Anthropic announced a landmark expansion of its Google Cloud TPU footprint, securing access to up to one million TPU chips with well over a gigawatt of capacity coming online in 2026, plus an additional 3.5 gigawatts via a tripartite agreement with Google and Broadcom set to launch in 2027. The combined TPU expansion represents a 4.5x increase in Anthropic's compute base within 18 months. The investment is structured as a multi-year commitment worth tens of billions of dollars, with Google Cloud CEO Thomas Kurian highlighting TPU price-performance and efficiency as the deciding factors for Anthropic's selection over alternatives.
Anthropic's compute strategy is explicitly multi-cloud: the company trains Claude across Google's TPUs, Amazon's custom Trainium 2 chips (via Project Rainier), and NVIDIA GPUs, allowing it to match workloads to the chip best suited for each task. This diversified approach reduces vendor lock-in and stretches every dollar of compute further than single-vendor architectures. The company also continues its primary partnership with Amazon Web Services, which remains its chief cloud provider and training infrastructure partner. Anthropic's annual revenue run rate has surpassed $30 billion, up from approximately $9 billion at the end of 2025—a tripling in roughly 12 months.
For the AI infrastructure space, Anthropic's TPU expansion signals that custom ASICs have crossed a durability threshold: frontier labs can now justify multi-gigawatt commitments to NVIDIA alternatives. The Broadcom-Google partnership supplies the custom silicon backbone for this expansion, locking in billions in revenue for Broadcom across multiple generations through 2031. For architects, the message is clear: GPU capacity is no longer the only game in town. Multi-cloud, multi-chip strategies are becoming standard practice for scaling frontier models.