Anthropic's $2T IPO Has a $518B Problem

📖 Read the full version with charts and embedded sources on ComputeLeap →

Anthropic's leaked IPO prospectus makes the case for a $2 trillion valuation — the largest in IPO history — on the back of revenue that grew twelvefold to $4.6 billion in 2025 and hit $11.5 billion in Q2 2026 alone. The headline numbers are staggering. But buried inside the 261-page document is a five-word admission that should make every investor pause: enterprises can walk at any time.

The prospectus warns that "many of its largest customers are not tied into long-term contracts and could reduce or stop spending." Two unnamed customers each generated 12% of Anthropic's 2025 revenue. That is nearly a quarter of the company's top line sitting on handshake agreements in a market where OpenAI, Google, and a dozen open-weight alternatives are one API swap away.

View original article on Fortune →

Here is the asymmetry that nobody is underlining: Anthropic locked itself into $518 billion of non-cancelable compute commitments over the next decade. The revenue to pay for that compute? Month-to-month.

The $518 Billion Bet — And Who's Really Locked In

The prospectus details infrastructure commitments with six partners that read more like a sovereign debt schedule than a startup's cost base:

• Google: $111.1 billion (through 2033, with makeup-payment obligations)
• Amazon: $110 billion (through 2036, similar makeup terms)
• Broadcom equipment leases: $161.2 billion (largely non-cancelable)
• xAI: $84.5 billion for Nvidia-based chips (90-day cancellation notice)
• Microsoft: $31.4 billion (through 2033)
• AMD compute capacity: $20+ billion

About 80% of that $518 billion total is non-cancelable or payable regardless of whether Anthropic actually uses the capacity. In Q1 2026, the company spent 71 cents on compute for every dollar of revenue — a ratio that makes cloud margins look luxurious.

ℹ️ The asymmetry in a single line: Anthropic owes its infrastructure partners $518B whether customers show up or not. Its customers owe Anthropic nothing.

The company's defense, buried in the prospectus, is that demand for advanced AI will be "limited principally by the availability of compute." Translation: they are betting the entire company that demand will grow faster than their already-massive supply commitments. It is a reasonable bet if you believe in exponential AI adoption. It is a catastrophic one if enterprise demand plateaus or shifts to cheaper alternatives.

View original article on PYMNTS →

The Churn Risk Nobody Wants to Talk About

Enterprise revenue concentration is the prospectus's quiet time bomb. Here are the numbers that matter:

• 80% of revenue comes from enterprise customers (vs. ~40% for OpenAI)
• Two customers generated nearly 25% of 2025 revenue
• 1,000+ businesses spend over $1 million annually (as of April 2026, doubled in under two months)
• Enterprise contracts lack long-term commitments in many cases

The bull case frames this as a feature: enterprise revenue is "stickier" than consumer subscriptions. Anthropic wins 70% of head-to-head enterprise deals against OpenAI. The product is clearly good enough that customers choose it on merit.

But "sticky" and "contractually committed" are not the same thing. ServiceNow reportedly exceeded its full-year Claude budget due to price increases — a data point that cuts both ways. Usage is high (bullish), but so is price sensitivity (bearish). When Anthropic's post-IPO board starts pushing for quarterly revenue growth, the temptation to squeeze pricing on captive-but-uncommitted customers will be enormous.

Forrester's analysts saw this coming. In a note titled "Anthropic's Proposed IPO Will Change The Economics Of Enterprise AI," they advised enterprise buyers to "resist long-term contracts to limit vendor lock-in and identify where you're exposed to usage-based cost increases." They also noted that some organizations have "already pulled back [from Claude Code] because of consumption costs."

View original article on Forrester →

That is not a theoretical risk. It is already happening.

The Prediction Markets Have Already Priced the Split

Polymarket tells a nuanced story that the headline coverage misses. As of this writing:

• Best AI model, October 2026: Google 65%, Anthropic 36% (down 53% this week on Gemini 4 Argon)
• Best AI model, end of 2026: Anthropic 52%, Google 38%

The near-term and long-term markets are in open disagreement. Traders give Google the month but hand Anthropic the year. The Anthropic IPO is timed to ride the long-dated sentiment — get public while the market still prices dominance, before the commodity-pricing war narrows the window.

But here is what the prediction markets cannot price: the gap between "best model" and "best business." Anthropic may have the best model and still face churn if enterprises decide that 90%-as-good alternatives (Gemini, GPT-6.1 Sol, open-weight Qwen/Llama) at 30% of the cost are good enough for production workloads.

The IPO Math: Growth Rate vs. Contract Duration

Let us stress-test the $2 trillion valuation against the churn-risk disclosure.

Anthropic's annualized revenue run-rate hit $65 billion by July 2026, with investors expecting $100-120 billion by year-end. At a $2 trillion valuation, that implies roughly 20x forward revenue — aggressive but defensible for a company growing 10x+ year-over-year.

But that multiple assumes the revenue is durable. Traditional SaaS companies trade at high multiples because they have contractual recurring revenue with 120%+ net retention. Anthropic has neither long-term contracts nor, based on the prospectus, a way to prevent its largest customers from cutting spend overnight.

The Series H round at $965 billion was priced on growth trajectory. The IPO at $2 trillion is being priced on the assumption that growth trajectory continues. But every dollar of that growth is earned quarter-by-quarter from customers who can leave at the end of any billing cycle.

⚠️ Contrarian Corner: The bull case is stronger than it looks. AI model switching costs are real even without contracts — enterprises build evals, integration layers, prompt libraries, and fine-tuned workflows around a specific model's behavior. Ripping out Claude and replacing it with Gemini is not a weekend project for a Fortune 500 company. The churn-risk disclosure might be legally required boilerplate rather than an operational signal. Anthropic's 70% win rate on enterprise deals suggests the product moat is deep enough that formal lock-in is unnecessary.

The Loss Problem: $42 Billion and Counting

Revenue growth is only half the equation. The prospectus reveals an operating loss of $8.06 billion in 2025 (up from $2.98 billion in 2024) and a staggering net loss of $41.97 billion — a figure inflated by accounting treatment of equity and convertible instruments, but still an attention-grabber on a cover sheet.

The path to profitability appeared in Q2 2026, with analysts estimating Anthropic posted its first-ever operating profit of about $559 million. But that margin is razor-thin against the $518 billion infrastructure bill looming over the next decade.

If enterprise customers can walk, and the compute bill cannot be renegotiated, Anthropic is running the most expensive game of musical chairs in tech history. The music sounds great right now — $11.5 billion quarterly revenue, 80% enterprise mix, first-ever profit. But the chairs (non-cancelable compute contracts) outnumber the players (uncommitted enterprise customers) by a factor that should concern any IPO investor.

What the Community Is Saying

The Anthropic IPO has drawn sharp reactions across the financial and tech commentary ecosystem. On Hacker News, the leaked prospectus thread drew 150 comments and 143 points, with developers dissecting the compute-to-revenue ratio and questioning whether enterprises would stay locked in without contracts.

View discussion on Hacker News →

A separate HN thread titled "Anthropic's IPO Prospectus Is a Fucking Doozy" captured the more visceral reaction, with commenters zeroing in on the 80 pages of risk factors and the existential-risk disclosure.

View discussion on Hacker News →

Substack analysts at Heavy Lifting were among the first to flag the risk-factor asymmetry, noting that 80 of the prospectus's 261 pages — nearly a third — are dedicated to risk disclosures, including the existential warning that AI "could pose catastrophic or existential risks to humanity."

View original article on Substack →

On X, Paul Graham observed that "AI startups grow really fast — that's why valuations got so high," framing the $2T number as a natural consequence of growth velocity rather than irrational exuberance.

View original post on X →

Bill Gurley countered with his emphasis on how Anthropic thinks about consciousness and enterprise responsibility — a more sober perspective that highlights the gap between "mature enterprise security" and "braggadocios press releases."

View original post on X →

Forbes' analysis noted that at a 30x revenue multiple, Anthropic's implied valuation would exceed $3 trillion — making the $2 trillion target look like a discount. The tension between the growth story (justify $3T) and the risk story (maybe $2T is too much) is the IPO's defining dynamic.

What This Means for You

If you are an enterprise AI buyer:

• Follow Forrester's advice: resist long-term contracts while you still can. Post-IPO Anthropic will face quarterly revenue pressure and will push hard for annual or multi-year commitments.
• Build your AI stack to be model-portable. Invest in abstraction layers (LiteLLM, custom routers) that let you swap providers without rewriting your application layer.
• Monitor pricing closely. The 71-cents-per-dollar compute ratio means Anthropic has very little room to absorb cost increases — they will pass them through.

If you are evaluating Anthropic as an investment:

• Stress-test the $2T valuation against actual contract duration, not just ARR. Ask: what is the revenue if the top 10 customers cut spend by 50%?
• Watch the Q3 2026 earnings for net revenue retention metrics — the prospectus conspicuously omits them.
• Compare Anthropic's enterprise contract structure to traditional SaaS companies. Snowflake, Databricks, and Palantir all publish net retention rates because they have long-term contracts. Anthropic's silence on this metric is itself a data point.

If you are building in the AI space:

• The Anthropic vs. OpenAI platform war is entering its financial phase. Both companies will be public by 2027, and Wall Street's quarterly cadence will reshape how they price, package, and compete for enterprise deals.
• The AWS dependency is a double-edged sword: Amazon is both Anthropic's largest distribution partner and its second-largest infrastructure creditor. That alignment works until it does not.
• Understanding the valuation trajectory from $1T to $2T reveals how much of the premium is model quality vs. market timing.

💡 The practitioner takeaway: The IPO prospectus is the best argument yet for multi-model architectures. Any enterprise that goes all-in on a single AI vendor — whether Anthropic, OpenAI, or Google — is accepting concentration risk that the vendors themselves admit they cannot contract around. Build for portability now, while you still have leverage.

The Bottom Line

The Anthropic IPO prospectus is a masterpiece of strategic framing. It leads with the growth story ($11.5B quarterly revenue, 12x annual growth, 80% enterprise mix) and buries the structural risk (no long-term contracts, extreme customer concentration, $518B in non-cancelable compute).

The market will probably reward the growth story — at least initially. AI is the trade of the decade, Anthropic has the best model (per prediction markets), and the revenue trajectory is genuinely unprecedented.

But the churn-risk admission is the canary in the coal mine. When the next frontier model from Google or Meta or a well-funded open-weight competitor arrives, Anthropic's enterprise customers will not need to break any contracts to leave. They will just stop renewing. And Anthropic will still owe $518 billion to Google, Amazon, Broadcom, xAI, Microsoft, and AMD.

That is not a risk factor. That is the business model.

Originally published at ComputeLeap