Anthropic S-1 Reveals $80.7 Billion Half-Year Loss and Up to $750 Billion in Planned Compute Spending

Anthropic's draft S-1 prospectus, viewed by the Wall Street Journal, lays bare the staggering financial scale of frontier AI development. The company posted an $80.7 billion net loss in the first half of 2026, up sharply from a $13 billion loss in the same period a year earlier, though the loss was largely driven by accounting charges. Revenue reached $16.3 billion in that half, and Anthropic spent nearly $9 billion on computing power alone in the same stretch.

The forward-looking commitments are even more striking. Anthropic is on the hook for roughly $450 billion in cloud computing agreements over the next seven to ten years, with $111 billion earmarked for Google Cloud, $31 billion for Microsoft, and up to $86 billion for renting Nvidia chips. Total planned compute spending could reach between $600 billion and $750 billion depending on the source, and the prospectus includes a warning about the risk of a catastrophic incident. Separately, Anthropic is reported to be in talks for another $45 billion deal that would bring its total Nscale commitment to roughly $90 billion, a development that analysts say would make Nvidia a major beneficiary.

The Anthropic disclosures land on the same day that Business Insider reported Amazon's own AI infrastructure bind: AWS is receiving an initial $5 billion shipment of Nvidia GB300 racks, with roughly $3 billion more expected each month, but lacks sufficient data center power to turn much of the hardware on. The result is billions of dollars in GPU systems sitting idle in storage facilities, underscoring that power availability, not chip supply, has become the binding constraint on AI expansion.

Both companies are simultaneously pushing to reduce their dependence on Nvidia's software ecosystem. AWS launched TorchNeuron for its Trainium chips, and Google open-sourced TorchTPU, its long-awaited PyTorch implementation for TPU hardware. The moves are aimed at pulling AI developers away from the default assumption that production workloads must run on Nvidia GPUs.