Anthropic's S-1 confirms the price collapse, in its own words

category

Perspectives

date

11/17/2026

author
No items found.

This is a follow-up to "The AI price collapse has only just begun"‍

A few weeks ago, I argued that token prices are falling off a cliff, open-weight models are good enough for most real work, and any business whose margin is a toll on token throughput has an expiry date.

‍

Then Anthropic filed its S-1. I expected the pushback to come from the frontier labs. Instead, one of them confirmed the thesis under oath, with page numbers.

The lab says the premium is at risk

The risk factors warn that customers "may conclude that incremental performance gains at the frontier do not justify the additional cost" and may shift to "smaller, cheaper, or open-source alternatives." Open-weight models can "exert pricing pressure." And competition "could increasingly depend on access to large-scale compute rather than on model development capability."

‍

That last line is the whole story. A frontier lab is telling the SEC that building the best model may soon matter less than owning the most data centers. Intel learned this lesson the hard way: once rivals could buy the same fabs and tooling, chip design alone stopped guaranteeing margin. The moat moved from the lab to the balance sheet.

Consumption has no floor

Eighty percent of Anthropic's revenue is metered per token. Revenue rises with usage and falls with it just as faithfully. The S-1 calls this "pay-for-intelligence." Your electric utility calls it a meter.

‍

And the price per unit is dropping on Anthropic's own say-so. Opus 5.5 "reduced customer costs per task by approximately 40%." Fable 5.1 cut typical workloads another 25%, up to 45% for agentic tasks. Two cuts of roughly 40% in a year, with more promised. Anthropic frames this as a growth engine: after Opus 4.5 shipped, quarterly revenue growth jumped from 48% to 132%. Tomasz Tunguz called it Jevons paradox in the real world.

‍

Jevons works for the coal mine. It's less kind to the coal reseller. When Anthropic cuts price per task, it collects less per token and makes it up on volume. The company reselling those tokens with a nice interface on top just collects less, and has to hope the volume lands at its door rather than a competitor's.

No switching costs, confirmed from the inside

Tunguz leads his bear case with "consumption-based with no lock-in." Mostly Metrics marveled at 900% net revenue retention sitting next to "a commodity product with little to no switching costs." Both are true. Customers spend nine times more than a year ago because the models got better and cheaper, not because they're stuck.

‍

It's worth noting that 48% of first-half revenue came through AWS, Google Cloud, and Azure marketplaces. Those same clouds host Qwen, DeepSeek, and Llama on the next shelf. The filing names this as a conflict-of-interest risk. For an enterprise buyer, switching models is a dropdown menu on the same invoice. That's the grocery aisle, not the App Store.

Compute is the moat, and it costs $637 billion

Analysts summed the named compute deals with Google, Amazon, Microsoft, Broadcom, SpaceXAI, Nscale, and Lambda to roughly $637 billion. About two-thirds is non-cancelable, with true-up clauses that bill Anthropic for capacity it doesn't use. The S-1 leaves the total blank. First-half revenue was $16.3 billion.

‍

Forbes put it well: if the research edge narrows, the balance sheet becomes the moat, and the balance sheet is carrying the compute bill. Anthropic has $131.7 billion in cash to back it, so this isn't reckless. But it's the economics of a railroad or a telecom, not a social network. Railroads were magnificent businesses that reshaped the world. They also never traded like Facebook, because laying track is something a well-funded rival can always do.

What this means for Builders

If you resell tokens, the clock runs faster than you thought. Your supplier just cut your input cost 40% twice and told investors to expect more. Your margin on the AI layer shrinks on a schedule you don't control. The question stands: what part of your product isn't a commodity?
‍

The labs are climbing the stack toward you. The S-1 positions Claude Code and Cowork, not raw API access, as the future, and announces Ode with Anthropic, a services arm built with private equity firms to deploy Claude into mid-sized companies. Microsoft did this to the independent software vendors of the 1990s, one bundled feature at a time. If your edge is a friendlier wrapper on Claude, your supplier is now your competitor.
‍

Multiples are about to get a public-market reality check. Reports put the IPO near $2 trillion. Public investors will decide whether Anthropic trades like a software platform or like a capital-intensive utility selling a deflating commodity at enormous scale. Whatever they decide becomes the comp for every AI company behind it. If you're raising in 2027, expect the question: what does this business look like when compute is 90% cheaper?
‍

Cheap intelligence is still the opportunity of a generation. Bandwidth getting cheap didn't kill the internet, it killed the toll booths and made Netflix possible. Anthropic's own numbers show usage exploding as cost per task falls. The winners will capture that usage with something a cheaper model can't copy: proprietary data, deep workflow integration, or outcomes customers pay for regardless of what tokens cost.
‍

The company with the most to gain from expensive intelligence just told the world, in a legal filing, that intelligence is getting cheap. Plan accordingly.

‍

‍

Sources: Anthropic Form S-1 (SEC EDGAR, filed October 26, 2026) · Forbes — The First IPO Where Investors Should Read the Risk Factors First · Tomasz Tunguz — Anthropic S-1 Analysis · Agentic Bytes — What the IPO Filing Says About AI's Economics · Mostly Metrics — Anthropic IPO S-1 Breakdown · CMS Investments — Anthropic's S-1: The Numbers That Matter · CNBC — Anthropic IPO live updates

Related insights