Nikunj Kothari 拆解 Anthropic S-1 中的算力合同

The main thing I wanted to dig into @AnthropicAI's S-1 was the compute contracts..

A few things that stood out (built in conjuction with Opus 5.5):

> At June 30, Anthropic had $265.6B of purchase commitments and $321.4B of leases that hadn’t started. That’s a $587B subtotal, payable over years, including some conditional leases. The latest aggregate total is still blank in the filing.

> The contracts are very different. Google and Amazon require minimum spending regardless of usage. SpaceXAI’s agreements have a potential value of $120.5B, but either party can cancel on 90 days’ notice after initial locked periods end in January and March 2027.

> Anthropic Direct separates chips, buildings, power and operations into different contracts. It lets Anthropic reserve a building early and commit to chips later. It also means one supplier can start billing before the others are ready. Equipment rent can begin before the compute is usable.

> Broadcom’s $42B facility never hands Anthropic cash. Notes settle specific equipment rent bills. The facility excludes each equipment pod’s first quarterly payment and any rent due before it is ready for service. Anthropic has to find a way to fund those payments elsewhere.

> Those notes carry 9% interest, added to the balance quarterly, and can convert at 90% of the trailing 30-day average share price after IPO. New notes issued during specified revenue-breach periods get 13% interest and an 85% conversion price. The share-price measure is volume-weighted.

> That interest sits on top of the equipment lease’s own financing charge, whose rate is confidential. The investor chooses whether to convert the notes. Otherwise, principal and accumulated interest come due in cash.

> Anthropic treats Trainium, TPUs, NVIDIA and AMD GPUs as a shared fleet across model development, paying customers and internal use. It doesn’t disclose a fixed physical split between training and inference. The supplier bill can be fixed while the work assigned to the hardware changes.

> Revenue per dollar of total compute and infrastructure expense rose from $0.57 in Q2 2025 to $2.08 in Q2 2026. Pricing, utilization, technical improvements and when capacity arrives all affect that ratio. It’s a substantial improvement in monetizing the compute base.

> About 80% of H1 revenue was consumption-based. Most compute commitments however run for years. With quarterly releases, this bifurcation between usage and subscription will be interesting to monitor.

Source: https://t.co/97KqkyyfpS