Originally published on deanlee.info.
The financial disclosures inside Anthropic's initial public offering prospectus, reported by Reuters this week, show a commercial ramp with very few precedents in enterprise software. Annual revenue jumped twelve-fold in 2025, climbing from roughly $386 million in 2024 to nearly $4.6 billion. Almost no enterprise vendor has compounded top-line sales at that pace after crossing the billion-dollar mark.
Even the headline $42 billion GAAP net loss is less alarming to an institutional desk than it looks in a general news summary. Roughly $34 billion of that figure comes from a non-cash accounting charge tied to earlier fundraising rounds. Because Anthropic issued financing instruments that convert into equity, and because its private valuation kept climbing toward a $965 billion mark in May and a targeted $2 trillion public valuation later this year, accounting rules forced the company to book the rising fair value of those liabilities as an expense. Strip out that paper revaluation, and the 2025 operating loss was just over $8 billion against $20.28 billion in year-end cash and short-term investments. On pure cash runway, a growth buyer can argue that top-line velocity covers the burn.
The harder test for public equity buyers sits in the contract terms on either side of the income statement. In 2025, Anthropic spent $7.33 billion on compute and infrastructure alone, triple its 2024 outlay. That single line item accounted for 58 percent of the company's $12.65 billion in total operating expenses and equaled roughly 159 percent of gross revenue. Before paying a single research scientist or sales engineer, Anthropic handed $1.59 to cloud and hardware providers for every dollar a customer paid to use Claude.
Looking forward, the prospectus discloses plans to spend $518 billion on cloud, computing, and infrastructure obligations in coming years. Pair that half-trillion-dollar commitment schedule with two admissions in the risk factors. Nearly a quarter of Anthropic's 2025 revenue came from just two customers, and many of its largest enterprise clients are not locked into long-term contracts and can reduce or stop spending whenever they choose.
In derivatives terms, Anthropic is running an unhedged fixed-for-floating swap between its input obligations and its customer receivables. On the pay leg, a frontier model lab must sign multi-year, fixed-commitment capacity agreements with hyperscale cloud landlords years before a model checkpoint finishes pre-training. On the receive leg, enterprise customers buy intelligence by the million tokens on short-term usage terms.
During the cloud software wave of the 2010s, enterprise SaaS companies went public with the opposite contract polarity. Firms like Workday and Salesforce spent heavily upfront on sales and implementation to win an enterprise account, then locked that customer into multi-year seat subscriptions backed by high database migration costs. Meanwhile, their marginal hosting cost on public cloud infrastructure was around twenty cents on the dollar, leaving 75 to 80 percent gross margins to amortize the acquisition burn over the life of the contract.
Standalone model labs have inverted both ends of that ledger. Their primary cost is leased accelerator capacity, priced by suppliers with immense balance-sheet power. Their product is an API endpoint that corporate engineering teams deliberately wrap inside model-agnostic routing layers so they can switch providers in an afternoon.
The past three weeks showed how unforgiving that spot-market revenue structure is in practice. Earlier this month, CEO Dario Amodei publicly urged the global AI industry to slow capability releases after controlled internal tests showed autonomous models sabotaging code and assisting fraud. Yet once OpenAI launched GPT-6 Astra and began pulling enterprise developer traffic, Anthropic could not rely on multi-year contract lock-in to hold its customer base. Last week, the company rolled out Claude Opus 5.5 to defend its daily token volume ahead of its roadshow. When your largest clients can re-route workloads without penalty, sitting out a capability cycle means watching revenue leave while the $518 billion infrastructure meter keeps running.
The destination of those infrastructure dollars shows who holds the bargaining power in the AI stack. Anthropic's earliest and largest strategic partners are Amazon and Google. Both tech giants invested billions of dollars into the startup while simultaneously serving as the primary cloud landlords training and hosting Claude.
For a hyperscaler, this arrangement is the cleanest vendor-financing loop in corporate finance. A cloud provider invests cash into a frontier lab at a rising private mark and books an unrealized equity gain as the lab targets a $2 trillion valuation after the November midterm elections. At the same time, the cloud provider locks in tens of billions of dollars in guaranteed, high-utilization compute backlog. If the standalone model layer eventually finds durable pricing power, the hyperscaler participates through its equity stake. If frontier inference commoditizes into a price war between Anthropic, OpenAI, and open-weight releases, the hyperscaler still collects its contracted compute payments at the top of the operating waterfall.
Until now, that circular trade was funded in private markets by venture syndicates, sovereign wealth funds, and the cloud landlords themselves. Taking Anthropic public after the November elections changes the buyer base. Institutional asset managers and retail index funds do not receive AWS or Google Cloud revenue rebates when they buy common shares in an IPO. They only own the residual cash flow left over after the $518 billion compute obligation is serviced.
SpaceX's June debut at a $1.77 trillion valuation proved that public markets can still digest mega-cap listings under elevated interest rates, with shares trading around $147 after pricing at $135. SpaceX, however, owns the physical launch pads and orbital spectrum rights that generate its cash flows. A standalone AI lab seeking a $2 trillion valuation at more than 400 times trailing revenue owns model weights whose commercial premium decays the week a rival finishes its next training run, while its cloud suppliers hold the long-term contracts and sit senior in the cash flow waterfall.
Top comments (0)