Nvidia is moving beyond chips into AI model gateways, while Chinese models are expanding their global reach, as competition shifts to other areas
輝達正從晶片領域拓展至人工智慧模型網關，而中國模型則加速擴大其全球影響力，競爭重心正轉向其他領域
Chip giantNvidia's(NVDA.US) new agreement to acquireHugging Facefor $12.93 billion is much more than the purchase of a website that hosts AI models. Nvidia has long dominated the AI chip market and the Cuda software ecosystem, while Hugging Face connects more than 18 million developers and about 200,000 companies with over 3 million models. If completed, the transaction will take Nvidia beyond its roots in computing infrastructure into model discovery, testing, downloading and deployment, bringing it closer to the gateway where companies decide which models to use and where to run them.
芯片巨頭輝達（NVDA.US）新近達成協議，以$129.3億收購Hugging Face，此舉遠不止是收購一個託管AI模型的網站。輝達長期以來主導着AI芯片市場和CUDA軟件生態系統，而Hugging Face連接了超過1,800萬名開發者以及約200,000家公司，平台上擁有超過300萬個模型。若交易完成，該收購將推動輝達超越其在計算基礎設施領域的根基，進入模型發現、測試、下載和部署環節，使其更接近企業決定使用那些模型以及在何處運行這些模型的網關。
Nvidia's move is not isolated.Microsoft(MSFT.US) offers more than 11,000 models through Microsoft Foundry, whileAmazon's(AMZN.US) Bedrock andAlphabet's(GOOG.US) Google Model Garden also combine models, development tools and cloud services into enterprise deployment platforms.Meta(META.US), meanwhile, is using its open-weight Llama models to expand its developer ecosystem. The technology giants are pursuing different strategies, but the competition has expanded beyond model capabilities to developers, computing power and enterprise deployment gateways.
輝達的這一舉措並非孤例。微軟（MSFT.US）通過Microsoft Foundry提供超過11,000個模型，而亞馬遜（AMZN.US）的Bedrock和Alphabet（GOOG.US）的Google Model Garden也將模型、開發工具和雲服務整合爲企業部署平台。與此同時，Meta（META.US）正利用其開源權重的Llama模型擴展其開發者生態系統。這些科技巨頭採取了不同的戰略，但競爭已不僅侷限於模型能力本身，還延伸至開發者、算力以及企業部署網關。
Chinese open-weight models are emerging as key players in this shift. A report released by Hugging Face in March showed that models developed in China accounted for 41% of downloads on its platform over the past year, surpassing the share of U.S. models. The Qwen family of models fromAlibaba(9988.HK; BABA.US) has spawned more than 113,000 derivatives, exceeding the combined total for Google and Meta. Models includingDeepSeek, Qwen, GLM fromZ.AI(2513.HK), Kimi fromMoonshot AIand those fromMiniMax(0100.HK) are increasingly being used by developers for fine-tuning, compression and redeployment, with some beginning to serve as the technical backbone of overseas projects.
中國開源權重模型正在這一轉變中嶄露頭角。Hugging Face今年3月發佈的一份報告顯示，在過去一年中，中國開發的模型在其平台上的下載量佔比達41%，超過了美國模型的份額。阿里巴巴（9988.HK；BABA.US）的Qwen系列模型已衍生出超過113,000個變體，數量超過谷歌和Meta的總和。包括DeepSeek、Qwen、Z.AI（2513.HK）的GLM、月之暗面（Moonshot AI）的Kimi以及MiniMax（0100.HK）等公司的模型正越來越多地被開發者用於微調、壓縮和重新部署，其中一些已開始作爲海外項目的技術骨幹。
Chinese companies primarily provide foundation models, Hugging Face connects models with developers, and Nvidia supplies chips and development tools. The three do not compete directly, but collectively influence how companies adopt and deploy models, while everyone seeks to capture developers, customer relationships and revenue within the industry's value chain.
中國企業主要提供基礎模型，Hugging Face將模型與開發者連接起來，而輝達則提供芯片和開發工具。這三方並不直接競爭，但共同影響着企業如何採用和部署模型，同時各方都在努力在行業價值鏈中獲取開發者、客戶關係和收入。
From downloads to overseas projects
從下載量到海外項目
Humain, an AI company under Saudi Arabia's Public Investment Fund, recently unveiled the Arabic-language humain-m3 model. Built on MiniMax-M3, the model was further pre-trained on more than 1 trillion native Arabic tokens and has 428 billion parameters. It is currently available in research preview on the Humain Node. Humain plans to make the model weights available after completing safety training and alignment.
沙烏地阿拉伯阿拉伯公共投資基金旗下的AI公司Humain近期推出了阿拉伯語humain-m3模型。該模型基於MiniMax-M3構建，並在超過1兆個原生阿拉伯語詞元上進行了進一步預訓練，參數量達4,280億。目前該模型已在Humain Node上提供研究預覽版。Humain計劃在完成安全訓練和對齊後開放模型權重。
For Saudi Arabia, adopting an existing Chinese model as its foundation can help Humain shorten development time and focus its resources on Arabic-language data, localized applications and enterprise services. It also provides Chinese AI companies with another path into overseas markets. However, Humain still controls the local platform and customer relationships, while cloud service providers and chipmakers can also generate revenue from deployment demand. How much value MiniMax can capture will depend on licensing, joint development and ongoing service arrangements.
對沙烏地阿拉伯阿拉伯而言，採用現有的中國模型作爲基礎有助於Humain縮短開發周期，並將資源集中於阿拉伯語數據、本地化應用和企業服務。這也爲中國AI公司開闢了另一條進入海外市場的路徑。然而，Humain仍掌控本地平台和客戶關係，而云服務提供商和芯片製造商也可從部署需求中獲得收入。MiniMax能捕獲多少價值，將取決於授權許可、聯合開發和持續服務安排。
Nvidia's revenue rose 106% year-on-year to $96.2 billion in its latest quarter, including $89 billion in data center revenue. But major customers, including Microsoft, Amazon, Google and Meta, are all developing their own AI chips. Acquiring Hugging Face would give Nvidia earlier access to demand for model development and deployment, while allowing it to reinforce its hardware advantage through software and inference services. Nvidia has pledged that Hugging Face will remain open and continue supporting different clouds and computing hardware, without requiring users to adopt Nvidia chips. Still, combining the two companies could bring more development activity into Nvidia's software and computing ecosystem.
輝達最新一季度營收同比增長106%，達到$962億，其中包括$890億的數據中心收入。但包括微軟、亞馬遜、谷歌和Meta在內的主要客戶都在開發自己的AI芯片。收購Hugging Face將使輝達更早接觸到模型開發和部署的需求，同時通過軟件和推理服務強化其硬件優勢。輝達已承諾Hugging Face將繼續保持開放，並繼續支持不同的雲平台和計算硬件，不要求用戶必須採用輝達芯片。儘管如此，兩家公司的結合可能會將更多開發活動引入輝達的軟件和計算生態系統。
Turning influence into revenue
將影響力轉化爲收入
Chinese companies have begun trying to generate revenue from application programming interfaces, private deployments for enterprises, model fine-tuning, AI agent tools and cloud computing. Alibaba has a relatively complete commercial ecosystem, with Qwen expanding its developer base and Alibaba Cloud providing the computing power and enterprise services needed to run the models.
中國企業已開始嘗試通過應用程序編程接口（API）、企業私有化部署、模型微調、AI智能體工具和雲計算來創造收入。阿里巴巴擁有相對完整的商業生態系統，其中通義千問（Qwen）不斷擴大其開發者基礎，阿里雲則提供運行這些模型所需的算力和企業服務。
Alibaba's revenue from AI-related products reached 12.4 billion yuan ($1.85 billion) in the quarter through June, marking its 12th consecutive quarter of triple-digit growth. While the company didn't disclose how much of that revenue came directly from Qwen, the complete value chain formed by its models, cloud computing and enterprise services provides a clearer path to monetization.
截至6月的本季度，阿里巴巴來自人工智能相關產品的收入達到124億元（$18.5億），標誌着其連續第12個季度實現三位數增長。儘管該公司未披露其中有多少收入直接來自通義千問（Qwen），但其模型、雲計算和企業服務所形成的完整價值鏈爲商業化提供了更清晰的路徑。
The challenge is greater for independent model developers such as MiniMax, Z.AI and Moonshot AI. Open weights can help them rapidly grow usage, derivative models and brand influence, but training, inference and overseas services remain costly. If developers download their models and deploy them on other cloud platforms, these companies may gain little beyond name recognition, while chipmakers, cloud platforms and systems integrators capture the revenue. Overseas licensing, joint development and enterprise deployment projects therefore provide a better test of their business models than download figures alone.
對於MiniMax、Z.AI和Moonshot AI等獨立模型開發商而言，挑戰更大。開放模型權重有助於它們快速擴大使用量、衍生模型數量和品牌影響力，但訓練、推理和海外服務仍然成本高昂。如果開發者下載它們的模型並在其他雲平台上部署，這些公司可能除了獲得知名度外收益甚微，而芯片製造商、雲平台和系統集成商則攫取了大部分收入。因此，與單純的下載量相比，海外授權、聯合開發和企業部署項目更能檢驗其商業模式的有效性。
Nvidia's acquisition of Hugging Face shows that value in the AI industry is being redistributed among models, platforms, computing infrastructure and enterprise services. Chinese models have already established an advantage in downloads and developer adoption, while humain-m3 shows that they are beginning to enter large overseas projects.
輝達對Hugging Face的收購表明，人工智能行業的價值正在模型、平台、計算基礎設施和企業服務之間重新分配。中國模型在下載量和開發者採用方面已建立優勢，而humain-m3項目表明它們正開始進入大型海外項目。
The next test is whether model developers can retain control of customer relationships and revenue as platforms, computing providers and cloud giants compete to capture value. Download figures can demonstrate influence, but only sustained spending by enterprise customers can translate such influence into a durable business.
下一個考驗在於，當平台、計算提供商和雲巨頭競相爭奪價值時，模型開發商能否保持對客戶關係和收入的控制權。下載量可以體現影響力，但只有企業客戶的持續支出才能將這種影響力轉化爲可持續的業務。 (Benzinga)