英伟达并不缺少挑战者，但到目前为止，他们都很难称得上是 英伟达的对手，难以撼动其领导地位。不过，未来这一点未必 不会改变。 作者： 刘劲等 封图：图虫创意 算力是人工智能最重要的基础设施和发展引擎。AI算力的代表企业英伟达（NVIDIA）凭借性能先 进的产品和难以复制的生态，在AI训练及推理芯片领域建立起了近乎垄断的领导地位，成为地球上 价值最高的上市公司。截至2025年11月，英伟达的市值约为4.5万亿美元，2025年第三季度营收 的同比增长约为62%。 在大模型发展的初期和中期，训练算力是核心瓶颈，决定了模型的"高度"，是算力芯片的战略制 高点。因此，我们在此着重讨论训练。 英伟达在训练算力上有统治性的地位。这种优势来自两个方面：先进的技术和生态的垄断。 主流大模型的参数规模已达千亿、万亿级别，训练时要对海量数据进行大规模计算，单机算力早已 远远不够，必须依托大规模芯片集群完成训练；要令这复杂而成本高昂的训练易于展开、效率高、 稳定可靠，还需要一整套的软件系统和工具来作为连接训练工程师、算力芯片和模型的桥梁。 因此，我们大致可以将训练对算力芯片的要求拆解成单芯片性能（单卡性能）、互联能力和软件生 态三部分 ...

Core Insights - AI computing power is the most critical infrastructure and development engine for artificial intelligence, with NVIDIA establishing a near-monopoly in the AI training and inference chip market, becoming the highest-valued public company globally, with a market capitalization of approximately $4.5 trillion by November 2025 and a year-on-year revenue growth of about 62% in Q3 2025 [2] Competitive Landscape - NVIDIA faces challengers from traditional chip giants like AMD and Intel in the U.S., as well as self-developed computing power from tech giants like Google and Amazon, and emerging players like Cerebras and Groq, but none have significantly threatened NVIDIA's leadership position yet [2] - The AI computing chip market has two main application scenarios: training and inference, with training being the core bottleneck that determines the model's capabilities [3] Training Power Dominance - NVIDIA holds a dominant position in training power due to advanced technology and a monopolistic ecosystem, as training large models requires massive data computation that single-chip power cannot provide [5] - The requirements for training chips can be broken down into single-chip performance, interconnect capabilities, and software ecosystem [6] Technical Advantages - NVIDIA excels in single-chip performance, with competitors like AMD catching up in key performance metrics, but this alone does not threaten NVIDIA's lead in AI training [7] - Interconnect capabilities are crucial for large model training, and NVIDIA's proprietary technologies like NVLink and NVSwitch enable efficient interconnectivity at a scale of tens of thousands of chips, while competitors are limited to smaller clusters [8] Ecosystem Strength - NVIDIA's ecosystem advantage is primarily software-based, with CUDA being a well-established platform that enhances developer engagement and retention [8] - The strong network effect of NVIDIA's ecosystem makes it difficult for competitors to challenge its dominance, as many AI researchers and developers are already familiar with CUDA [9][10] Inference Market Dynamics - Inference requires significantly fewer chips than training, leading to reduced interconnect demands, which diminishes NVIDIA's ecosystem advantage in this area [11] - Despite this, NVIDIA still holds over 70% of the inference market share due to its competitive performance, pricing, and overall value proposition [11] Challenges to NVIDIA - Competitors must overcome both technical and ecosystem barriers to challenge NVIDIA, with options including significant technological advancements or creating protective market conditions [13] - In the U.S., challengers are primarily focused on technological advancements, such as Google's TPU, while in China, the market has become "protected" due to U.S. export bans on advanced chips [16] Geopolitical Implications - The U.S. government's restrictions on NVIDIA's chip sales to China have created a challenging environment for Chinese AI firms, but also present significant opportunities for domestic chip manufacturers [17] - The recent shift in U.S. policy allowing NVIDIA to sell advanced H200 chips to China under specific conditions indicates a recognition of the need to maintain NVIDIA's competitive edge while managing geopolitical tensions [19] Strategic Considerations - The competition in AI technology should not solely focus on domestic replacement strategies, as this could lead to a cycle of technological isolation [20] - Huawei's decision to open-source its CANN and Mind toolchain reflects a strategic move to build a competitive ecosystem that can attract global developer participation [21]

Core Insights - Alphabet's CEO Sundar Pichai faces declining stock prices, prompting Nvidia to assert its industry leadership, emphasizing the superiority of GPUs over Google's TPU technology [2] - Berkshire Hathaway's investment in Alphabet marks a significant shift, coinciding with Meta's consideration of deploying Google's TPU in its data centers by 2027 [2] - Google continues to collaborate with Nvidia, highlighting its commitment to supporting both TPU and Nvidia's GPU technologies [2] TPU Development History - The TPU project was initiated in 2015 to address the unsustainable power consumption of Google's data centers due to the increasing application of deep learning [3] - TPU v1 was launched in 2016, proving the feasibility of ASIC solutions for Google's core services [4] - Subsequent versions (v2, v3) were commercialized, with TPU v4 introducing a supernode architecture that significantly enhanced performance [5][6] Transition to Commercialization - TPU v5p marked a turning point, entering Google's revenue-generating products and doubling performance compared to v4 [6][7] - The upcoming TPU v6 focuses on inference, aiming to become the most cost-effective commercial engine in the inference era, with a 67% efficiency improvement over its predecessor [7][8] Competitive Landscape - Google, Nvidia, and Amazon are at a crossroads in the AI chip market, each pursuing different strategies: Nvidia focuses on GPU versatility, Google on specialized TPU efficiency, and Amazon on cost reduction through proprietary chips [19][20][22] - Google's TPU strategy emphasizes vertical integration and system-level optimization, contrasting with Nvidia's general-purpose GPU approach [21][22] Cost Advantages - Google's vertical integration allows it to avoid the "CUDA tax," significantly reducing operational costs compared to competitors reliant on Nvidia GPUs [26][27] - The TPU service enables Google to offer lower-priced inference capabilities, attracting businesses to its cloud platform [27][28] Strategic Importance of TPU - TPU has evolved from an experimental project to a critical component of Google's AI infrastructure, contributing to a significant increase in cloud revenue, which reached $44 billion annually [30][31] - Google's comprehensive AI solutions, including model training and monitoring, position it favorably against AWS and Azure, enhancing its competitive edge in the AI market [32]