, well above the United States threshold for the Hart-Scott-Rodino Act. According to the 2026 criteria, transactions in excess of $537 million must be reported to both the FTC and the Ministry of Justice, pending a period of silence of at least 30 days.
The process was chosen by Young Weida, who wrote the time line in the statement — expected to be completed in the first half of 2027. This gives the regulator sufficient time to take stock of the transaction.
Reuters reports that the Ministry of Justice ' s enquiries have been extended to the pattern of mergers and acquisitions, outward investments, sales over the past few years, as well as pressure on clients. Hugging Face is only the latest link, preceded by the acquisition of SchedMD and Run:ai, and the technical authorization transaction with Groq.
Young Waida's amulet is "Keep Open," but how open is the open source?
The response is direct: development or deployment through Hugging Face does not require the use of Britain's computing resources.
The logic of this sentence is correct, but one key variable — the default path — is missing. The neutrality of open-source platforms has never depended on "technology being able to use other people's products" but on "what the developers see first when they encounter problems." Search for rankings, recommended algorithms, priority sequences for documents, default options for preconfiguring environments — these are the true catchers of soft locking.
One of the developers who hosted the model on Hugging Face revealed that their team had found that the family model had continued to decline in its ranking in the relevant search, and had to advertise paid to return to the front page. “I don’t know why, now look back, it might be a slight adjustment of the platform’s rules.”
This is not conspiracy theory; it is the basic logic of the platform economy. The cost of "compatibility with and support for chips" changes quietly when its largest shareholder becomes a GPU sales company.
Worse still, open-source communities have been working for 10 years to produce 18 million developers and 3 million models, and have been bought by a cheque. Community creation of value, acquisition of value by giants is repeated over and over again.
At the end of last year, Clément Delange, founder of Hugging Face, also said a beautiful thing: "We rejected an exclusive investment of $500 million." "In less than a year, the amount of the deal has more than doubled, and it is no longer an investment, but a full-scale purchase.
The response of the developers ' community can be summarized in one word: tired.
Location: If Hugging Face really changes his last name
It's not if it's going to be swallowed, it's if it's the open source. Rights
Whether or not the transaction will eventually be dismantled is not a matter of the utmost concern. What is more worth asking is what leverage the developers have when the pricing rights of open-source platforms have disappeared and when similar platforms appear next.
The answer is almost no. The bargaining power of open-source communities derives from two things: the scarcity of alternative and alternative platforms for users. But when the channel itself became the subject of mergers and acquisitions, the group of developers no longer had any bargaining tools. You can't change the platform, because there's no second distribution hub, and you can't vote with your foot because your model is there for traffic.
The regulatory level is clearly aware of this problem. Reuters's analysis suggests that the focus of FTC's and Justice's concerns is not "will Yvette monopolize the chip" but "will Yvette consolidate the chip monopoly by controlling distribution channels"? This is a classic issue of vertical monopolies.
The retreat of Ingginga may be based on two points: on the one hand, proof that Hugging Face’s openness will not be substantially weakened; and on the other hand, the inclusion of the deal in the broader narrative of “Promoting AI Innovation.”
The latter has entered into force at some levels. In response, the spokesperson for the Department of Justice mentioned that under the current administration, priority will continue to be given to the promotion of super-smart development for the benefit of the American people. This sentence is not clear, but the translation is that we do not want antimonopoly enforcement to slow down the United States competitiveness in the AI field.
But that does not mean that the review will relax. The endpoint of compliance is often a trade-off rather than unconditional release.
If you are an entrepreneurial company that relies on the Hugging Face distribution model, there is only one thing that can be done now: to assess the local backup and alternative channels of key data once the platform rules change the level of reliance on you, and before delivery is complete. This is not a pessimism, but a fundamental function of risk management.
Open sources will not die, but the entrance to open sources is becoming the focus of a new round of infrastructure wars.
References
-[1] According to the news, the U.S. Department of Justice is asking about the $12.33 billion acquisition of Hugging Face - IT House. www.ithome.com/1/019/272.h...
- [2] The acquisition of Hugging Face by Young Weidar of $12.9 billion was questioned by the United States Department of Justice - Electronic Engineering Album. www.eet-china.com/news/202610...
U.S. Department of Justice is investigating the acquisition of Hugging Face - Financial Information Network. www.cw.net.cn/article?id=...]
- [4] U.S. Department of Justice (DOJ) inquiries regarding the acquisition of Hugging Face by Inweida.
- [5] Hugging Face official network statement (3 September 2026): YVD announced its agreement to purchase Hugging Face.
Extension entrance
- Original archive: tobemagic.github.io/ai-magician...
- WeChat Official Account: Computer magician
Auto-fetched and translated by AI Atlas. All rights belong to the original author.