Anthropic Signs Multi-Gigawatt TPU Deal With Google and Broadcom for 2027

Anthropic Signs Multi-Gigawatt TPU Deal With Google and Broadcom for 2027

Anthropic, the AI safety company behind the Claude family of models, has finalized a landmark agreement with Google and Broadcom to secure multiple gigawatts of Tensor Processing Unit computing capacity scheduled to come online starting in 2027. The deal represents one of the largest single infrastructure commitments in the history of artificial intelligence and signals Anthropic’s ambition to remain at the forefront of frontier model development as the industry enters an era of unprecedented scale.

The Scale of the Agreement

While exact financial terms have not been publicly disclosed, industry analysts estimate the deal is worth tens of billions of dollars over its multi-year duration. The agreement gives Anthropic priority access to Google’s next-generation TPU v6 processors, manufactured in collaboration with Broadcom’s advanced semiconductor packaging technology. The computing capacity secured would be sufficient to train models significantly larger and more capable than anything currently in existence, with some estimates suggesting the infrastructure could support training runs requiring over 10 zettaflops of sustained compute.

Strategic Implications for Anthropic

The TPU deal addresses what has been a critical bottleneck for AI companies: access to sufficient computing resources to train increasingly large models. Anthropic’s decision to partner with Google for TPU capacity reflects a strategic diversification of its compute supply chain. This approach reduces dependency on any single hardware provider and gives Anthropic access to TPU-specific optimizations that can improve training efficiency. The partnership also deepens the already significant relationship between Anthropic and Google, which has invested billions in the AI safety company.

What This Means for the AI Industry

The sheer scale of the deal underscores the enormous infrastructure investments required to develop next-generation AI systems. As models continue to grow, the companies that can secure reliable access to computing resources will have a decisive competitive advantage. The partnership highlights the growing importance of custom silicon in AI development, as general-purpose GPUs face increasing competition from purpose-built AI accelerators designed to optimize specific workloads.

Timeline and Future Capabilities

The infrastructure is expected to become operational in phases beginning in mid-2027, with full capacity available by early 2028. Anthropic has indicated that the new computing resources will be used to train the next several generations of Claude models, as well as to support expanded research into AI safety, interpretability, and alignment. The additional capacity will also enable more competitive pricing for API services, making advanced AI capabilities more accessible to developers and organizations worldwide.