Anthropic's $518B Compute Bet: Who Wins From the AI Buildout?
Anthropic confidentially filed for its IPO back in June. But the bigger story for AI investors isn't the listing — it's the prospectus.
According to the confidential prospectus reviewed by Reuters, Anthropic has committed at least $518 billion to cloud, compute, and AI infrastructure over roughly the next decade. About 80% of it is non-cancelable — they pay whether they use the capacity or not.
That is one of the largest AI buildout commitments on record. And it's coming from a company that did $4.6B in revenue last year.
The growth story
Anthropic generated about $4.6B in 2025 revenue — roughly 12x the prior year. But growth is coming at enormous cost: compute and infrastructure spending hit $7.33B (more than half of its $12.65B in operating expenses), the operating loss topped $8B, and the reported net loss was about $42B — though roughly $34B of that was a financing-related accounting charge, not cash out the door.
Anthropic's thesis is simple: future AI development will be "limited principally by the availability of compute." So it's moving beyond rented cloud into dedicated data centers and directly leased chips — GPUs, TPUs, ASICs — to control its own supply.
Where the $518B goes
The commitments span six partners, and the numbers are staggering:
- $GOOGL — at least $111.1B over 7–10 years. Pay-or-play: "if our actual spend falls short, we must pay Google the difference."
- $AMZN — at least $110B, same take-or-pay structure.
- $MSFT — $31.4B, non-cancelable except for Microsoft's uncured breach.
- $AVGO — $161.2B in equipment lease obligations, largely non-cancelable. The single largest disclosed exposure.
- SpaceX — up to $84.5B through 2029 at the Colossus 1 facility (220,000+ Nvidia GPUs on 300MW of power), per the updated prospectus.
This is the $5 trillion compute thesis playing out in real time: the winners sit in layers — cloud ($GOOGL, AMZN, MSFT), chips and custom silicon (NVDA, AMD, $AVGO), memory and networking ($MU, $MRVL), and eventually data centers and power ($BE).
How big a bet is this, really?
| 2025 revenue | $4.6B |
| 2025 compute spend | $7.33B |
| 2025 operating loss | >$8B |
| Future commitments | $518B+ |
The duration mismatch is the whole story. The compute cost is locked in for a decade; the revenue to pay for it isn't. Nearly a quarter of 2025 revenue came from just two customers, and many large customers aren't on long-term contracts.
If Claude keeps scaling at 12x, buying compute early is a competitive moat. If AI pricing falls, models get dramatically more efficient, or demand growth slows — those non-cancelable commitments become a very expensive anchor.
The takeaway
Investors have tracked AI demand through hyperscaler capex. Anthropic just extended the visibility a decade out — and it's the labs themselves now signing the checks, not just the clouds. That's another major validation for the AI infrastructure trade.
But read it both ways: much of the next decade's compute is already spoken for. Now Claude's revenue has to grow into $518 billion. That's the bet.