Why Nvidia May Spend $13bn on Hugging Face’s Doorway to AI Demand
Why Nvidia May Spend $13bn on Hugging Face’s Doorway to AI Demand¶
- Source: SemiVision
- Author: SEMIVISION
- Published: 2026-08-30T13:38:26+08:00
- URL: https://tspasemiconductor.substack.com/p/why-nvidia-may-spend-13bn-on-hugging
中文翻譯與英文原文（雙語）¶
(2) 為何輝達可能斥資 130 億美元收購 Hugging Face，以掌握 AI 需求的入口 --- (2) Why Nvidia May Spend $13bn on Hugging Face’s Doorway to AI Demand
Why Nvidia May Spend $13bn on Hugging Face’s Doorway to AI Demand¶
為何輝達可能斥資 130 億美元收購 Hugging Face，以掌握 AI 需求的入口
Original Article By SemiVision Research [Reading time: 7 mins]¶
原文作者：SemiVision Research [閱讀時間：7 分鐘]
Aug 30, 2026
Share分享
Why Nvidia May Spend $13bn on Hugging Face’s Doorway to AI Demand
為何輝達可能斥資 130 億美元收購 Hugging Face，以掌握 AI 需求的入口
Nvidia has reportedly agreed to acquire Hugging Face for $12.9bn. Neither company has formally confirmed the transaction, so the deal may yet change or disappear. But the reported price is revealing.
據報導，Nvidia 已同意以 129 億美元收購 Hugging Face。雙方公司尚未正式確認此交易，因此該案仍可能生變或告吹。然而，這項傳聞中的價格卻極具啟發性。
Hugging Face generates annualised revenue of roughly $150m. Nvidia would therefore be paying about 86 times sales—a multiple that looks absurd if Hugging Face is treated as an ordinary software company.
Hugging Face 的年化營收約為 1.5 億美元。換言之，Nvidia 將支付相當於其營收 86 倍的對價；若將 Hugging Face 視為一般軟體公司，此估值倍數顯得荒謬至極。
It makes more sense if Nvidia is not buying the revenue. Nor, for that matter, is it buying the models. It is buying the doorway through which much of the open-AI world enters.
但若理解 Nvidia 並非在購買營收，甚至也不是在購買模型本身，此事便較合理。Nvidia 真正要買的是通往開放式 AI 世界的主要入口。
The model shopfront模型的展示門面
Hugging Face is commonly described as a repository for open models. That understates its importance. It is closer to a mixture of GitHub, an app store and a package-management system for artificial intelligence.
Hugging Face 常被描述為開放模型的儲存庫，但這低估了其重要性。它更接近於 GitHub、應用程式商店與人工智慧套件管理系統的混合體。
By the end of 2025, the platform had 13m users, more than 2m public models and over 500,000 public datasets. By August 2026, its collection was approaching 3m models. More than 30% of Fortune 500 companies have verified accounts on the platform. Hugging Face’s own data show that it is no longer merely a home for researchers. It is increasingly part of the enterprise-AI workflow.
截至 2025 年底，該平台已擁有 1,300 萬用戶、超過 200 萬個公開模型及逾 50 萬個公開資料集。到了 2026 年 8 月，其模型收藏量已逼近 300 萬個。超過 30% 的《財富》500 強企業在該平台上擁有驗證帳戶。Hugging Face 自身的數據顯示，它已不再僅是研究人員的聚集地，而是日益成為企業級人工智慧工作流程的一部分。
Developers arrive at Hugging Face to discover models, compare downloads, inspect licences and evaluate benchmarks. They then decide which model to fine-tune, quantise or deploy.
開發人員來到 Hugging Face 以發掘模型、比較下載量、檢視授權條款並評估基準測試結果。隨後，他們決定要對哪個模型進行微調、量化或部署。
That decision may look like software selection. In reality, it travels down the entire computing stack:
此項決策看似軟體選型，實際上卻貫穿整個運算堆疊：
Model architecture → memory requirements → inference framework → accelerator choice → GPU count → network design → power consumption.
模型架構 → 記憶體需求 → 推論框架 → 加速器選擇 → GPU 數量 → 網路設計 → 功耗。
SEMIVISION @_@ 是一份由讀者支持的出版物。若要接收最新文章並支持我的工作，歡迎考慮成為免費或付費訂閱者。
7月25日7 月 25 日
Not every download produces a GPU order. Many small models run perfectly well on CPUs, smartphones or competing accelerators. But once an experiment becomes a production service, choices concerning parameter count, mixture-of-experts architecture, context length and numerical precision begin to determine spending on GPUs, HBM, networking and electricity.
並非每次下載都會產生 GPU 訂單。許多小型模型在 CPU、智慧型手機或競爭性加速器上都能完美運行。但一旦實驗轉化為生產服務，參數數量、混合專家（Mixture-of-Experts）架構、上下文長度以及數值精度等選擇，便開始決定在 GPU、HBM、網路與電力方面的支出。
Hugging Face sits upstream of that spending. It is where demand begins to take shape.
Hugging Face 位於該支出的上游，是需求開始成形的地方。
NVIDIA and Hugging Face to Connect Millions of Developers to Generative AI Supercomputing
NVIDIA 與 Hugging Face 攜手將數百萬開發者連結至生成式 AI 超級運算
Nvidia already owns the kitchen
Nvidia 已掌握廚房
Nvidia’s hardware loop is largely complete. It supplies GPUs, CPUs, NVLink, InfiniBand, Spectrum-X Ethernet, systems and increasingly entire racks. Around them sit CUDA, TensorRT-LLM, Dynamo, NIM and AI Enterprise.
Nvidia 的硬體生態系已基本完備，提供 GPU、CPU、NVLink、InfiniBand、Spectrum-X 乙太網路、系統，並日益擴展至整機櫃解決方案。圍繞這些硬體的是 CUDA、TensorRT-LLM、Dynamo、NIM 與 AI Enterprise。
The company is no longer selling a single type of chip. It is selling an AI factory.
該公司不再只銷售單一類型的晶片，而是銷售一座 AI 工廠。
以餐廳術語來說，Nvidia 已擁有烤箱、蒸爐、冰箱、外送箱，以及廚房作業系統的大部分。客戶只需加入資料、電力與資本，token 便會開始產生。
What Nvidia still lacks is control of the menu at the entrance.
Nvidia 目前仍欠缺的是對入口處選單的控制權。
Hugging Face provides that menu. Owning it would allow Nvidia to see which models are attracting developers, which architectures are gaining traction and which experimental workloads may soon become commercial deployments.
Hugging Face 提供了這樣的選單。擁有它將使 Nvidia 能夠洞察哪些模型正吸引開發者、哪些架構日益受到青睞，以及哪些實驗性工作負載可能即將轉為商業部署。
Such information would be useful far beyond advertising or subscription sales. It could influence the design of future GPUs, memory configurations, networking topologies and inference software. Hugging Face could become an early-warning system for changes in computing demand.
此類資訊的價值遠超廣告或訂閱銷售。它可能影響未來 GPU 的設計、記憶體配置、網路拓撲以及推論軟體。Hugging Face 可成為預測運算需求變化的早期預警系統。
The $12.9bn price may thus cover four assets:
129 億美元的價格可能涵蓋四項資產：
the global index of open models, datasets and AI applications;
開放模型、資料集與 AI 應用的全球索引；
the community and habits of millions of developers;
數百萬開發者的社群與使用習慣；
an increasingly important enterprise deployment workflow; and
日益重要的企業部署工作流程；以及
early signals about the future direction of AI workloads.
關於 AI 工作負載未來走向的早期訊號。
Together, these could create a feedback loop running from model discovery to hardware procurement.
這些要素結合起來，可形成從模型發掘到硬體採購的回饋迴路。
Poolside makes the recipes; Hugging Face distributes them
Poolside 負責打造配方；Hugging Face 則負責分發這些配方。
Nvidia’s recent transactions make more sense when viewed as pieces of the same puzzle.
若將 Nvidia 近期的交易視為同一拼圖的各個片段，其背後邏輯便更為清晰。
The company has reportedly agreed to pay $6bn for a non-exclusive licence to Poolside’s Model Factory, invest another $1bn in the firm and offer jobs to 109 employees. Poolside itself will survive, but much of its model-building software and expertise will move closer to Nvidia. The reported structure resembles an acquisition stripped of the company shell.
據報導，該公司已同意支付 60 億美元，取得 Poolside Model Factory 的非專屬授權，並額外投資 10 億美元，同時為 109 名員工提供職位。Poolside 本身將繼續存續，但其大部分模型建構軟體與專業知識將更貼近 Nvidia。此項據報的交易架構，形同剝離公司外殼的收購案。
If Hugging Face is the menu, Poolside provides the recipes and kitchen procedures.
如果 Hugging Face 是菜單，Poolside 則提供食譜與廚房作業流程。
Nemotron, Nvidia’s family of open models, performs another function. Nvidia publishes model weights, datasets and training techniques, allowing developers to download, modify and deploy the models. Commercial users can later be guided towards NIM, TensorRT-LLM and AI Enterprise. Nvidia says Nemotron models can be downloaded from Hugging Face and used free in production.
Nemotron 是 Nvidia 的開放模型系列，具備另一項功能。Nvidia 發布模型權重、資料集與訓練技術，讓開發者能夠下載、修改並部署這些模型。商業用戶隨後可被引導至 NIM、TensorRT-LLM 與 AI Enterprise。Nvidia 表示，Nemotron 模型可從 Hugging Face 下載，並免費用於生產環境。
Run Hugging Face Models Instantly with Day-0 Support from NVIDIA NeMo Framework
透過 NVIDIA NeMo Framework 的首日支援，立即執行 Hugging Face 模型
Giving models away is not an act of charity. A free model optimised for Nvidia’s software stack can create demand for the hardware beneath it. CUDA followed a similar logic: make the programming environment indispensable, then allow the economics to surface elsewhere.
免費提供模型並非出於慈善。一個針對 Nvidia 軟體堆疊優化的免費模型，能夠為其底層硬體創造需求。CUDA 也遵循類似的邏輯：先讓程式開發環境變得不可或缺，再讓經濟效益在其他層面顯現。
Hugging Face would connect these components. Poolside supplies model-making machinery. Nemotron provides the open weights. Hugging Face distributes them. Nvidia’s infrastructure runs them.
Hugging Face 將串連這些元件。Poolside 提供模型建構機制，Nemotron 提供開放權重，Hugging Face 負責分發，而 Nvidia 的基礎設施則負責執行。
Perplexity occupies a different part of the stack. Nvidia is reportedly considering an investment in the AI-search company at a valuation exceeding $30bn. It is seeking exposure, not control. Reuters reported that Perplexity’s annualised revenue had risen above $750m.
Perplexity 佔據堆疊中不同的位置。據報導，Nvidia 正考慮以超過 300 億美元的估值投資這家 AI 搜尋公司，其目的在於取得曝險而非控制權。路透社報導指出，Perplexity 的年化營收已突破 7.5 億美元。
That distinction is deliberate. Nvidia may want a seat in one of the busiest restaurants without buying the restaurant itself.
這項區別是刻意的。Nvidia 可能希望在最繁忙的餐廳中佔有一席之地，卻無需買下整間餐廳。
Why Nvidia does not want to become ChatGPT
為何 Nvidia 不想成為 ChatGPT
The Hugging Face deal could be mistaken for an attempt to build Nvidia’s answer to ChatGPT. That would probably be the wrong interpretation.
與 Hugging Face 的交易可能被誤解為打造 Nvidia 版 ChatGPT 的嘗試。這很可能是一種錯誤的解讀。
Nvidia’s greatest strategic advantage is neutrality. It can sell infrastructure to OpenAI, Anthropic, xAI, Microsoft, Amazon, Google, sovereign governments and smaller AI laboratories, even as they compete ferociously with one another. So long as all of them require more computing power, Nvidia can profit regardless of which model wins.
Nvidia 最大的戰略優勢在於其中立性。即使 OpenAI、Anthropic、xAI、Microsoft、Amazon、Google、各國政府與小型 AI 實驗室彼此激烈競爭，Nvidia 仍能向他們銷售基礎設施。只要這些企業都需要更多運算能力，無論哪種模型勝出，Nvidia 都能獲利。
Buying a large consumer application would complicate that position. Nvidia would stop being merely the supplier of picks and shovels and begin competing with the miners. Its largest customers would have an even stronger incentive to design their own chips or support alternative accelerators.
收購大型消費級應用程式將使該立場變得複雜。Nvidia 將不再僅僅是提供鏟子與鋤頭的供應商，而開始與礦工競爭。其最大客戶將有更強烈的動機去設計自有晶片或支援替代加速器。
6月1日6 月 1 日
Applications are also fickle. Today’s leading chatbot or AI-search service may not remain dominant. Model registries, software libraries, version histories, enterprise integrations and developer habits are more persistent.
應用程式也變化無常。今日領先的聊天機器人或 AI 搜尋服務，未必能持續保持主導地位。相較之下，模型註冊表、軟體函式庫、版本歷史記錄、企業整合方案以及開發者的使用習慣則更為持久。
Nvidia therefore appears to be assembling a deliberately incomplete full stack:
因此，Nvidia 似乎正在刻意組建一個不完整的全棧架構：
licences and hiring deals secure technology and talent;
授權協議與人才聘用確保了技術與人才的獲取；
Nemotron fills the open-model layer;
Nemotron 填補了開放模型層；
Hugging Face provides distribution;
Hugging Face 提供分發；
minority investments offer exposure to applications; and
少數股權投資可提供應用程式的曝險；且
Nvidia’s hardware and software capture the resulting computing demand.
Nvidia 的硬體與軟體能掌握由此產生的運算需求。
The application layer remains open not because Nvidia has forgotten it, but because filling it would turn customers into competitors.
應用層之所以保持開放，並非因為 Nvidia 遺忘了它，而是因為填補該層將使客戶轉變為競爭對手。
Financing the customers為客戶提供資金
Nvidia is extending this loop beyond technology and into finance.
輝達正將此循環從技術層面延伸至金融領域。
In August 2026 it disclosed residual-value guarantees connected to an OpenAI data-centre campus in Ohio. Nvidia’s potential payment obligation under the initial agreements is capped at $105bn. If OpenAI defaults on its leases, Nvidia could be required to assume them, find another tenant or help dispose of the assets. The SEC filing describes an arrangement covering approximately 4.25 gigawatts of IT load.
2026 年 8 月，輝達揭露了與俄亥俄州 OpenAI 資料中心園區相關的殘值保證。根據初始協議，輝達的潛在付款義務上限為 1,050 億美元。若 OpenAI 違約其租賃合約，輝達可能被要求承接這些租約、尋找其他承租者，或協助處置相關資產。美國證券交易委員會（SEC）的文件描述了一項涵蓋約 4.25 吉瓦（GW）IT 負載的安排。
Colette Kress, Nvidia’s finance chief, has said that demand from AI laboratories which the company expects to support with its balance sheet could account for roughly a quarter of its business next year. Nvidia’s earnings-call transcript also notes that the company has invested nearly $50bn in frontier laboratories.
Nvidia 財務長 Colette Kress 表示，公司預期將以資產負債表支持的 AI 實驗室需求，明年可能佔其業務約四分之一。Nvidia 的財報會議紀錄亦指出，該公司已向前沿實驗室投資近 500 億美元。
Nvidia is therefore doing three things at once. It is making models easier to build, making them easier to discover and helping customers finance the infrastructure required to run them.
因此，Nvidia 同時在做三件事：讓模型更容易建構、更容易被發現，並協助客戶籌措運行這些模型所需的基礎設施資金。
In effect, it is helping arrange the electricity bill and the lease before waiting for the racks to arrive.
實際上，它是在機架送達之前，就先協助安排好電費帳單與租賃合約。
Hugging Face and NVIDIA to Accelerate Open-Source AI Robotics Research and Development
Hugging Face 與 NVIDIA 攜手加速開源 AI 機器人技術的研發
Hugging Face’s LeRobot open-source framework combined with NVIDIA AI and robotics technology will enable researchers and developers to drive advances across a wide range of industries.
Hugging Face 的 LeRobot 開源框架結合 NVIDIA AI 與機器人技術，將助力研究人員與開發者推動各行業的進步。
The neutrality paradox中立性悖論
Hugging Face’s greatest asset is also the deal’s biggest vulnerability.
Hugging Face 最大的資產，也是這筆交易最大的弱點。
Developers use the platform partly because it is regarded as relatively neutral. Models from Meta, Google, Microsoft, Mistral, Chinese companies and independent researchers can be placed side by side. Users remain free to choose their preferred cloud, framework and accelerator.
開發者使用該平台的部分原因，在於其被視為相對中立。來自 Meta、Google、Microsoft、Mistral、中國公司及獨立研究人員的模型可以並列展示。使用者仍可自由選擇偏好的雲端服務、框架與加速器。
If search rankings, recommendations, deployment tools or performance optimisation begin to favour CUDA too visibly, competitors may fund alternative repositories. AMD, Intel, Google, Amazon and sovereign-AI programmes would have stronger reasons to support rival distribution channels.
若搜尋排名、推薦機制、部署工具或效能優化過於明顯地偏向 CUDA，競爭對手可能會資助替代的程式庫。AMD、Intel、Google、Amazon 以及主權 AI 計畫將有更強烈的理由支持 rival 分發管道。
Regulators may also ask whether the dominant supplier of AI accelerators should control one of the main gateways through which models are discovered and deployed. Model files are portable. The surrounding social graph, download histories, organisation accounts, documentation and developer trust are much harder to reproduce.
監管機構也可能質疑，作為 AI 加速器的主導供應商，Nvidia 是否應掌控模型發現與部署的主要入口之一。模型檔案具有可攜性，但周圍的社交圖譜、下載歷史、組織帳戶、文件記錄以及開發者的信任，則難以複製。
The deal therefore contains a paradox. Nvidia must control Hugging Face without making developers feel that Hugging Face is controlled.
因此，這項交易蘊含一個悖論：Nvidia 必須控制 Hugging Face，同時又不能讓開發者覺得 Hugging Face 已被其控制。
Turning the platform into an obvious Nvidia showroom would destroy part of what Nvidia is paying for. The entrance is valuable because everyone believes they may use it.
若將該平台轉變為明顯的 Nvidia 展示廳，將會摧毀 Nvidia 所付費購買的部分價值。此入口之所以珍貴，正因所有人都相信自己可以使用它。
A toll-free gate worth $12.9bn
價值 129 億美元的免費通行門戶
Nvidia would not be spending $12.9bn on nearly 3m model files. Nor would it be paying merely for $150m in annualised revenue.
輝達不會為了近 300 萬個模型檔案而耗資 129 億美元，也不會僅僅為了每年 1.5 億美元的營收而支付如此鉅額。
It would be buying influence over how models are discovered, evaluated and deployed—the point at which software choices begin their journey towards GPU, HBM, networking, rack and power orders.
這將是買下對模型如何被發現、評估與部署的影響力——而這些環節正是軟體選擇開始轉化為 GPU、HBM、網路設備、機櫃與電力訂單的起點。
Poolside offers the model-making process. Nemotron supplies open weights. Hugging Face provides distribution. Perplexity offers application exposure. Capital investments and lease guarantees help convert demand into AI-factory construction.
Poolside 提供模型製作流程；Nemotron 供應開放權重；Hugging Face 負責分發；Perplexity 則帶來應用曝光。資本投資與租賃擔保有助於將需求轉化為 AI 工廠的建設。
Whether that is worth 86 times revenue remains uncertain. This is not a conventional software valuation. It is the price of a potential control point.
此舉是否值得以營收的 86 倍定價，仍屬未知。這並非傳統的軟體估值，而是潛在控制點的價格。
The measure of success will not be how much subscription revenue Nvidia extracts from Hugging Face. It will be whether the company can quietly guide more workloads towards CUDA, NIM, TensorRT, NVLink and Nvidia’s AI factories without undermining the platform’s openness.
成功的衡量標準不在於 Nvidia 能從 Hugging Face 抽取多少訂閱收入，而在於該公司能否在不損害平台開放性的前提下，悄然引導更多工作負載流向 CUDA、NIM、TensorRT、NVLink 以及 Nvidia 的 AI 工廠。
Nvidia may appear to be buying a cabinet containing millions of models. What it is really buying is the parts catalogue and the map of the shop.
Nvidia 看似要買下藏有數百萬模型的櫃子，但實際上它購買的是零件目錄與店鋪地圖。
Developers will remain free to choose what they want. Nvidia, however, may learn what they are about to choose before anybody else does.
開發者仍可自由選擇所需內容，但輝達或許能比任何人都更早掌握他們即將做出的選擇。
Thanks for reading SEMIVISION @_@! This post is public so feel free to share it.
感謝閱讀 SEMIVISION @_@！本文為公開文章，歡迎自由分享。
Share分享
Previous上一頁