← Back to feed

Anthropic announced on April 6, 2026 that it has signed a new agreement with Google and Broadcom for multiple gigawatts of next-generation TPU capacity, expected to come online starting in 2027. The new capacity will power frontier Claude models and serve what Anthropic describes as “extraordinary demand from customers worldwide.”

The deal is notable for its stated scale. Gigawatts — not megawatts — is an unusual unit of compute procurement to publicize. For context, training runs for current frontier models are estimated to require tens to hundreds of megawatts of sustained power. A multi-gigawatt commitment implies either very large training runs, massive inference scale-out, or both. Anthropic said the vast majority of the new compute will be sited in the United States, framing the deal as an extension of its November 2025 commitment to invest $50 billion in strengthening US computing infrastructure.

The Google Cloud Relationship

This is not Anthropic’s first Google Cloud TPU announcement. The company also expanded its TPU access in October 2025. Today’s deal builds on that prior expansion and goes substantially further. The structure involves two parties: Google Cloud (for services and TPU access) and Broadcom (for Google-built TPUs supplied through Broadcom). The three-way arrangement — Anthropic as customer, Google as cloud provider, Broadcom as chip supplier — reflects the vertical structure of Google’s custom silicon strategy, where Broadcom manufactures TPU ASICs at scale.

Thousands of enterprise customers already access Claude through Google Cloud, including Coinbase, Cursor, Palo Alto Networks, Replit, and Shopify.

Business Context

Anthropic’s demand signal is consistent with this commitment. The company raised $30 billion in February 2026 at a $380 billion valuation and reports $14 billion in annualized revenue. In the weeks around this announcement, Anthropic noted that the number of business customers spending over $1 million annually on Claude has more than doubled in under two months. Revenue is accelerating, which directly drives the need for inference capacity ahead of training capacity.

The timing also follows a period of public compute anxiety in the AI industry. Several labs have cited compute constraints as a ceiling on model development timelines. Locking in gigawatt-scale TPU capacity two years ahead — at a time when AI compute demand is outpacing supply from Nvidia and AMD — is a meaningful strategic hedge.

What This Is Not

This is compute-as-a-service, not custom chip development. A separate, earlier story reported Anthropic exploring its own custom AI chips. The Google-Broadcom deal is the opposite direction: doubling down on Google’s TPU stack rather than building proprietary silicon. Whether Anthropic pursues both paths simultaneously remains an open question, but the gigawatt TPU deal represents the near-term compute strategy while any custom chip program remains longer-horizon.

Key Numbers

• Capacity: Multiple gigawatts (next-generation TPUs)
• Estimated online: Starting 2027
• Counterparties: Google Cloud, Broadcom
• Geography: Majority sited in United States
• Context: Extends November 2025 $50B US infrastructure commitment
• Claude enterprise customers spending $1M+/year: More than doubled in under 2 months