Anthropic Locks In 3.5GW of TPU Compute via Google and Broadcom Deal
Summary: Anthropic signed a multi-year deal for 3.5 GW of Google TPU capacity brokered through Broadcom, with most infrastructure slated for the US and online by 2027.
Key Facts
- The deal provides ~3.5 GW of next-generation compute through a multi-year Broadcom arrangement; new capacity expected starting 2027, mostly US-based
- Anthropic is currently using 1 GW of Google TPUs, per Broadcom CEO Hock Tan
- Anthropic's annualized revenue run rate crossed $30B in 2026 (vs. ~$9B at end-2025); enterprise customers spending $1M+ annually more than doubled in under two months
- The deal deepens Anthropic's reliance on Google's TPU stack, with Broadcom handling silicon implementation, rack integration, and production scaling
Why It Matters
Compute capacity is the hard ceiling on AI growth. Locking in 3.5 GW — three-and-a-half times current usage — signals that Anthropic views its demand surge as structural, not cyclical. It also cements Google as Anthropic's primary infrastructure partner, creating a significant strategic entanglement even as Anthropic eyes an IPO. For the broader industry, the scale of this TPU commitment is a data point on how the AI hyperscalers are systematically hedging away from pure Nvidia GPU dependency.
Read More
- TechCrunch original — TechCrunch
- Deal details — Yahoo Finance
How this story unfolded
- Google Ships Nano Banana 2 Lite and Gemini Omni Flash: 4-Second Images, Video Generation
- Anthropic Claude Goes GA on Azure AI Foundry, Powered by NVIDIA GB300 Blackwell Ultra
- OpenAI and Broadcom Unveil 'Jalapeño' — First Custom AI Inference Chip
- Google and Blackstone Form $5B Joint Venture for TPU-Powered AI Data Centers