Anthropic-Google TPU deal
Last reviewed
Jun 3, 2026
Sources
13 citations
Review status
Source-backed
Revision
v1 · 1,585 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Jun 3, 2026
Sources
13 citations
Review status
Source-backed
Revision
v1 · 1,585 words
Add missing citations, update stale details, or suggest a clearer explanation.
The Anthropic-Google TPU deal is a set of escalating compute and investment agreements between the AI lab Anthropic and Google Cloud, built around Anthropic running Claude on Google's tensor processing units (TPUs). The arrangement grew through several steps. In October 2025 the two companies committed to up to about 1 million TPUs and well over a gigawatt of capacity coming online in 2026. In April 2026 they expanded it sharply: a three-way deal with chip supplier Broadcom added multiple gigawatts of next-generation TPU capacity from 2027, and Google agreed to invest up to about $40 billion in Anthropic. Reported figures put the combined commitment in the tens of billions of dollars, with one widely cited estimate of roughly $200 billion in Anthropic spending on Google Cloud over five years.[1][2][3][4]
The agreement matters because it is one of the largest deployments of non-NVIDIA silicon for training and serving a frontier AI model, and because Google is both a major investor in Anthropic and a direct competitor through its own Gemini models.
The partnership came in distinct stages, and the numbers attached to each are easy to confuse. The table below separates them.
| Milestone | Date | Reported terms |
|---|---|---|
| Initial TPU expansion | October 23, 2025 | Up to ~1 million Google TPUs; well over 1 GW of capacity online in 2026; value in the "tens of billions" of dollars[1][3] |
| Google-Broadcom expansion | April 6-7, 2026 | Multiple gigawatts of next-generation TPU capacity from 2027; Broadcom SEC filing specifies ~3.5 GW, contingent on Anthropic's commercial performance[5][6][7] |
| Google equity investment | April 24, 2026 | Up to ~$40 billion (about $10 billion cash now, ~$30 billion tied to milestones); 5 GW of Google Cloud capacity over five years[2][8] |
| Reported five-year spend | May 5, 2026 | ~$200 billion committed by Anthropic to Google Cloud over five years, beginning 2027 (reported by The Information)[4] |
The October 2025 announcement was the headline event for the TPU count itself. Anthropic said it would access "up to one million TPUs" and that "well over a gigawatt of capacity" would come online in 2026.[3] Neither company disclosed an exact price, but Google confirmed to CNBC that the deal was worth tens of billions of dollars.[1] Industry estimates put a one-gigawatt AI data center at roughly $50 billion all in, with a large share of that going to the chips themselves, which is one way analysts arrived at the tens-of-billions framing.[9]
The April 2026 steps changed the scale again. On April 6, Anthropic, Google, and Broadcom announced that Anthropic had committed to "multiple gigawatts" of next-generation TPU capacity expected to come online starting in 2027.[5] A Broadcom securities filing then put a number on it: roughly 3.5 gigawatts of Google TPU capacity from 2027, on top of the 1 GW already arriving in 2026, with Broadcom also agreeing to design and supply future TPU generations and networking components for Google through 2031.[6] Broadcom's filing noted that Anthropic's use of the expanded capacity is contingent on its continued commercial performance.[6]
On April 24, Google said it would invest up to $40 billion in Anthropic in a mix of cash and compute. About $10 billion arrives upfront as cash at a $350 billion valuation, the same valuation Anthropic carried in a February 2026 round, with the remaining roughly $30 billion contingent on Anthropic hitting certain performance targets.[2][8] As part of the same arrangement, Google Cloud agreed to deliver 5 gigawatts of TPU-based capacity over five years, with room to add more.[8]
Anthropic has run workloads on Google's TPUs since 2023, and Google Cloud CEO Thomas Kurian framed the 2025 expansion as a vote of confidence in the hardware. "Anthropic's choice to significantly expand its usage of TPUs reflects the strong price-performance and efficiency its teams have seen with TPUs for several years," he said.[3] The latest generation, TPU Ironwood, is purpose-built for the kind of large transformer training and inference that Claude requires.
Anthropic does not rely on a single chip. The company trains and serves Claude across three platforms: Google TPUs, Amazon Trainium (including the Project Rainier supercluster of hundreds of thousands of chips with Amazon), and NVIDIA GPUs.[3] Anthropic CFO Krishna Rao described the strategy as deliberate rather than opportunistic: "Anthropic and Google have a longstanding partnership and this latest expansion will help us continue to grow the compute we need to define the frontier of AI."[3] Spreading across vendors gives Anthropic negotiating leverage, supply resilience, and the ability to route each workload to whichever silicon is cheapest for it. The approach stands in contrast to rivals who concentrate enormous commitments on a single architecture.
For Google Cloud, the deal is a marquee anchor tenant for its custom silicon. NVIDIA GPUs still dominate AI training, so every additional gigawatt that a frontier lab runs on TPUs strengthens Google's argument that its chips are a credible alternative rather than an in-house curiosity. The Broadcom partnership reinforces that: Broadcom acts as the implementation partner that turns Google's TPU designs into manufacturable chips and supplies the surrounding networking and packaging.[6][7]
The financial structure also de-risks Google's own capital spending. By tying a large investment in Anthropic to a commitment that Anthropic spend heavily on Google infrastructure, Google effectively guarantees demand for the data centers it is building. The Information reported on May 5, 2026 that Anthropic had committed to spend roughly $200 billion on Google Cloud over five years beginning in 2027, a figure that, if accurate, would make Anthropic responsible for a large share of Google's disclosed cloud revenue backlog.[4] That creates what several commentators called a circular arrangement, where Google invests in Anthropic and Anthropic spends much of that money back on Google.[4]
NVIDIA's stock reaction to the April news was muted. Analysts read the expanded TPU commitment as incremental validation of custom silicon rather than evidence of weakening GPU demand, since Anthropic continues to use NVIDIA chips alongside TPUs and Trainium.[10]
The deal fits Anthropic's broader push to lock in compute well ahead of demand. The company's annualized revenue run rate climbed from about $9 billion at the end of 2025 to more than $30 billion by April 2026, and the number of business customers spending over $1 million a year more than doubled in the early months of 2026.[5] That growth is what Anthropic points to when it justifies multi-gigawatt commitments: the capacity is meant to both train next-generation Claude models and serve a rapidly expanding customer base.[5] The vast majority of the new infrastructure is being sited in the United States.[5][6]
The compute is also tied to Anthropic's model roadmap. Reporting around the April investment noted that Anthropic had not yet released its most powerful planned model, Claude Mythos, and the TPU capacity is part of how it expects to train and deploy frontier systems through the rest of the decade.[11]
Coverage treated the deal as a landmark for both companies. Data Center Dynamics described the October 2025 agreement as confirmation of a "massive" 1 GW-plus cloud deal with up to 1 million TPUs.[12] Several analysts focused on the unusual position of Google investing in a company it competes with. A Motley Fool analysis argued the April investment looked cheap for Google, noting that Anthropic had roughly tripled its revenue in four months and that the $10 billion came in at the same $350 billion valuation employees had recently been willing to hold rather than sell at.[11]
Skeptics raised the circularity of large AI-infrastructure deals, where chipmakers, cloud providers, and model labs invest in one another and book each other's spending as revenue. The $40 billion investment and the reported $200 billion spend commitment became a frequently cited example of that pattern, alongside similar arrangements involving OpenAI and its suppliers.[4][13] Supporters counter that the commitments reflect genuine demand for Claude and a real bet that custom silicon can compete with NVIDIA at scale.