In November 2025, the global AI chip sector was rocked by a major upheaval as Google’s TPU (Tensor Processing Unit) made a powerful breakthrough, sparking intense discussions about the reshaping of industry dynamics. That month, Alphabet, Google’s parent company, saw its market value surge by approximately $530 billion, approaching the $4 trillion mark; meanwhile, NVIDIA, the long-standing hegemon in the AI chip space, witnessed a market value evaporation of $620 billion (equivalent to about 4.39 trillion RMB) and a stock price drop of nearly 12.59%. At the core of this striking divergence lies industry speculation that Meta plans to invest billions of US dollars in purchasing Google’s TPU chips in 2027. As a key customer of NVIDIA, Meta’s potential “defection” directly threatens NVIDIA’s nearly 85% market share, a development that has become a focal point in the latest AI news.
Technical Showdown: The Tug-of-War Between Specialization and Versatility
Google’s TPU has undergone seven generations of iteration over a decade, with its seventh-generation product, Ironwood, achieving a major leap forward. Equipped with 192GB HBM memory per chip, the cluster computing power reaches 42.5 exaFLOPS, and its energy efficiency is twice that of the previous generation and 2-3 times higher than that of contemporary GPUs. Tailored for deep learning training and inference, it is particularly well-suited for complex tasks such as large language models. As Google’s first externally sold TPU product, it leverages a “systolic array” core structure to deliver advantages of low latency and high cost-effectiveness in AI-specific scenarios.
NVIDIA’s GPU, on the other hand, stands out for its “versatile all-round capabilities.” Built on the CUDA platform launched in 2006, it has established a mature ecosystem with 13 million developers, supporting diverse fields including graphics rendering, scientific computing, and AI research. The two form a sharp contrast: the TPU is an “AI specialist,” focusing on segmented high-efficiency scenarios; the GPU is a “versatile all-rounder,” dominating the market with its ecological moat. In terms of business models, GPUs are sold directly as hardware to cover the entire chain from cloud vendors to enterprises and research institutions, while TPUs have primarily relied on cloud computing power leasing and now enter the market with the first external sales model.
Wall Street’s Debate: Is NVIDIA’s Moat Shaking?
Regarding the industry landscape, Wall Street has split into two camps. The “win-win faction” argues that the future AI infrastructure market, expected to reach trillions of US dollars in scale, is large enough to accommodate the coexistence and mutual success of multiple giants. Bank of America predicts that by the end of this decade, the global AI data center market will grow from $242 billion to $1.2 trillion. Even if NVIDIA’s market share drops to around 75%, it will still maintain a leading position.
The “threat faction,” however, expresses concerns about Google’s full-stack vertical integration capabilities. Its comprehensive ecosystem spanning chips, networks, and AI models could potentially form a closed yet highly efficient system. Nevertheless, NVIDIA’s CUDA ecosystem remains a core barrier. While Google has lowered the threshold for use through programming languages like JAX, dislodging CUDA’s “standard” status remains a daunting task in the short term. Currently, NVIDIA is actively responding by investing to bind potential customers such as OpenAI and Anthropic, and emphasizing its platform’s “one-generation lead in the industry” and “support for full-scenario AI computing” to counter the limitations of TPU’s specialization. Meanwhile, Google’s collaborations with Meta and Anthropic signal that TPU has become a crucial alternative for hyperscale customers, and competition in the trillion-dollar AI chip track is set to intensify further.
