There's a certain poetry to watching a company sprint toward an IPO while its biggest expense line sprints faster. Anthropic, in a Monday update to its S-1 filing, gave the world a fresh look at the economics of compute - and the picture is less "scaling efficiently" and more "paying through the nose because everyone wants what we're selling."

The original S-1, released last week, only disclosed finances through June 30. The new filing extends the window to September 30, revealing a quarter that was, on the revenue side, genuinely impressive: $17 billion, up from $11.5 billion the quarter before. That's the kind of growth that makes investors salivate. Unfortunately, the operating loss grew right alongside it, largely because compute costs are doing their own impressive growth story.

Here's the number that should make any IPO prospectus reader wince: compute ate up 58% of revenue in the new quarter, up from 48% the quarter before. At the exact moment Anthropic could be wooing public-market investors with a story about improving margins, the margins are moving in the wrong direction. A chart in the filing tracks revenue per dollar spent on compute, which peaked at just over 2x this past quarter before sliding back to 1.72x. Since the start of 2025, that ratio had been climbing steadily - until it wasn't.

So what happened? Demand. Anthropic's compute bill hit $10 billion for the quarter, including what the filing describes as "$5.0 billion of compute under short-term commitments, which currently costs significantly more than comparable compute procured under our multi-year commitments." Translation: when you need compute right now and you haven't locked in long-term rates, you pay the premium tier. Anthropic has been paying SpaceX $1.25 billion a month for compute, alongside multibillion-dollar deals with Nscale, Akamai, and Volta - the industry's "neoclouds," which offer flexible deals at higher rates than the hyperscalers.

The company is trying to rein this in. It's building out its own compute base under a project called "Anthropic Direct," but when demand is growing at nearly 50% a quarter, internal efforts can't keep pace. The neoclouds remain the best option available, and the filing is candid about the trajectory: "We expect compute and infrastructure expenses to increase in absolute dollars and remain our largest operating expense for the foreseeable future as additional committed compute becomes available to us under multi-year arrangements with CSPs, and as we continue to invest in additional compute."

In fairness, these are champagne problems. Revenue growing at nearly 50% a quarter tends to quiet investor concerns about high compute costs, and if demand ever slows, the compute picture would presumably get sunnier fast. But the filing is a reminder of how thoroughly AI labs depend on the trillion-dollar infrastructure buildout - and how exposed they are to its hiccups. If bond markets get skittish or local opposition to data centers intensifies over the next year, Anthropic can't do much about either risk. And either one would turn a messy compute picture into a considerably messier one.

For now, the future of Anthropic and its peers remains at the mercy of compute prices - which is a polite way of saying the AI revolution runs on electricity, real estate, and the willingness of bond traders to keep the party going.