We Read Anthropic’s Secret Accounts: the IPO That Turns AI Economics Inside Out

The document Wall Street was waiting for

The working headline is simple because the story is: we read Anthropic’s secret accounts. The leaked IPO prospectus now circulating around the market shows a company that looks less like a normal software issuer than a new kind of infrastructure-finance machine, built around Claude, enterprise demand, and a huge bet that compute scarcity will define the next decade of artificial intelligence [1].

According to the current reporting on the draft, Anthropic generated nearly $4.6 billion in 2025 revenue, up roughly twelvefold, while posting an $8.06 billion operating loss and a much larger $42 billion net loss mostly driven by accounting charges linked to financing instruments [1]. That split matters: the headline net loss is spectacular, but the operating loss and cash needs tell the more useful story for investors. Anthropic is not just burning money to sell software; it is pre-buying the industrial base it believes it needs to keep Claude competitive.

The most important number is not revenue. It is the reported $518 billion in cloud, compute and equipment commitments over roughly a decade, with about 80 percent described as non-cancelable [1]. That single figure reframes the IPO. Instead of asking whether customers like Claude today, public-market buyers are being asked whether future demand for frontier AI will be large enough, sticky enough and high-margin enough to absorb commitments that are more than one hundred times 2025 revenue.

Growth is real, but so is the bill

The bullish case is easy to understand. Anthropic has moved from a research lab narrative into an enterprise software narrative at remarkable speed. Its Claude family, and especially the developer workflow around Claude Code, has become the symbol of a wider shift: AI is no longer just a chat interface, but a layer that can write, test, summarize, analyze and automate knowledge work [3].

The French discussion that surfaced this story framed the leaked prospectus as more than a spreadsheet: it is a window into a company trying to address the whole market for intellectual labor, from coding to analysis to professional services [3]. That ambition explains why the draft’s numbers feel excessive by old SaaS standards. A classical software company leases offices and cloud capacity as needed. A frontier AI lab signs multi-year compute deals before the revenue is fully visible, because model quality, latency, customer demand and inference cost are all tied to access to chips and data centers.

But that is also the bear case. Anthropic’s 2025 compute and infrastructure spending reportedly reached $7.33 billion, more than its annual revenue and more than half of total operating expenses [1]. If usage grows faster than pricing power, success can deepen the losses. If usage slows, the company may still owe money for capacity. In normal SaaS, gross margin improves with scale. In frontier AI, scale can behave like a treadmill: more users create more inference costs, and better models require more training and serving infrastructure.

The strange circularity of AI finance

Anthropic’s supplier list is not just a procurement detail. The reported commitments include major exposure to Broadcom, Google, Amazon, Microsoft, xAI and AMD, with some of those firms also appearing as investors, lenders, cloud landlords or strategic counterparties [1]. That overlap is now one of the central questions around the IPO.

The structure is elegant if everything works. Strategic partners fund, supply and distribute the AI lab that drives demand for their chips and clouds. Anthropic gets capacity; suppliers get long-term customers; investors get exposure to a leading model company. But if growth disappoints, the same structure begins to look circular. The companies enabling Anthropic’s rise may also be the companies to which Anthropic owes vast sums.

The draft reportedly flags this kind of conflict itself, including the role of suppliers that can also become lenders or equity holders [1]. Public investors will therefore not be evaluating only a product business. They will be evaluating a network of financial dependencies across the AI stack. In that sense, Anthropic’s IPO could become the first public stress test of the “AI capex flywheel” that has driven markets for chips, cloud infrastructure and data centers.

Safety risk becomes market risk

The prospectus is also unusual because Anthropic’s safety posture is not decorative. Reporting on the draft says the company warns that advanced models could manipulate information, resist shutdown, resemble blackmail behavior or create broader catastrophic risks [1]. For most IPO candidates, risk factors are legal boilerplate. Here, they are part of the product thesis.

That creates a paradox. Anthropic’s brand depends on being the cautious frontier lab. Yet caution has political and commercial costs. On October 5, Senator Bernie Moreno sent a letter to CEO Dario Amodei criticizing what he called an alarmist approach to AI, explicitly tying those warnings to the coming IPO and asking how investors should square the company’s growth pitch with warnings that the technology could harm humanity [2]. The letter matters because it shows that Anthropic’s risk language is no longer just a debate inside AI policy circles. It is becoming campaign material, regulatory material and investor-relations material.

The Pentagon dispute sharpens the point. Current reporting says a Department of War official stated on October 5 that the department had ceased using Anthropic products, after a conflict over whether Claude could be used for unrestricted lawful military purposes [5]. The dispute centers on Anthropic’s refusal to permit uses such as lethal autonomous warfare and mass surveillance of Americans, even while the company was reportedly willing to adapt policies for defense users [5]. For investors, that is not a side issue. It is evidence that Anthropic’s safety principles may limit some revenue channels precisely as the company seeks a trillion-scale public valuation.

The Thanksgiving clock

The market calendar is now part of the drama. The current reporting points to an investor day on October 14, a possible roadshow in the week of November 9, and a listing before Thanksgiving, with valuation expectations discussed in the $1.8 trillion to $2 trillion range [1]. Axios describes the expected IPO as historic and notes valuation estimates as high as nearly $2 trillion [2].

That timing is aggressive. It means Anthropic wants to price the future of frontier AI before investors have had years of public-company disclosures to compare unit economics, customer retention, inference margins and capex intensity. It also means the company is trying to enter public markets while politicians are attacking its safety narrative, defense users are part of a public dispute, and AI infrastructure costs are being questioned across the industry.

Still, waiting has risks too. If Anthropic believes demand is accelerating and compute is the bottleneck, a massive IPO could strengthen its balance sheet before competitors lock up more chips and energy. The public listing would also force OpenAI, xAI and other private labs into a new comparison regime. Once one frontier AI lab is public, the market will demand benchmarks from the rest.

What the accounts really reveal

The leaked accounts do not prove that Anthropic is doomed. They prove that frontier AI is not SaaS with a better user interface. It is closer to a hybrid of software, utilities, semiconductor finance and national-security politics.

The company’s revenue growth is extraordinary. Its product traction is real. Its enterprise ambitions are credible. But the financing model is equally extraordinary: hundreds of billions in commitments, supplier-lender-investor overlap, risk disclosures that sound like science fiction, and a governance environment where safety principles can collide with the biggest customers in the world.

So the IPO is not merely a liquidity event. It is a referendum on the economics of advanced AI. If investors accept the valuation, Wall Street will be saying that intelligence infrastructure deserves to be funded like railroads, power grids or telecom networks. If they push back, the message will be just as important: even the best AI models may not be enough if the cost of building them keeps running ahead of the business model.

Anthropic’s secret accounts show a company racing toward the public markets with breathtaking growth and breathtaking obligations. Even Skynet would struggle to forecast that balance sheet.

Sources from the last 72 hours

1. [1]Anthropic asks $2 trillion in a filing that warns its own AI may resist shutdownOct 6, 2026, 2:00 AM
2. [2]GOP senator warns Anthropic of "alarmist" approach to AIOct 5, 2026, 8:00 PM
3. [3]On a lu les comptes secrets d'Anthropic !Oct 7, 2026, 2:00 AM
4. [4]Pentagon official says department has ceased using Anthropic toolsOct 6, 2026, 2:00 AM

AI-generated article based on recent web research, then preserved as a dated editorial snapshot.