The curtain has risen on an IPO that will test Wall Street's confidence in the AI boom. On October 26 local time, U.S. artificial intelligence large-model company Anthropic (NASDAQ: ANT) submitted a public version of its S-1 filing to the U.S. Securities and Exchange Commission (SEC), planning to list on the Nasdaq under the ticker "ANT," with Goldman Sachs and Morgan Stanley leading the underwriting. The offering price, number of shares, and fundraising size have yet to be determined, but market expectations for its valuation have already surged to between $1.5 trillion and $2 trillion. If completed at the upper end, this would become one of the largest IPOs in history. How can a company founded just five years ago achieve such scale?
The prospectus offers a direct answer: revenue is "sprinting," quarterly operating profit has already turned positive, and enterprise customers are paying real money for tokens. But the same document also lays bare the other side — hundreds of billions of dollars in compute bills for the coming years have already been signed, yet nearly 80% of revenue is settled based on actual usage, and most top customers have not yet signed long-term contracts. From Wall Street to Silicon Valley to the entire AI industry, the market is watching this IPO closely. What truly needs to be priced is not whether AI can make money, but whether growth can continue to outrun costs when a decade of bills has been signed in advance. "If Anthropic performs poorly after going public, investors will have to reprice AI companies," said Brad Bernstein, managing partner at FTV Capital, an institution focused on investing in growth stocks.
A Prospectus That Proves AI Can Make Money
Anthropic's revenue curve looks impressive. Full-year revenue was $41 million in 2023, $386 million in 2024, and $4.585 billion in 2025. In the first half of 2026, it jumped directly to $16.286 billion, roughly 14 times the $1.163 billion in the same period a year earlier. On a quarterly basis, revenue was $4.731 billion in the first quarter of this year and rose to $11.555 billion in the second quarter, a sequential increase of about 144%. There is no official confirmed data for the third quarter yet, but estimates put it at about $16.8 billion, with sequential growth cooling to around 45% — still extremely fast, but no longer doubling quarter after quarter.
More critical is the lower half of the income statement. In the second quarter of 2026, Anthropic achieved operating profit of $974 million, an operating margin of about 8%; after excluding stock-based compensation, non-cash charitable donations, and other items, adjusted operating profit reached $2.235 billion, a margin of about 19%. By comparison, OpenAI's adjusted operating margin in the same period was about -122%. What supports these numbers is the speed of enterprise customer adoption. As of the end of September 2026, more than 11,900 customers had contributed more than $100,000 in revenue over the past 12 months, compared with only about 950 a year earlier; net revenue retention was about 900% — meaning the same group of enterprise customers (excluding individual subscriptions) spent nine times as much in a quarter as they did a year earlier. At the same time, compute efficiency continued to improve: revenue generated per dollar spent on compute and infrastructure rose from $0.40 in the first quarter of 2025 to $2.08 in the second quarter of 2026. Second-quarter revenue grew 1368% year over year, while compute spending increased only about 300%. Anthropic's formal sprint toward an IPO has ignited the long-quiet global AI market, with the Philadelphia Semiconductor Index surging 2.82% overnight. "This will be the first major opportunity for investors to directly invest in the companies driving the AI economy, and it is of enormous importance," said Gil Luria, head of technology research at U.S. financial services firm DA Davidson.
A Decade of Bills Signed in Advance
The sharpest contradiction in the prospectus is not in the income statement, but in the contracts. In the first half of 2026, Anthropic's compute and infrastructure expenses were $8.945 billion, about 55% of revenue, already exceeding the full-year 2025 figure of $7.326 billion. The company does not mainly rely on building its own data centers, but instead obtains compute through long-term procurement, equipment leases, and third-party data center partnerships, with contracts typically locked in 12 to 18 months in advance. Based on binding force from strongest to weakest, Anthropic's compute commitments can be divided into three tiers: rigid procurement commitments from Google, Amazon, and Microsoft total about $269.4 billion, plus about $161.2 billion in five-year equipment leases from Broadcom, for a combined total of about $430.6 billion; if agreements with SpaceX AI, Nscale, Lambda, and others are included, the seven named suppliers total about $637 billion; and if the various line items in the prospectus notes are added together, the broadest measure is about $850 billion. The prospectus shows that among Anthropic's compute spending commitments to six partners, many contracts have terms of the next 7 to 10 years and are mostly non-cancelable, or require payment regardless of whether the capacity is used. For reference, Anthropic's revenue over the past 12 months was about $35.3 billion; calculated on the $850 billion basis, its compute commitments are about 24 times that.
Compute commitments are rigid costs locked in years in advance, while revenue only occurs after customers call on the capacity based on usage. About 80% of the company's revenue comes from enterprise pay-as-you-go API and token usage, with customers paying for what they use. In the first nine months of this year, the top five customers contributed about 25% of revenue, but most top customers have not signed long-term cooperation agreements. This relationship is gradually creating a mismatch: revenue depends on how much AI customers are willing to use in the future, while costs depend on how much compute the company has already committed to buy today. If demand exceeds expectations, pre-locked compute is a barrier; if usage growth slows, rigid bills will instead devour profits and cash flow.
Book losses further amplify outside concerns. In the first half of the year, Anthropic recorded a GAAP net loss of $80.748 billion, of which about $72.388 billion came from fair value changes in convertible notes, derivative liabilities, and SAFEs — the higher the company's valuation, the more valuable these convertible notes become, and the larger the book loss. Although this portion is not a current cash outflow, and the related losses largely stop after conversion upon listing, the book loss does contain real costs. As of the end of June, about $26.6 billion in stock-based compensation expenses had not yet been recognized, and are expected to be fully recognized over about 2.1 years; net operating cash inflow in the first half was $3.9 billion, which also included the add-back of about $2.2 billion in non-cash expenses. Its cash-generating ability is improving, but it is not yet thick enough to ignore the bills. Guosen Securities analyzed that about 80% of Anthropic's revenue comes from token-consumption-based billing, with net revenue retention near 900% and no single customer contributing more than 10%, a revenue quality and stickiness rarely seen among large-model companies, but reliance on three major cloud channels for distribution of nearly half of revenue is a major concern.
It is worth noting that Anthropic also directly wrote "what it will not do" into the prospectus. As a Delaware public benefit corporation, it pledges not to do advertising or video generation and prohibits the use of its models for fully autonomous weapons and mass surveillance. In April of this year, the company trained Claude Mythos Preview, which had overly strong cybersecurity capabilities, and judged that broad release without safeguards would pose too much risk, so it delayed public release and opened it only to a small number of institutions for cyber defense purposes. The company acknowledged in the prospectus that this means giving up substantial revenue. At the same time, government contract revenue currently accounts for only about 1%, and relations with Washington have also seen friction.
A Stress Test for the AI Valuation Chain
Anthropic's listing performance affects the nerves of the entire AI industry, and if IPO pricing falls short of expectations, the shock will ripple outward layer by layer. Jai Das, co-founder of venture capital firm Sapphire Ventures, said: "If its listing performance falls below expectations, it will be a bit like a house of cards." For OpenAI, Anthropic's IPO is also a mirror — the former has already secretly applied to go public but said it may not actually list until 2027. If Anthropic's listing performance is lackluster, it could further affect OpenAI's IPO and prompt investors to re-examine the high valuations of AI companies. At the same time, Federal Reserve rate hikes and rising U.S. Treasury yields also mean it is becoming more expensive for companies to issue debt to raise funds.
However, some views hold that market worries are excessive. Das compared current AI infrastructure investment to the telecommunications and internet bubble of the late 1990s — at that time fiber optic equipment flooded the market and many companies collapsed, but the infrastructure was ultimately fully utilized. "In the long run, there is still no shortage of demand for AI infrastructure in the market." Rohit Kulkarni, a partner at Burnham Park Capital Markets, said Anthropic's financial condition is "better than the market had feared, but more negative news cannot be ruled out afterward." Snowflake CEO Sridhar Ramaswamy expressed support for Anthropic's listing: "Anthropic and OpenAI have now proven that they have the ability to stay at the forefront of AI and continue to drive innovation, and we strongly support their listings." That said, Snowflake is also testing low-cost open AI models from China's Moonshot AI, DeepSeek, and Zhipu. David Trainer, CEO of financial research firm New Constructions, wrote in a report last month that the AI boom is built on the foundation of cheap credit and "always available" access to capital markets. If AI capital liquidity dries up and there are clear signs of a 20% sharp downside risk for the S&P 500, most of that would be concentrated in the valuations of AI-related stocks. PIMCO analysis pointed out that as of September 23, AI-related issuers already accounted for nearly a quarter of dollar-denominated non-financial corporate bond issuance, up from less than 13% in 2025 and about 4% in 2024; against the backdrop of rising U.S. Treasury yields and higher financing costs, AI companies are still crowding into debt issuance, and the concentration of AI-related debt in the credit market continues to rise. JPMorgan Asset Management estimates that complete data center construction may require about $5 trillion in funding by 2030, of which about $2 trillion may be financed through the investment-grade credit market.
Anthropic has already proven that frontier large models can quickly generate enormous revenue, and that a single quarter's operations can cover currently recognized costs. But quarterly profitability proves only that the business model has the potential to deliver profits; it has not yet proven that such profitability can survive a growth cycle. What it must prove next is whether revenue growth can continue to outrun the ladder of compute costs, whether customer usage can justify long-term investment, whether improved model efficiency can offset pricing pressure, and whether safety governance can remain stable under profitability requirements. This IPO is not just pricing one company; it is also testing a business model: when AI companies sign the next decade's bills in advance, will the public market be willing to keep paying for the assumption that "growth outruns costs"?
Comments