Anthropic's IPO Math: $65B Run Rate, $42B Loss • The Weighted Average

Anthropic expects its initial public offering to match or exceed SpaceX’s record raise, and it is bringing an unusual pair of numbers to the pitch. Its annualized revenue run rate hit $65 billion at the end of July, CNBC confirmed with three sources familiar with the investor update — a sevenfold jump in a year, against roughly $10 billion of full-year 2025 revenue. In the same window, reporting on the IPO preparation puts the company’s 2025 net loss at almost $42 billion, about five times the prior year, per The Next Web’s summary of the Bloomberg reporting on the filing plans.

Put those side by side and you get the figure the roadshow will spend its time on: 65 cents of loss for every dollar of annualized revenue, comparing a 2025 loss to a mid-2026 run rate. The mismatch of periods is the point. The bull case is that the loss belongs to a training era the revenue has already outrun; the bear case is that the loss line scales with the revenue line because both are functions of compute. Nothing in the public record settles it, which is precisely why the offering is being pushed while momentum is legible.

A run rate is not revenue, and the market knows it

The $65 billion figure deserves the caveat its own coverage attaches. It came from unnamed sources rather than a filing, Anthropic declined to comment, and as The Next Web noted, a run rate extrapolates a short period into a year — a metric shared with investors as part of a pitch. The audited number underneath it is preliminary second-quarter revenue of more than $11.5 billion, against $787 million in the same quarter of 2025. Annualize $11.5 billion naively and you get $46 billion, not $65 billion; the gap is the growth the company expects to keep compounding through the second half.

There is a second reason to distrust smooth extrapolation. Anthropic’s revenue line has already survived events that had nothing to do with demand: in June the company temporarily disabled access to Claude Fable 5 and Mythos 5 to comply with a government export directive, restoring them after roughly two weeks of negotiation, per the same CNBC reporting. A supplier whose flagship models can be switched off by policy carries a risk that no run-rate chart displays. We covered the shape of that exposure when Anthropic first shipped Fable and Mythos as a guarded release.

The comparison set matters for anyone modeling their own vendor costs. OpenAI’s annualized run rate recently reached $40 billion by CNBC’s account, meaning Anthropic now claims a 62% lead on its closest rival while sitting on a $965 billion private valuation from its May round. Enterprise buyers should read that as pricing power arriving before the lockup expires, not after. We have tracked this trajectory since Anthropic’s $900 billion valuation topped OpenAI’s, and each step has been financed by capital that expects returns from the same enterprise budgets doing the buying.

The debt stack behind the equity story

The equity raise is the visible half. The other half is leverage. Broadcom is negotiating a financing package with private lenders that includes a senior secured tranche of $60 billion to $70 billion plus a junior tranche of roughly $30 billion, potentially totaling $100 billion, according to Stocktwits’ account of the Bloomberg report. The proceeds secure custom chips and infrastructure for firms including Anthropic; the structure runs through a special-purpose vehicle with Broadcom backstopping part of the senior debt, extending an earlier $35 billion vehicle whose investors funded chips leased directly to Anthropic.

This is the same mechanism Nvidia has been industrializing, and the reason its moat is shifting from chips to capital. A guarantee from an investment-grade supplier turns speculative AI capacity into paper a pension fund can hold. It also means a model vendor’s cost base is now partly a credit product, with covenants and refinancing dates that have nothing to do with token demand.

Note what that does to the cost curve an operator is implicitly renting. If chips arrive through leveraged vehicles rather than retained earnings, the interest expense sits somewhere in the stack, and the only place it can ultimately land is token pricing or equity dilution. Investment-grade guarantees keep the rate low today; they do not remove the obligation.

For operators the practical consequence is contractual, not philosophical. A supplier heading into public markets with a loss of this size has three levers: raise prices, cut discounting, or push customers toward higher-margin tiers. All three show up first in renewal terms. Teams that negotiated multi-year Claude commitments during the private era should re-read their price-protection clauses now, before an S-1 makes gross margin a quarterly obligation. The same logic argues for keeping a live second supplier, which is exactly why today’s lead on Harvey’s decision to post-train its own model reads as a hedge rather than a vanity project.

What would change the verdict: the public prospectus. Confidential filings hide the two numbers that decide this — gross margin by product line, and how much of the run rate is committed contract versus usage that can evaporate. If the S-1 shows enterprise commitments carrying the majority of revenue, the loss is an investment. If it shows consumption-based usage funding a training bill that grows with it, the 0.65× ratio is a structural feature, and every buyer is helping finance the next run.