Anthropic announced on 6 April a significant expansion of its use of Google Cloud infrastructure, securing multiple gigawatts of TPU capacity to scale development of foundation models, AI agents, and enterprise applications. The capacity, delivered through Google Cloud services and Google-built TPUs supplied through Broadcom, is expected to come online starting in 2027.
The scale of the deal signals where Anthropic sees the compute requirements heading. Multiple gigawatts of TPU capacity represents an extraordinary amount of processing power — the kind of infrastructure needed to train and serve models substantially larger and more capable than the current Claude Opus 4.6. For context, a single gigawatt can power roughly 750,000 homes. Anthropic is clearly building for a future where AI agents run continuously at enterprise scale, not just responding to individual queries.
The partnership already has deep roots. Thousands of customers currently access Claude models through Google Cloud, including Cursor, Replit, Coinbase, Palo Alto Networks, and Shopify. Anthropic also uses Google Cloud's BigQuery, Cloud Run, and AlloyDB services as part of its operational infrastructure. This expansion moves the relationship from cloud services consumption to a strategic infrastructure commitment that will shape Anthropic's model development roadmap for years.
For context engineers, the practical implication is straightforward: Anthropic is investing heavily in the compute needed to keep Claude at the frontier. Larger training runs mean more capable models, and dedicated serving infrastructure means better availability and lower latency for the agentic workflows — Claude Code sessions, MCP server chains, multi-agent orchestrations — that developers rely on daily. The 2027 timeline also suggests that whatever Anthropic is planning for its next generation of models will require compute at a scale that current infrastructure cannot support.
