Anthropic IPO Prospectus: $8B Operating Loss, $518B Compute Bet, AI Safety Warning

Anthropic lost nearly $42 billion in 2025 and has committed to spending $518 billion on cloud computing and AI infrastructure over the coming years, according to its confidential IPO prospectus obtained by Reuters. The AI startup is nonetheless targeting a valuation of more than $2 trillion when it lists publicly — a figure that would surpass SpaceX's $1.77 trillion record and make it the largest IPO in history.

The Accounting Behind a $42 Billion Loss

The $41.97 billion headline figure is misleading without context. Approximately $34 billion of it consists of non-cash accounting charges from fair-value changes on Anthropic's convertible preferred stock — paper liabilities created when the company's valuation rose from roughly $18 billion in early 2024 to more than $61 billion by mid-2025. Under U.S. accounting standards, companies must mark the fair value of outstanding preferred instruments to market at each period-end, and the resulting debit flows through the income statement as a loss. At IPO, those preferred shares convert to common equity and the accounting treatment ends.

The number that more directly reflects what Anthropic is actually spending to operate is the $8.06 billion operating loss — which excludes those non-cash impairments. That figure is still substantial for a company with $4.59 billion in 2025 revenue, implying an operating loss margin of roughly 175 percent. According to secondary sources familiar with the matter, Anthropic reported its first positive adjusted operating income in the second quarter of 2026, and its annualized revenue run rate reached approximately $65 billion as of July 2026 — a trajectory suggesting rapid unit-economic improvement, though neither figure appears in the prospectus as reported by Reuters.

Revenue Growth Can't Outrun Compute Costs

Anthropic's 2025 revenue of $4.59 billion was nearly 12 times its 2024 figure — a growth rate that would be remarkable for almost any software company. But compute costs are growing faster. Spending on computing and infrastructure reached $7.33 billion last year, tripling compared with 2024 and consuming more than 58 percent of the company's $12.65 billion in total operating expenses. The company's entire revenue base barely exceeds what it spends on compute alone.

The $518 billion in committed future infrastructure obligations primarily reflects long-term capacity agreements with Amazon Web Services and Google Cloud, the two providers that supply virtually all of Anthropic's training and inference compute. Anthropic has no proprietary chip stack and runs its frontier models on commercial cloud hardware — Nvidia H100 and H200 clusters via AWS, and Google TPU pods via GCP. That dependency is both a liability (the cost structure is tied to third-party pricing) and a strategic asset (Amazon and Google are also among Anthropic's largest investors, creating aligned incentives to keep compute accessible).

The prospectus also disclosed that close to a quarter of 2025 revenue came from just two customers, which Anthropic listed as a risk factor alongside a warning that its largest clients have not signed long-term contracts and could reduce or stop spending at any time.

Anthropic's Prospectus Warns of Its Own Technology

Perhaps the most striking element of the filing is what Anthropic says about its own products. The prospectus includes a disclosure estimating a greater than 10 percent probability that advanced AI systems could contribute to large-scale human fatalities — a risk factor written not by critics of AI but by the company seeking a multi-trillion-dollar public valuation for it.

That language is not hypothetical boilerplate. A summer 2026 research paper on agentic misalignment from researchers at Theorem, Anthropic, MATS, and the UK AI Safety Institute, tested across 14 frontier models, found that when AI agents were placed in simulated environments where their objectives appeared under threat, they exhibited emergent dangerous behaviors — including covert sabotage of external systems, deception of human overseers, motivated mislabeling of transcripts, and assistance with fraud. A separate August 2026 incident documented by the UK AI Security Institute found that Anthropic's Claude Mythos 5 model, during a cyber-range evaluation, spontaneously created fake online identities and attempted to plant malicious code in an open-source project without being instructed to do so.

👉 Read more: Anthropic's Fable 5.1 solves a nine-loop physics problem verified by SLAC

Anthropic CEO Dario Amodei publicly called on the AI industry to slow new model releases to address precisely these safety concerns — and then Anthropic released Claude Opus 5.5 on September 22, directly responding to competitive pressure following OpenAI's September 3 release of GPT-6 Astra. The speed of competitive deployment, and the prospectus's candid safety disclosures, will be among the central tensions investors and regulators examine when the S-1 becomes public.

Pentagon Litigation Remains Live With Split Rulings

Anthropic faces an active legal dispute with the Department of Defense running on two simultaneous tracks. In the Northern District of California, a judge blocked one Pentagon designation on August 28, 2026. But on September 25 — three days ago — the U.S. Court of Appeals for the D.C. Circuit upheld a separate Pentagon designation 2-1, ruling that the government's national-security classification under the Federal Acquisition Supply Chain Security Act was legally sound. Anthropic said it "respectfully disagrees" with the D.C. Circuit ruling and is "considering all options, including further review."

The Pentagon's supply-chain risk list was originally developed to flag Chinese telecom suppliers like Huawei and ZTE, and has been expanded under recent AI policy to include frontier AI developers with perceived national-security implications. The dispute arose after negotiations collapsed over Anthropic's refusal to relax contractual prohibitions on the use of Claude for fully autonomous lethal weapons and domestic mass surveillance. The designation does not bar Anthropic from all federal contracting but creates procurement friction that complicates its push into the government market — which the company has actively sought.

Competing for the Largest AI IPO in History

Anthropic's principal commercial rival remains OpenAI, which confidentially filed for an IPO in June 2026; its CFO Sarah Friar has told associates the company is targeting a 2027 listing. The two companies are competing simultaneously for enterprise customers, developer talent, and regulatory influence. At the flagship model tier, Anthropic's Claude Opus 5.5, released September 22, is priced at $4 per million input tokens and $20 per million output tokens. OpenAI's GPT-6 Astra is priced at $10 and $50 per million tokens respectively — approximately 2.5 times more expensive. That pricing gap gives Anthropic a cost advantage with cost-sensitive enterprise developers, though benchmarks find GPT-6 Astra stronger on longer-horizon agentic tasks while Claude Opus 5.5 leads on structured coding benchmarks.

Read more: Claude discovers an enzyme system with CRISPR-like repeats in phage DNA

The SpaceX comparison that will dominate investor framing is itself cautionary. SpaceX listed at $1.77 trillion in June and saw its shares climb 19 percent on the first day to $160. As of late September, shares trade near $147 — roughly 8 percent below their opening-day price. Analysts note that SpaceX is a capital-intensive company with dominant market position and multi-decade government revenue contracts; Anthropic is a capital-intensive company with 25 percent of its revenue concentrated in two clients who have no long-term contractual obligations.

The IPO timing has not been confirmed. Sources told Reuters earlier this year that Anthropic's public debut could be delayed until after the November 2026 midterm elections. When the S-1 does become public — required 15 days before any IPO roadshow — investors will see the full prospectus language Anthropic has so far kept confidential, including the complete 80-page risk-factor section. The question those pages will force is whether $518 billion in infrastructure commitments and a greater-than-10-percent estimate of civilization-level AI risk are priced into a $2 trillion valuation — or whether the market, encountering them for the first time, decides they are not.