Expanding Our Use of Google Cloud TPUs and Services

Source: Anthropic Newsroom (official) — primary corporate announcement Published: 2025-10-23 (capacity lands in 2026) | Ingested: 2026-06-24 Authority: technical-report (first-party company announcement). The deal terms are stated by a direct party (strong); business-scale metrics are self-reported.

Summary

Anthropic's commitment to ~1 million Google TPUs and well over a gigawatt of capacity online in 2026 is the cleanest demand-side evidence for the custom-silicon inflection the KB tracks. It shows that a frontier lab is committing gigawatt-scale capital to a hyperscaler ASIC (Google's Ironwood TPU) — not to merchant NVIDIA GPUs — while explicitly keeping a multi-vendor strategy. This is the "pull" behind the TrendForce finding that Google TPUs are "increasingly being sold to external clients such as Anthropic."

Key Claims

• Up to 1 million TPUs committed. Evidence: strong (first-party)
• Well over a gigawatt of capacity online in 2026. Evidence: strong (first-party)
• Tens of billions of dollars in contract value. Evidence: strong (first-party)
• TPU generation = Ironwood (Google 7th-gen). Evidence: strong
• Three-platform compute strategy — Anthropic uses Google TPUs, Amazon Trainium, and NVIDIA GPUs across training and development; primary partnership remains with Amazon (Project Rainier, "hundreds of thousands of AI chips" across multiple US data centers). Evidence: strong

Follow-on (April 2026 expansion — cross-source)

In April 2026, Anthropic signed an additional multi-gigawatt deal with Google and Broadcom for up to 5 GW of next-generation TPU capacity starting in 2027 (the five-year commitment spans ~5 GW of TPU capacity). Anthropic's run-rate revenue surpassed ~$30B by April 2026 (up from ~$9B end-2025). Evidence: moderate (multi-source — separate announcement, summary-derived).

Executive Quotes

"expanded capacity ensures we can meet exponentially growing demand" — Krishna Rao, CFO, Anthropic

TPU "price-performance and efficiency" has been "strong … for several years" — Thomas Kurian, CEO, Google Cloud

Significance for the KB

• Demand-side proof that the ASIC inflection is not just a supply-side / cost-engineering story — frontier labs are voting with gigawatt-scale commitments.
• Confirms the multi-vendor reality: even Anthropic, the largest external TPU buyer, still runs Trainium and NVIDIA in parallel — consistent with the concept page's argument that GPU absolute volume keeps growing.
• Connects two KB entities (Google Quantum AI's parent Google, and the merchant layer Broadcom) to a concrete compute deal.

Caveats

• "Up to 1 million" is a ceiling, not a floor; realized deployment depends on Google's Ironwood supply.
• The April-2026 5 GW / 2027 expansion is summary-derived from a separate announcement; the 2026 1 GW / 1M-TPU figure is the directly-fetched, first-party number.