Anthropic passed OpenAI in quarterly revenue for the first time in 2026, filed a draft prospectus aiming at a valuation above $2 trillion, and disclosed in that same filing that it lost roughly $42 billion last year. Six months ago, the open question here was whether the frontier labs could ever convert revenue into enough cash to cover their own compute bill. The answer has arrived, and it is not a clean yes. It is that the companies selling the chips have started lending the labs the money to buy them.
When I last wrote about this in April, Anthropic had just crossed a $19 billion annualized run-rate and was throttling paying Pro and Max subscribers during peak morning hours because its capacity could not keep up with demand. That looked like the crisis. It was the overture. The structural tension, where every new customer costs more to serve than the last tranche of revenue covers, did not resolve. It scaled, and the industry’s answer was financial engineering.
The revenue flip that rewrote the headline
The single biggest change since spring is the order of the leaderboard. Anthropic overtook OpenAI in revenue for the first time in the second quarter of 2026, according to reporting from Yahoo Finance. Run-rate figures have been quoted anywhere from roughly $30 billion to above $65 billion depending on the month and the counting method, and that spread is not a rounding issue. As Axios explained, the two labs book a meaningful share of sales through cloud partners and calculate run-rate differently, so any head-to-head number deserves a caveat rather than a headline. The ranking flipped; the measurement is still soft.
What is not ambiguous is the trajectory and the valuation attached to it. Anthropic raised a roughly $30 billion Series G at a $380 billion valuation in February, closed a $65 billion Series H at a $965 billion post-money valuation in May, and is now targeting an IPO valuation above $2 trillion. That is more than a doubling of the private mark in a matter of months, on a company that its own prospectus shows is deeply unprofitable. (For the governance and timing side of that filing, see our coverage of the OpenAI delay and Anthropic’s $2 trillion sprint.)
Note what the filing does to a claim that circulated earlier this year: the idea that Anthropic was “about to post its first profit.” The actual document tells a different story. Revenue grew roughly twelvefold to about $4.6 billion in 2025, but operating losses exceeded $8 billion, and the headline $42 billion net loss includes around $34 billion in non-cash accounting charges tied to financing instruments that can convert into shares, not cash the company actually spent, per Fortune’s reporting on the prospectus. Strip the accounting noise and you still have a business losing billions in cash to grow. The growth is real. The margin is the problem, and it is the same problem it was in April.
The margin problem never left
OpenAI is the cleaner illustration because its numbers leak more often. In the first quarter of 2026 the company burned about $3.7 billion on $5.7 billion of revenue, according to reporting on its Q1 financials. Its reported gross margin improved to roughly 39 percent from 33 percent a year earlier, which sounds like progress until you read the direction of the adjusted figure. One analysis puts adjusted gross margin back down near 33 percent against a 46 percent internal target, because inference costs are climbing as usage grows. Inference spend alone ran about $8.4 billion in 2025 and is tracked near $14.1 billion for 2026.
That is the mechanism the whole crisis turns on. A model company’s cost of goods sold is the compute it takes to answer a query, so its margin gets worse exactly as its best customers lean in harder. The accounts running agents, not single chats, are the most valuable and the most expensive to serve at the same time. Analysts tracking the unit economics have OpenAI losing on the order of $1.22 for every dollar it earns and do not model cash-flow breakeven until 2029 or 2030. Revenue leadership, in this industry, is not a path to profit. It is a larger bill.
The crisis didn’t get solved. It got financed
Here is the development that actually changes the thesis. On October 1, Broadcom agreed to lend Anthropic up to $42 billion to lease Broadcom-built chips, an arrangement confirmed in filings and reported by Semafor and others. The $42 billion covers roughly a third of the $125.2 billion Anthropic has committed under a five-year lease for tensor-processing capacity, and the debt can convert into Anthropic shares. Broadcom is assembling a $60 billion financing package around it, with Blackstone leading an $18 billion junior tranche. We unpacked the deal mechanics in a separate piece on the vendor-financing structure.
Read that structure slowly. The chip vendor is lending its customer the money to rent the vendor’s own chips, and taking equity in the customer as part of the deal. Anthropic is on track to become Broadcom’s largest custom-silicon client by 2027, which is also why it is designing its own chips to pull inference costs down rather than rent its way out forever. And this is not a one-off. Nvidia spent the summer negotiating to backstop OpenAI’s build-out, floating a guarantee as large as $250 billion, trimming it below $120 billion, and settling near $105 billion when the partnership was signed. AMD struck its own equity-linked capacity deal with Anthropic. Nvidia had already disclosed well over half a trillion dollars of such arrangements across 2026.
The industry has a name for this now: circular financing. The vendor funds the customer, the customer buys the vendor’s product, the vendor books the revenue, and the customer’s valuation rises on the strength of the spending. Analysts have mapped more than $800 billion of these entangled commitments across the supply chain. The warnings are getting louder in step with the dollars: Michael Burry has publicly called it a circular web with Nvidia at the center, per 24/7 Wall St., and Nvidia’s five-year credit-default-swap spread has roughly doubled in two months as debt markets start pricing the risk.
The April version of this story was a company rationing access to protect margins. The October version is a company that cannot cover the bill from revenue and has instead borrowed the cost of the bill from the company that sends it. The compute crisis did not get fixed. It got refinanced, and the refinancing is now a systemic question rather than a quarterly one.
Cheaper tokens, bigger bills: the paradox held
The counterintuitive part I flagged in April is, if anything, more pronounced. Per-token prices keep collapsing. Blended token prices fell roughly 67 percent year over year, from about $18.40 to $6.07 per million tokens between the first quarters of 2025 and 2026, according to data McKinsey highlighted. Measured over three years, inference has fallen on the order of a thousandfold, from roughly $30 per million tokens at GPT-4’s launch to well under a dollar.
Enterprise AI bills went up anyway. Spend tripled to around $37 billion over the same window the prices were falling, because consumption outran the discount. Inference now accounts for roughly 85 percent of enterprise AI budgets, and the share of large organizations running agents in production climbed to about 40 percent from 27 percent a year earlier. An agentic workflow that chains dozens of model calls to finish one task consumes far more tokens than any chat ever did, so a 67 percent price cut gets swallowed whole by a volume increase several times larger. Falling prices are not relief here. They are the on-ramp to higher total spend, the same way cheaper bandwidth never once lowered a carrier’s data bill; it just moved more data. This is the compute-equals-output logic I wrote about in why every software company now runs on tokens, playing out on the cost side of the ledger.
What a financed subsidy means if you’re the one buying
Zoom out to the capital. Goldman Sachs now sees hyperscaler AI capital spending rising about 50 percent to $1.2 trillion, with the big five cloud and platform companies committing north of $600 billion in 2026 alone. Amazon is guiding to roughly $200 billion, Microsoft toward about $190 billion, Alphabet to $180 to $190 billion, Meta to $125 to $145 billion, and Oracle near $50 billion. That spend is what makes today’s token prices possible. Someone is paying to keep inference cheap, and increasingly that someone is a lender rather than a profitable operating business.
Running operations at a large telecom, I learned to distrust any input price a supplier was visibly subsidizing, because the subsidy is a decision someone can reverse, and you inherit the reversal. The AI version is sharper than the usual vendor-discount game, because the discount is now underwritten by convertible debt between a handful of mutually entangled companies. If you are putting AI into production, three things follow.
Budget for inference, not training: it is the 85 percent, it grows with your success rather than your experiments, and the fastest lever on it is architectural, not contractual (start with disciplined prompt caching). Keep at least two frontier providers genuinely integrated, not as a slide but as running code, because availability and pricing shift with each lab’s capacity and each financing deal, and single-provider dependence inherits someone else’s margin pressure and someone else’s lender. And when you model 2027 spend, assume the per-token discount you enjoy today carries an expiration date, because the entities funding it are carrying losses and debt to do so. The labs are not pricing to their costs. They are pricing to win share while other balance sheets absorb the gap.
The compute wars I described in April are still running. What changed is who is funding the fight. The companies winning on revenue have not learned to cover their own compute. They have learned to borrow it from the companies that make the compute, and that is a more interesting problem than throttled usage ever was.