Anthropic has filed an S-1 registration statement with the US Securities and Exchange Commission, putting the frontier lab's financials in front of public investors for the first time. The Verge framed the filing on October 29 as the financial justification for the company to spend as much as it possibly can on compute. The move follows a $13 billion Series F in September 2025 at a $183 billion post-money valuation, later reinforced by strategic commitments of up to $5 billion from Microsoft
The timeline is straightforward. Anthropic filed its S-1 with the SEC, formally starting the initial public offering process, disclosing business model, revenue mix, cost structure and risk factors as required. The Verge's reading seized on the central logic: if public markets are willing to value frontier model companies on revenue multiples, then converting capital into GPUs and TPUs as fast as possible is the rational move rather than an失控 one. The filing supplies the financial justification for exactly that.
Behind the logic sits a set of already-public commitments. Anthropic's agreement with Google covers up to one million TPUs and more than one gigawatt of new capacity in 2026. Amazon has invested $8 billion in total and built the Project Rainier cluster. In November 2025, Microsoft and Nvidia each committed up to $5 billion and $10 billion respectively. On the product side,
Why now. Since 2025, the capital structure of frontier labs has entered a collective repricing. OpenAI completed its restructuring in October 2025, placing a public benefit corporation under foundation control. CoreWeave listed on Nasdaq in March of the same year, giving compute infrastructure assets a valuation anchor. xAI and several model companies have been preparing their own windows. Appetite for AI assets on public markets swung from skepticism to competition during that stretch.
Cost dynamics are more direct. Single training runs have moved from tens of millions of dollars to the billion-dollar scale, while inference spending expands roughly linearly as coding and agent products
Liu Gong's take: within the next 12 months, Anthropic will most likely break out its long-term compute purchase commitments as a separately disclosed off-balance-sheet obligation in a subsequent S-1 amendment, because what this filing actually sells is not software subscriptions but a publicly priced compute curve.
From a technical standpoint, the cost structure of frontier models has inverted. A comparison using public parameters makes the scale clear: Anthropic's Google agreement covers up to one million TPUs and more than one gigawatt of added capacity in 2026, while a typical hyperscale campus draws roughly 100 megawatts at full load. One gigawatt is therefore equivalent to binding ten such campuses to a single model company. Capital expenditure at that magnitude has only two outlets, either raising in public markets or letting cloud providers convert investment into orders. The S-1 chooses the former, and the price is that every dollar of compute gets interrogated through depreciation in quarterly reports.
On the business side, the real difficulty is who defines gross margin. Claude Code crossing $1 billion in annualized revenue shows genuine willingness to pay, yet inference costs scale with usage and margin depends heavily on assumed depreciation life. Stretching GPU depreciation from four years to six can add more than ten percentage points of reported gross margin, exactly the question public investors are best equipped to press. The governance comparison is equally telling: after OpenAI's restructuring, Microsoft holds roughly 27% and the foundation about 26%, while Anthropic's Long-Term Benefit Trust holds zero economic interest. Outside shareholders buy cash flow with almost no leverage over control.
For the industry, the demonstration effect outweighs the raise itself. If Anthropic prices successfully on a revenue multiple, the compute race shifts from who can raise the most to who can keep adding capacity under quarterly scrutiny, and funding costs for second-tier labs rise in step. In industrial machine vision, the calculus Liu Gong sees daily is different: manufacturing customers ask first whether per-inspection cost amortizes within 18 months, not where a model ranks on a benchmark. When public markets demand quarterly returns while factory floors pay back over years, that mismatch of clocks is the first friction an IPO would expose.
- – Anthropic raised $13 billion in a September 2025 Series F at a $183 billion post-money valuation, roughly 26x annualized revenue of about $7 billion at the time, versus the 8-12x range common for enterprise software.
- – The Google agreement covers up to one million TPUs and more than one gigawatt of new capacity in 2026, equivalent in site power to about ten typical 100-megawatt hyperscale campuses.
- – Microsoft and Nvidia committed up to $5 billion and $10 billion respectively in November 2025, a combined $15 billion that exceeds Anthropic's 2023 and 2024 rounds added together.
- – Claude Code passed $1 billion in annualized revenue by late 2025, going from zero to that mark in under a year,
For the industry, the near-term effect within three months is a shift in the private-market pricing anchor. Once Anthropic's roadshow range is set, other frontier labs will negotiate against the same ruler, and investors will demand matching disclosure of compute contracts. Over a one-year horizon, competition among leading labs shifts from benchmark rankings to revenue conversion per unit of compute, while circular investment-for-orders arrangements with cloud providers will face repeated scrutiny from regulators and auditors.
For ordinary users, the near-term impact shows up in price and availability: quarterly pressure pushes vendors to tighten free tiers and move expensive reasoning
Competitors responded with restraint but clarity. OpenAI has long argued publicly that compute must be purchased ahead of demand, with executives describing it as a strategic reserve, a narrative closely parallel to Anthropic's filing. xAI and several second-tier labs prefer to stress training efficiency, pulling the conversation back toward cost per unit rather than capital expenditure. Cloud providers are the most enthusiastic: Google and Amazon are simultaneously shareholders and suppliers, and the compute agreements are themselves order books.
Expert opinion divides on accounting. Some analysts argue that off-balance-sheet treatment of multi-year compute commitments obscures real
Questions worth exploring:
- How the accounting treatment of compute commitments in the S-1 will shape gross margin in subsequent quarterly reports
- What the public benefit corporation plus Long-Term Benefit Trust structure actually leaves public investors with
- How a successful IPO reprices funding costs and compute access for second-tier model developers
Further reading: 2026 Frontier LLM Comparison: A Practical Guide — a durable, evergreen reference. Bookmark it.
This item comes from The Verge AI (published 2026-10-29 13:30:00). Wenfeng Radar provides a Chinese/English summary and commentary on overseas AI developments; the original content is owned by its author.
Daily aggregation of overseas AI developments with in-depth analysis. Bookmark us to never miss an important signal; back to homepage.