BEIJING — In the quiet way that tectonic shifts announce themselves, Broadcom disclosed two deals this week that together represent the most significant structural challenge to Nvidia's dominance of AI computing infrastructure.

The first: a long-term agreement to develop and produce Google's custom tensor processing units — TPUs — through 2031. [1] The second: a deal to supply Anthropic with access to approximately 3.5 gigawatts of TPU-based AI computing capacity, built on Broadcom-manufactured chips, beginning in 2027. [2]

The numbers deserve unpacking. Google's $180 billion in planned AI capital expenditure for 2026 flows partly through this deal. [3] Anthropic, now operating at an annualized revenue run rate of $30 billion according to Bloomberg, has committed to what one analyst described as "the largest single compute procurement in AI startup history." [2] The 3.5 gigawatts of capacity is not a metaphor — it is enough electricity to power a midsize American city, dedicated entirely to training and running frontier AI models.

What makes these deals strategically significant is not their size but their architecture. Nvidia's business model depends on being the default chip for AI training. Nearly every frontier lab — OpenAI, Anthropic, Google DeepMind, Meta AI — has trained its largest models on Nvidia's H100 and B200 GPUs. The Broadcom-Google partnership creates an alternative stack: Google designs the TPU, Broadcom manufactures it, and the resulting chips go not only to Google's own data centers but to third parties like Anthropic. [4]

Reuters reported the deal straightforwardly: Broadcom "signed a long-term deal to develop Google's custom AI chips." [1] CNBC went further, noting that the Anthropic component was an "expanded deal" that positioned Broadcom-manufactured TPUs as a direct alternative to Nvidia hardware for frontier model training. [5]

Broadcom's stock gained over 3% on the announcement. BofA reiterated its buy rating with a $450 price target, citing the Google and Anthropic deals as evidence that Broadcom's custom silicon business is structurally undervalued. [6]

The deeper question is whether this creates a real alternative or merely a second source. Nvidia's CUDA software ecosystem — the programming framework that makes its chips useful — is a moat that hardware specs alone cannot cross. Google's TPUs run on a different software stack (JAX and TensorFlow), and Anthropic's willingness to commit to that stack at scale suggests the company believes the software gap is closable.

For Nvidia, the threat is not immediate. Its backlog stretches years. But the Broadcom-Google-Anthropic axis establishes something that did not exist before: a vertically integrated alternative to Nvidia for frontier AI training, backed by the world's largest AI infrastructure spender (Google) and the world's fastest-growing AI company (Anthropic).

The chip war within the AI war has its first serious front.

-- DAVID CHEN, Beijing