# Anthropic's signed compute deals vs its compute bill today
Latent Equity · published Oct 26, 2026 · AI capital ledger
Post: https://x.com/Latent_Equity/status/2114795037309763861

Our read: Anthropic's revenue now covers its compute bill about 2x over, but the compute deals it has signed work out to roughly 3x to 4.6x what it spends on compute today (our calc).
Signed compute deals work out to about $103B a year, vs about $22B a year of compute spending today (our calc).

What it means for a regular investor: that gap is a bet that Anthropic's revenue keeps compounding. Our read: part of the AI build at $AMZN, $GOOGL and $MSFT leans on customers like this paying for years.

A caveat first: this S-1 (the filing a company makes before going public) is preliminary. Its headline totals are blank, so the figures below are built from the deal-by-deal numbers it does fill in.

Today:
- Q2 2026 revenue was $11.6B, against $5.6B of compute and infrastructure costs. That compute figure covers research and model training as well as serving customers.
- Revenue per $1 of compute went from $0.40 in Q1 2025 to $2.08 in Q2 2026 (the filing's own figures).

Signed:
- Cloud deals: $AMZN at least $110.0B, $GOOGL about $114.5B, $MSFT at least $44.9B. Non-cancelable except for breach or undelivered capacity.
- Chip equipment leases with $AVGO: about $161B. That's leased hardware, not cloud backlog.
- A deal to rent computing power on $NVDA chips (not with $NVDA itself) of up to $120.5B that can be cancelled on 90 days' notice from early 2027.

Spread evenly over each contract, that's about $103B a year, or roughly 4.6x today's compute spending (our calc). Leave out the cancellable deal and it's about $66B a year, roughly 3x (our calc). These are payment commitments set against today's expense line, so it's a rough comparison, not like-for-like. Actual first-year spending will be lower, since deals ramp up.

Our read: these deals could be up to about a fifth of the backlog (signed future revenue) at $AMZN and $GOOGL, and about 7% at $MSFT. The filing doesn't confirm they sit in that backlog. This week's earnings may tell us.

What would prove us wrong:
- an amended S-1 fills in totals well below these figures, or revenue stops growing fast enough to keep up with compute costs as the deals kick in.

General research, the same for every reader. Not investment advice.