2025.12. 18 这场由巨头博弈引发的震荡，将一个核心议题推至台前：在以大模型为核心的AI时代，硬件的技术范 式是否正在从通用GPU转向专用芯片如TPU？这是否意味着一场结构性的变革已然来临？ 本文字数：1632，阅读时长大约3分钟 作者 | 第一财经 刘佳 这一悬念不仅关乎国际巨头的战略布局，也紧密牵动着中国AI算力产业链的神经。作为对标英伟达、 不久前刚刚上市的中国GPU厂商代表，摩尔线程创始成员、摩尔学院院长李丰与沐曦高级副总裁孙 国梁在今日腾讯contech大会上"同框"，并回应了对于两种路线的看法。 在李丰看来，争议背后其实是"通才与专才"的分工，而非简单的替代关系。 他分析，谷歌能做TPU，本质上是因为它是全栈整合公司。谷歌有强大的 Infra、基础模型与云服务 形成闭环，把模型跑在自家芯片上量身优化，实现成本性价比的最大化。"但绝大部分企业不具备这 样的垂直整合能力。" 他总结，GPU持续保持优势的原因有三个：灵活度是"甜点"、多模态时代的全功能性、生态的护城 河。 谷歌新一代AI模型Gemini 3系列的发布，在硬件领域投下一颗"重磅炸弹"——其自研TPU（张量处 理器）所展现的性能与成 ...

谷歌新一代AI模型Gemini 3系列的发布，在硬件领域投下一颗"重磅炸弹"——其自研TPU（张量处理 器）所展现的性能与成本优势，被外界解读为对英伟达GPU（图形处理器）霸主地位的强劲挑战。资本 市场反应迅速，英伟达市值一度蒸发超千亿美元。 生态更是关键壁垒，英伟达之所以在计算领域成为"王者"，得益于它所建立起的CUDA生态，能够联合 所有开发者建设生态。李丰称，摩尔线程也在搭建自己MUSA的生态。 沐曦创始人孙国梁的观点更加直接：任何芯片架构都没有高低优劣之分，关键还是看场景。 他认为，GPU和ASIC（专用芯片）的架构几十年前就存在，已经是超级稳态。而现在的大模型太卷 了，迭代速度非常快，达到了按周计、按月计的程度，任何基础模型远未到达收敛的时间点，通用 GPU 的泛化能力和适配性仍是核心竞争力，现在业界还很难把一个专用性的产品放在通用的场景里。 但与此同时，客户是多元化的，应用场景分散、层出不穷，因此GPU和类似TPU这样的ASIC会长期共 存。 这场由巨头博弈引发的震荡，将一个核心议题推至台前：在以大模型为核心的AI时代，硬件的技术范 式是否正在从通用GPU转向专用芯片如TPU？这是否意味着一场结构 ...

Group 1: Federal Reserve and Market Dynamics - The market is anticipating a significant shift towards interest rate cuts, with the probability of a December rate cut rising from below 50% to over 80% according to CME's FedWatch tool [5] - San Francisco Fed President Mary Daly highlighted the tightness of the U.S. job market, suggesting a potential spike in unemployment, while inflation appears to be temporarily subdued [3][5] - The absence of key economic data, such as Q3 GDP and September PCE index, creates uncertainty for the Federal Reserve, likening their current situation to "blind flying" [5][7] Group 2: Global Economic Concerns - Concerns are growing over the global liquidity situation, with warnings from Bank of America that the numerous rate cuts by central banks this year may soon reach their limits [7] - The decline in Japan's 30-year government bond prices and the weakening yen could trigger a significant outflow of capital, impacting global equity and bond markets [7] Group 3: Company Insights - Broadcom has seen a 60% increase in stock price this year, driven by its role in supplying core components for Google's AI initiatives, positioning it as a key player in the AI supply chain [8][10] - The tech sector is experiencing a surge in bond issuance, with major companies like Meta, Google, and Oracle contributing to a significant increase in overall debt levels [11][13] Group 4: Investment Strategies - Investors are advised to focus on "interest rate sensitive" assets, such as long-term government bonds and fundamentally strong mid-cap stocks, if rate cuts materialize [14][15] - Companies deeply integrated into the AI supply chain, like Broadcom, are seen as solid investment opportunities due to their stable order flow from major tech players [15] - High-leverage companies that rely on debt to finance operations should be approached with caution, as rising CDS prices indicate increasing concerns over default risks [13][17]

"未来OpenAI希望专注于直接与英伟达合作构建计算基础设施。"科莱特·克雷斯表示，英伟达与 Anthropic的合作目前是通过云服务厂商来进行的，但Anthropic不仅对云服务厂商提供算力这个方式感 兴趣，还考虑未来与英伟达直接达成1吉瓦算力的合作。 当地时间12月2日，英伟达参加了瑞银全球技术与AI大会。英伟达CFO科莱特·克雷斯（Colette Kress） 接受了瑞银分析师提问并谈及英伟达与OpenAI等大模型公司的合作关系。 9月，英伟达宣布向OpenAI投资最多1000亿美元，英伟达将帮助OpenAI部署至少10吉瓦的数据中心。关 于双方合作的细节，科莱特·克雷斯表示，英伟达与OpenAI的协议是一种超过十年的合作关系。今年及 明年英伟达看到Blackwell及Rubin系统的需求约5000亿美元，其中包括OpenAI通过云服务提供商来满足 计算需要。但这5000亿美元不包括英伟达与OpenAI协议下一阶段进行的工作。英伟达和OpenAI尚未敲 定最终的协议。 宣布投资OpenAI之后，英伟达又宣布投资大模型厂商Anthropic，合作也涉及Anthropic采购计算资源。 两个合作背后，科莱特 ...

Core Insights - Google's technological strength and strategic outlook are being rapidly re-evaluated by the market following the release of its next-generation flagship AI model, Gemini 3 [1] - Google is in negotiations with Meta for a large-scale TPU supply deal, which could disrupt NVIDIA's core customer base in the AI chip sector [1][2] - The release of Gemini 3 has garnered overwhelmingly positive reviews, positioning Google as a strong competitor in the AI landscape [5][6] TPU Supply Negotiations - Google is negotiating a multi-billion dollar TPU deal with Meta, marking a significant shift in strategy as it aims to enter the AI chip device market dominated by NVIDIA [2][4] - Meta plans to integrate Google's TPU into its global data centers starting in 2027, with initial TPU capacity rental from Google Cloud expected as early as next year [4] - If successful, this partnership could lead to a structural shift in NVIDIA's core customer relationships, as Meta has historically relied on NVIDIA GPUs [4] Gemini 3 Model Performance - The Gemini 3 model has received near-unanimous positive feedback, with claims of surpassing OpenAI's GPT-5.1 in key tasks such as code generation and logical reasoning [5][6] - Salesforce CEO Marc Benioff expressed a strong preference for Gemini 3 over ChatGPT after testing the model [5] - OpenAI's CEO acknowledged the strength of Gemini 3, indicating that the company is working to catch up in the competitive landscape [6] Financial Performance and Market Reaction - Google's third-quarter earnings exceeded expectations, with a 35% increase in earnings per share to $2.87 and a 16% rise in total revenue to $102.35 billion [6] - The company has raised its capital expenditure forecast for 2025 by 8% to $92 billion, primarily for AI data center and cloud computing expansion [6] - Following these developments, A-share stocks related to Google's supply chain have seen significant gains, with companies like Dekoli and Guangku Technology experiencing notable stock price increases [7][9] Market Outlook - Analysts suggest that Google's comprehensive technology ecosystem, from chips to applications, creates a strong AI moat and drives continued capital expenditure growth [9] - The recent shift in investment strategies, such as Berkshire Hathaway selling Apple shares to buy Google, reflects confidence in Google's position in the AI industry [9] - The performance of Gemini 3, trained on self-developed TPU, is expected to enhance ROI and establish a commercial loop, positioning Google as a leader in the ongoing AI revolution [9]

Core Insights - The construction of data centers in space is becoming a viable solution for addressing the energy supply constraints faced by terrestrial data centers, with predictions indicating that energy demand for U.S. data centers will nearly double by 2027 [1][3] Group 1: Industry Trends - Major tech companies, including Google and SpaceX, are exploring the feasibility of building scalable machine learning computing systems in space, with Google's "Suncatcher" initiative leading the charge [3][5] - SpaceX plans to expand its Starlink V3 satellite capabilities to facilitate the construction of data centers in space, while Jeff Bezos anticipates that within the next 10 to 20 years, humans will be able to build gigawatt-scale data centers in space [3][4] Group 2: Technological Advancements - Starcloud is set to launch a satellite equipped with NVIDIA H100 GPUs, marking the first instance of advanced data center GPUs being deployed in space, with the satellite expected to provide 100 times the GPU computing power of previous space computing facilities [4] - The potential for unlimited low-cost renewable energy in space is highlighted as a significant advantage, with Starcloud's data center projected to save 10 times the carbon dioxide emissions compared to terrestrial data centers [4][5] Group 3: Future Projections - Industry experts predict that within the next decade, space could emerge as a primary location for new data centers, with the cost of building these facilities expected to decrease significantly [6][7] - Historical data suggests that by the mid-2030s, launch costs could drop below $200 per kilogram, making the operational costs of space data centers comparable to those of ground-based facilities [6][7]

Core Viewpoint - The article discusses the increasing energy demands of AI data centers and the potential shift towards building data centers in space as a solution to energy constraints on Earth [3][4]. Group 1: Energy Demand and Constraints - FTI Consulting predicts that energy demand for data centers in the U.S. will nearly double by 2027, leading to significant strain on utility companies and grid capacity [3]. - The construction of data centers in space is being considered by several Silicon Valley tech companies due to the limited availability of power on Earth [4]. Group 2: Initiatives by Tech Companies - Google has launched a project called "Suncatcher" to explore scalable machine learning computing systems in space, as announced by CEO Sundar Pichai [6]. - SpaceX, led by Elon Musk, plans to build data centers in space using Starlink V3 satellites equipped with high-speed laser links [7]. - Jeff Bezos has indicated that within the next 10 to 20 years, humanity will be able to construct gigawatt-scale data centers in space [7]. Group 3: Technological Developments - Google and Planet Labs are collaborating to launch two satellites in early 2027 to explore the feasibility of large-scale space data center clusters [7]. - Starcloud plans to launch a satellite carrying NVIDIA H100 GPUs, marking the first advanced data center GPUs to enter space, with a projected performance increase of 100 times compared to previous space computing facilities [7]. Group 4: Advantages of Space Data Centers - Space data centers will benefit from abundant renewable energy, eliminating the need for water cooling and backup power sources [8]. - The lifecycle carbon emissions of space data centers could be ten times lower than those of terrestrial data centers [9]. - Solar energy in space can produce eight times more output than on Earth, providing continuous power without weather interruptions [9]. Group 5: Cost Considerations and Feasibility - High launch costs have historically been a barrier to large-scale space systems, but costs may drop below $200 per kilogram by the mid-2030s, making space data centers potentially cost-competitive with terrestrial counterparts [10]. - Google has conducted preliminary studies indicating that their next-generation TPUs have strong radiation resistance, although challenges such as thermal management and system reliability remain [10].

Core Insights - Anthropic has entered a significant partnership with Google, involving a multi-billion dollar deal for AI computing resources, including up to one million TPU chips and 1 gigawatt of power capacity [1][2] - This collaboration marks a critical phase in the AI infrastructure race, with estimated costs for building a 1 gigawatt data center around $50 billion, primarily for chip procurement [1] - Anthropic's rapid growth is evident, with annual revenue nearing $7 billion and a substantial increase in its customer base, particularly large clients [3] Company Developments - Anthropic's partnership with Google is built on a solid foundation of previous collaboration, including a $300 million investment from Google in February 2023 [2] - Despite deepening ties with Google, Amazon remains Anthropic's largest investor, with a total investment of $8 billion [2] - Anthropic's Claude product line has seen explosive growth, with a reported annual revenue of $500 million within two months of launch, making it the fastest-growing product in the company's history [3] Industry Context - The AI sector is witnessing a surge in high-value transactions, with concerns about a potential investment bubble reminiscent of the internet bubble in the early 2000s [4] - Analysts suggest that the current AI boom differs fundamentally from the internet bubble, as AI is transforming production tools rather than just production venues [5] - There is a noted disparity in AI adoption among businesses, with only about 10% of small enterprises and 40% of medium to large enterprises currently utilizing AI, indicating room for deeper integration [5]

Core Viewpoint - The actual controller of Tianpu Co., Ltd. has changed from You Jianyi to Yang Gongyifan of Zhonghao Xinying, following a series of complex transactions including share transfers and capital increases [1][2]. Group 1: Share Transfer Details - The first share transfer involves Tianxing Trading, Tianpu Holdings, and You Jianyi transferring a total of 10.75% of Tianpu's total share capital to Zhonghao Xinying for a total price of 346 million yuan, at a price of 23.98 yuan per share [1]. - The second share transfer involves Puen Investment and Tianxing Trading transferring 8.00% of Tianpu's total share capital to Fang Donghui for a total price of 257 million yuan, also at a price of 23.98 yuan per share [1]. - Before the transfers, the four companies held a combined 75% of Tianpu's shares, with the remaining 25% held by minority shareholders [1]. Group 2: Capital Increase and Control - Following the share transfers, Zhonghao Xinying, Hainan Xinfan, and Fang Donghui plan to increase capital in Tianpu Holdings by 619 million yuan, 395 million yuan, and 507 million yuan respectively [2]. - After the capital increase, Zhonghao Xinying will hold 30.52%, Hainan Xinfan 19.49%, and Fang Donghui 24.99% of Tianpu Holdings, while You Jianyi will hold 25% [2]. - Zhonghao Xinying and Hainan Xinfan will collectively hold 50.01% of Tianpu Holdings, allowing Yang Gongyifan to control Tianpu Holdings and, consequently, Tianpu Co., Ltd. [2]. Group 3: Financial Overview of Zhonghao Xinying - Zhonghao Xinying is one of the few companies in China that possesses core technology for TPU (Tensor Processing Unit) training and inference architecture [3]. - The projected revenues for Zhonghao Xinying from 2022 to 2024 are 81.69 million yuan, 485 million yuan, and 598 million yuan respectively, with net profits of -42.98 million yuan, 81.33 million yuan, and 88.91 million yuan for the same years [3].