Anthropic logo

Anthropic's confidential IPO prospectus has leaked, and the numbers are staggering. The Claude maker lost nearly $42 billion in 2025 while chasing a $2 trillion public market valuation — a gap so wide it makes every other AI startup's burn rate look quaint.

The financials, laid bare

Reuters obtained a copy of Anthropic's S-1 filing, first submitted confidentially to the SEC on June 1, 2026 (UTC). The document, reviewed by multiple outlets in the final days of September, offers the first unvarnished look at what it actually costs to run a frontier AI lab.

Metric 2025 2024 YoY Change
Revenue ~$4.6B ~$383M ~12x
Operating loss ~$8.1B ~$3.0B ~2.7x
Net loss ~$42B ~$8B ~5.3x
Compute & infrastructure spend ~$7.33B ~$2.5B ~2.9x
Total operating expenses ~$12.65B — —

Source: Anthropic S-1 prospectus via Reuters, Read Us 24x7, Evermx

The $42 billion net loss dwarfs the $8.1 billion operating loss because it includes non-cash charges — primarily writedowns of liabilities tied to previous fundraising rounds and stock-based compensation. The operating figure is the cleaner read on core business health, and even there, Anthropic is losing roughly $1.76 for every dollar of revenue.

Revenue composition tells another story. About $3.8 billion came from consumption-based customers paying per token, while subscription revenue hit $789 million. The company expects usage-based pricing to remain "the substantial majority" of revenue going forward — a model that scales with customer adoption but also scales with compute costs.

47% of revenue runs through Amazon and Google

Here's the part that should give investors pause. Nearly half of Anthropic's 2025 sales — $2.16 billion, or 47% — were routed through cloud marketplaces operated by Amazon and Alphabet's Google. Anthropic paid those platforms roughly $351 million in distribution fees, meaning the cloud giants collected about 16 cents on every dollar of those sales.

This isn't just a vendor relationship. Amazon and Google are simultaneously:
- Investors — having poured tens of billions into Anthropic
- Compute suppliers — selling the GPUs and cloud capacity Anthropic needs to exist
- Distribution partners — collecting 16% of marketplace revenue
- Direct competitors — running their own AI models (Bedrock/Amazon Nova, Gemini)

Anthropic acknowledged the conflict in the filing, warning that reliance on a limited number of partners "creates complex dynamics that could give rise to conflicts of interest and adversely affect our access to compute." The cloud providers also gain visibility into Anthropic's pricing and commercial terms — information they could use to allocate compute capacity or compete more aggressively.

A November 2025 cloud deal with Microsoft added a third giant to the mix, but the dependency structure remains the same: Anthropic cannot function without the very companies trying to put it out of business.

$518 billion in compute commitments

The infrastructure numbers are almost impossible to contextualize. At the end of 2025, Anthropic held $54.6 billion in non-cancellable hosting and computing commitments. By early 2026, total long-term commitments exceeded $417 billion, covering 3.5 gigawatts of dedicated computing capacity. The prospectus outlines $518 billion in planned infrastructure spending overall.

For scale: 3.5 gigawatts is roughly the electricity consumption of a mid-sized European country. The commitments include over $100 billion to AWS across a decade and a $15 billion annual agreement with SpaceX running through May 2029.

These are not optional expenses. They are contractual obligations that must be paid regardless of whether Anthropic's revenue grows fast enough to cover them. If the AI adoption curve flattens — if enterprise customers stop increasing token consumption — Anthropic is locked into paying for capacity it may not need.

The valuation question: $2T or $1.5T?

Anthropic is targeting a Nasdaq listing as early as mid-October 2026 (UTC), potentially before the U.S. midterm elections, with an asking valuation of approximately $2 trillion. Morgan Stanley, Goldman Sachs, and JPMorgan are leading the offering.

That $2 trillion figure would more than double the $965 billion post-money valuation from Anthropic's $65 billion Series H in May 2026 — a four-month doubling that assumes explosive revenue growth continues indefinitely.

The market isn't uniformly convinced. The Information reported on September 30 (UTC) that while Silicon Valley VCs are pushing for the $2T figure, two large Wall Street institutions consider $1.5 trillion more appropriate. That $500 billion gap reflects a fundamental disagreement: are you pricing Anthropic as a hyper-growth software company, or as a capital-intensive infrastructure business with razor-thin margins?

Seven co-founders will retain 50.1% voting power through a Founder LLC entity, meaning public investors will have limited say in governance even at a $2T valuation.

Why this matters

Anthropic's IPO is the first real test of whether public markets can absorb the financial profile of a frontier AI lab. Every other major AI company — OpenAI, xAI, Mistral — is watching. If Anthropic prices at $2T and holds, it unlocks a wave of AI public listings. If it prices at $1.5T or below, the entire private AI valuation complex reprices downward.

The $42 billion net loss, while inflated by non-cash items, signals something important: the economics of frontier AI are not yet proven at scale. Compute costs are growing almost as fast as revenue. The 12x revenue growth in 2025 is impressive, but it came alongside a 2.9x increase in infrastructure spending. If compute costs don't decline through efficiency gains, the path to profitability keeps receding.

The 47% cloud marketplace dependency is arguably the most underappreciated risk. Public market investors are being asked to value a company whose distribution, compute, and investor base are all controlled by its direct competitors. Amazon could decide tomorrow to prioritize its own Nova models on Bedrock. Google could shift Gemini to the top of Cloud AI search results. These aren't hypothetical scenarios — they're standard platform behavior.

The $518 billion in compute commitments creates a different kind of risk: a demand cliff. If enterprise AI spending slows — if the ROI on AI agents doesn't materialize as fast as CFOs hoped — Anthropic is stuck paying for 3.5GW of capacity while revenue growth decelerates. That's the worst possible combination for a public company.

The skeptical read

There's a legitimate argument that the $2T valuation is a marketing number designed to anchor negotiations, not a realistic clearing price. The gap between Silicon Valley enthusiasm ($2T) and Wall Street skepticism ($1.5T) is $500 billion — larger than the entire market cap of most S&P 500 companies. When the buy-side and sell-side disagree this sharply, the IPO price typically lands closer to the lower bound.

The Founder LLC voting structure also deserves scrutiny. At a $2T valuation, public shareholders are putting up enormous capital while receiving less than 50% of voting control. This is increasingly common in tech IPOs, but it concentrates risk: if the co-founders make a strategic misstep — say, overcommitting to compute at the top of the cycle — public investors can't vote them out.

What to watch next

This isn't just Anthropic's IPO. It's a referendum on whether the economics of building god-like AI can ever produce a god-like profit margin.