Anthropic Inks Historic AI Compute Agreement, Sidestepping Nvidia
In a move that reshapes the artificial intelligence infrastructure landscape, Anthropic has finalized one of the largest compute deals in AI history, with the vast majority of its investment flowing to Google and Broadcom rather than Nvidia. The AI laboratory has committed to multiple gigawatts of next-generation Tensor Processing Unit (TPU) capacity from Google and Broadcom, in an agreement that could be valued at hundreds of billions of dollars, commencing in 2027.
This monumental deal represents a direct and large-scale endorsement of Google's in-house silicon, occurring at a pivotal moment when Nvidia's long-standing dominance in AI infrastructure faces its most significant challenge to date. Anthropic will gain access to nearly 5 gigawatts of new computing capacity over the coming years, with each gigawatt estimated to cost between $35 billion and $50 billion to develop, primarily allocated to chip expenditures.
Google's TPUs Gain a Major External Customer
For much of the past decade, Google's Tensor Processing Units were often overlooked as a serious competitor to Nvidia, which commands over 90% of the AI chip market. However, this perception is rapidly changing. Anthropic has been utilizing Google TPUs since 2023, and in October of last year, it announced a deal for more than a gigawatt of Google computing power, equivalent to up to one million TPUs. The new agreement significantly expands this commitment.
Broadcom plays a crucial role in this partnership by transforming Google's TPU designs into manufacturable chips, effectively enabling other companies to leverage Google's internal silicon at scale. Morgan Stanley analysts project that every 500,000 TPUs sold to external customers could generate approximately $13 billion in revenue for Google, positioning Anthropic as the largest external customer in this ecosystem.
Nvidia's Market Position Under Scrutiny
This is not the first instance where TPU momentum has unsettled Nvidia. Late last year, reports surfaced that Meta was considering using Google TPUs in its data centers, causing Nvidia's stock to drop over 3%. Nvidia responded by asserting its technological leadership, but Meta ultimately renewed its relationship with Nvidia while also developing its own in-house MTIA chips.
Anthropic's strategy diverges from Meta's approach of balancing Nvidia purchases with proprietary chip development. Instead, Anthropic is making a substantial, long-term commitment to third-party chips, a decision that is difficult to reverse. Concurrently, Reuters reports that Anthropic is exploring the design of its own chips, following a path similar to Meta, OpenAI, Amazon, and Google.
The Inference Era Challenges Nvidia's Dominance
Nvidia established its supremacy in AI model training, where large models are built on extensive GPU clusters. However, the industry has transitioned into the inference era, where models are deployed for daily use by hundreds of millions of users. This shift demands different economic considerations, favoring cheaper, specialized chips.
Google's latest Ironwood TPU reportedly offers a total cost of ownership roughly 30–44% lower than Nvidia's equivalent GB200 Blackwell server. Bank of America analysts estimate that inference will constitute 75% of AI data center spending by 2030, up from about 50% last year, prompting Google, Amazon, Microsoft, and Broadcom to develop chips tailored for this market.
Nvidia has responded by unveiling an inference-focused chip at GTC, leveraging technology from Groq, a startup it licensed for $17 billion in December. This marks a departure from Nvidia's traditional stance that its GPUs are equally effective for both training and inference.
Market Dynamics and Future Implications
Despite these developments, Nvidia remains a formidable player. Anthropic continues to maintain compute commitments with Nvidia and Microsoft Azure alongside this Google deal, and Nvidia's projected $1 trillion revenue opportunity through 2027 remains viable. However, every gigawatt allocated to Google TPUs represents capacity diverted from Nvidia.
The Broadcom-Google partnership now boasts a long-term anchor customer, a supply pact extending through 2031, and a functional commercial model. On the day of the announcement, Broadcom and Alphabet's stock rallies reflected the market's recognition of this strategic shift, signaling a potential reconfiguration of AI chip procurement and competition in the years ahead.
