1 招股书背后的 AI 经济账
Key Highlights
In this episode of The Verge's Decoder, Nilay Patel sits down with SemiAnalysis analyst Joey Brookhart for a deep teardown of Anthropic's S-1 filing. The two keep circling one central tension. On one side sits the inflated valuation propped up by an AGI narrative. On the other sit the hard numbers in the filing, enormous compute commitments and accounting losses written in black and white. Brookhart argues that if the AGI story actually pays off, Anthropic's long-term valuation could reach $10 trillion. But that figure rests on extremely optimistic assumptions about both technology and demand, and ignores how quickly sentiment can turn if growth stalls.
What Happened
The episode surfaces several concrete figures that are worth holding onto. Anthropic has formally filed its S-1 with Nasdaq and could list as soon as next month at a valuation between $1.5 and $2 trillion. Current annual recurring revenue exceeds $70 billion. Inference gross margin climbed from negative 94% two years ago to positive 85% today, with roughly $65 million of inference revenue per megawatt of deployed hardware, while the cost side is only about $10 million per megawatt. Non-GAAP EBITA margin sits around 11%. The eye-watering $80 billion loss reported for the first half of 2026 is a non-cash charge tied to the revaluation of convertible debt and warrants. Actual cash burn is closer to $10 billion, while the company holds more than $120 billion in cash and marketable securities.
Technical Details
Compute commitments are the other focal point. The S-1 discloses between $500 and $600 billion in lease commitments to hyperscalers and neoclouds over seven to ten years, with some analysts pegging the ten-year figure above $800 billion. Anthropic runs mixed Trainium, TPU, and GPU clusters to grow token throughput five- to tenfold every year, and uses a special-purpose vehicle structure to buy chips directly and lease them back. Only about 30% of its compute currently earns inference revenue. The rest feeds training and research and development, which means the asset base is enormously front-loaded against future demand that has not yet materialized.
Versus Competitors
The show notes that in net new ARR during the third quarter, OpenAI overtook Anthropic to take the majority share. Open-source inference cost ballooned from $1 billion in the first quarter to more than $100 billion in the third, and enterprises increasingly prefer on-prem or open-source deployments for data-privacy reasons while coding-switching costs stay low, squeezing pricing power. OpenAI is more consumer-facing, runs negative EBITA, and has delayed its own IPO to tidy up the financials first.
Industry Impact and Use Cases
This episode translates a dry prospectus into tangible business logic. Anthropic's valuation is half cash flow and half faith. It reminds listeners that when colossal compute commitments meet slowing growth or tightening financing, even a machine-god business can turn into a debt problem overnight. For practitioners trying to understand the resilience of the AI-IPO bubble, this is a high-quality plain-language briefing that connects the headline numbers to the underlying mechanics.
Practical Notes
The single most useful takeaway for readers is to separate the non-cash accounting losses from real cash burn. An $80B headline loss looks terrifying, but the underlying cash trajectory is what determines solvency. Watch the share of compute earning inference revenue, today only about 30%, because that ratio, more than any model benchmark, signals when the build-out starts paying for itself.
What to Watch
Keep an eye on the IPO pricing and the first post-listing quarters. If net new ARR leadership keeps shifting toward OpenAI and open-source inference keeps getting cheaper, Anthropic's pricing power, and therefore its multiple, will be tested. The $500 to $600B compute commitments are fixed for years, so any demand hiccup lands straight on the balance sheet.
Takeaway
Anthropic's story is the clearest window into how the frontier labs are valued, as options on AGI rather than as ordinary software companies. That makes them exhilarating and fragile at the same time, and the S-1 is the document that finally makes the fragility legible to ordinary investors.
Background
The broader context is that 2026 became the year AI labs stopped being private and started filing to go public, and Anthropic's S-1 is the template the others will follow. That is why a podcast teardown of its numbers draws so much attention from investors who previously only saw headline valuations. The tension between a ten-trillion-dollar faith-based ceiling and an eight-hundred-fifty-billion-dollar fixed floor of commitments is the defining risk of the whole sector, and it is worth sitting with before believing either the bulls or the bears.
Where to Look Next
For readers who want to go deeper, the original S-1 filing and SemiAnalysis's own written notes are the primary sources, and the Decoder episode is best treated as a guided tour of those documents rather than a standalone analysis. Track the IPO prospectus updates week by week once the roadshow begins, because the gap between the filed numbers and the priced valuation is where the real story will be told.
Closing Thoughts
The episode is a useful antidote to both hype and doom. It shows that the most valuable AI company in the pipeline is, underneath the narrative, a leveraged bet on perpetual growth, and that bet is only as strong as the next few quarters of demand.