Anthropic Filing Reveals $600B+ Compute Commitments

A draft IPO filing reviewed by the Wall Street Journal shows Anthropic growing revenue nearly fifteen-fold year on year while committing more than USD 600 billion to computing capacity. The disclosure, reported on 24 October, is the clearest public view yet of how much infrastructure a frontier AI lab must contract before it earns the corresponding revenue.

The prospectus states that around one fifth of USD 16.3 billion in first-half 2026 revenue came from just two customers, and warns that a significant portion of revenue will come from a limited number of customers for the foreseeable future. Commitments include USD 111 billion with Google Cloud, USD 110 billion with Amazon Web Services and USD 31 billion with Microsoft, described as payable regardless of actual usage, plus up to USD 86 billion to rent Nvidia chips from Nscale and Lambda through 2033 and 2034.

Chip leases add a further layer. Including an obligation of more than USD 160 billion to lease Google chips through a Broadcom arrangement, total future infrastructure costs exceed USD 600 billion, and the filing says the company is negotiating deals that could add up to USD 150 billion more. The Wall Street Journal also reported that total compute spending could reach USD 750 billion over the next few years.

Two structural details matter to hardware suppliers. First, the commitments are take-or-pay in part, which makes near-term demand for accelerators, HBM and packaging far more visible than the company's revenue. Second, the mix is explicitly multi-vendor: Google TPUs, AWS Trainium, Nvidia GPUs and AMD Instinct all appear, so component demand spreads across ASIC and GPU supply chains rather than concentrating in one.

The risk is funding, not intent. Anthropic is seeking a valuation of up to USD 2 trillion and has committed at least five percent of its shares to a public benefit fund. If the IPO prices below expectations, the pace of drawdown on those contracts, rather than their existence, becomes the variable that decides 2027 order books.

Suppliers should structure agreements around cancellation windows and step-in rights, and qualify parts that ship into both ASIC and GPU platforms. YouHuYou lists such parts in our catalogue, from 40A PMBus step-down converters to 85V digital power monitors and gigabit Ethernet PHYs for rack management networks.