Anthropic underwrites massive compute capacity

Anthropic is no longer just selling model access, it is underwriting its own power plant sized supply chain. By locking in Google TPU capacity through Broadcom, Amazon Trainium capacity through AWS, and additional cloud capacity with providers like Lambda, Anthropic makes sure it can train and serve Claude even if chips stay scarce. The tradeoff is that revenue now has to catch up to infrastructure bills that keep running whether customer demand arrives smoothly or not.1 2 3 4 5 6

• The burden is unusually rigid. Reuters reported at least $518B of buildout commitments with six partners, plus about $161.2B of Broadcom related equipment lease obligations that are largely non cancelable. That means a meaningful share of spend behaves more like rent on a giant factory than normal cloud usage.4 5
• These deals buy real operating leverage if demand materializes. Anthropic said Amazon is its primary cloud and training partner, with up to 5 GW of capacity and over one million Trainium2 chips already used to train and serve Claude. It separately expanded its Google and Broadcom partnership for multiple gigawatts of next generation TPU capacity starting in 2027.1 2 3
• The commercial race is therefore about filling reserved compute with high value workloads fast enough. Anthropic's revenue run rate rose from $1B at the end of 2024 to $9B at the end of 2025 and $65B by July 2026, which shows why management is willing to prebuy supply years ahead. The core job is turning coding, enterprise, and API usage into dense utilization of that committed capacity.6 7 8

Going forward, frontier model competition will look more like airline route planning or power procurement than pure software. The winners will be the labs that can keep utilization high, steer workloads across TPUs, Trainium, and GPUs, and finance the gap between infrastructure commitments today and customer cash flows tomorrow.1 2 4 6