AI AI Toolkit
AI Newsindustry

约 8500 亿美元 compute 承诺押注递归自我改进

The Verge:AI(RSS)2026-10-29T13:30:00.000Z

Key Highlights

On October 29 The Verge published a deep read of Anthropic's S-1 prospectus with a blunt title, Build God or die trying. The piece argues Anthropic has staked its future on extreme growth because only continued growth can cover its fixed costs. The grand machine-god narratives in the S-1 are essentially a pitch to Wall Street that this is a business that can pay for itself, not a bottomless money pit, and the article is unsentimental about the math.

What Happened

The good news is striking. Anthropic's revenue grew 14x to $16.3 billion between the first half of 2025 and the first half of 2026. Stripping out an $80B-class non-cash accounting loss and items like compensation leaves about $1.62 billion in adjusted income for the first half of 2026. Net revenue retention is 900%, and roughly three-quarters of first-half 2026 customers spend over $100,000 a year, showing a firm grip on the enterprise market. The bad news is that competitive and cost pressure is intense, with OpenAI and Google named as the only rivals and open-source models eating in as well.

Technical Details

The most staggering part is the commitment scale. The filing discloses $265.6B in minimum non-cancelable purchase commitments on one-to-twenty-year terms, $321.4B in not-yet-commenced leases, then about $245B in new compute and infrastructure agreements, plus $125 to $150B in unfinished transactions, a twenty-year minimum of roughly $850B, approaching $1 trillion if unfinalized deals are included. A PitchBook analyst models about $90B annual spend for 2027 to 2030, requiring $188B revenue to keep the cost ratio stable, which is a steep climb from today's base.

Versus Competitors

Against Google's multiple revenue lines, search and cloud, Anthropic's single frontier-model-plus-enterprise path is more fragile. It must keep training to stay ahead yet lacks Google's distribution advantage. The article sharply notes that recursive self-improvement is both a business necessity and a source of existential risk, capabilities advancing faster than safety validation, which is why the S-1 is littered with catastrophic and existential language.

Industry Impact and Use Cases

For investors and AI practitioners, this read is a risk lesson. Frontier-lab valuations rest on the assumption that growth never stops. Once demand saturates or financing tightens, the $850B commitment turns from a strategic asset into a debt noose. It also reveals a brutal reality that AI-safety investment and commercial speed are bound to the same wagon, and you cannot fully optimize one without constraining the other.

Practical Notes

Read the commitments, not the headline revenue multiple. A 14x growth story looks unstoppable until you map it against fixed multi-year compute obligations that must be serviced regardless of demand. The $90B annual spend modeled by PitchBook implies revenue must nearly double again just to hold the cost ratio, a target that leaves no room for a demand plateau.

What to Watch

Track whether enterprise net revenue retention stays near 900% as open-source alternatives mature. If customers start routing a meaningful share of inference to on-prem or open weights, Anthropic's pricing power erodes and the build-god-or-die math gets harder, with direct consequences for the IPO multiple.

Takeaway

Anthropic's S-1 is the clearest document yet showing that frontier labs are leveraged bets on perpetual growth, where a single demand plateau can turn a trillion-dollar story into a balance-sheet crisis almost overnight.

Background

The Verge's framing, build God or die trying, captures the existential math of frontier labs better than any analyst slide. They are not ordinary growth companies, because their cost base is a multi-year bet on demand that may or may not arrive, and the S-1 is the first time that math was laid bare in a public filing that ordinary investors can read.

Where to Look Next

The most useful follow-up is the post-listing quarterly report, because that is where the rubber meets the road. If net revenue retention holds near 900% and free cash flow trends positive, the build-god thesis holds and the valuation looks defensible. If either cracks, the die-trying half of the equation becomes the headline and the multiple compresses quickly.

Closing Thoughts

Anthropic's document is a gift to anyone who wants to understand how the AI boom is actually financed, and it should be read alongside the infrastructure-spending stories, because the labs and the cloud builders are two sides of the same leveraged bet.