Anthropic IPO: Investors Want Revenue-Per-Token Data

As Anthropic moves toward one of the most anticipated technology IPOs in years, its would-be public investors are asking for something the company has never had to disclose: the unit economics of running a frontier AI lab. According to reporting circulating this week, potential buyers are pressing Anthropic for granular operating metrics — including revenue per token and revenue per gigawatt of compute — that go well beyond the headline growth figures the company has released so far.

The demand is a sign of how AI valuations are maturing. Revenue growth alone impressed the private markets that funded Anthropic’s rise; the public markets want to know whether that revenue is efficient.

The headline numbers everyone already has

Anthropic’s top-line trajectory is extraordinary and well documented. The company disclosed an annualized revenue run rate of roughly $65 billion at the end of July 2026, up from about $9 billion at the end of 2025 — a pace of growth with almost no precedent at this scale. Preliminary second-quarter revenue came in near $11.5 billion, up sharply from $4.73 billion in the first quarter.

Those figures sit behind a private valuation of roughly $965 billion, set by the $65 billion Series H the company closed earlier this year, and Anthropic filed its draft registration confidentially with the SEC in June. A public filing has been anticipated for late September 2026, with Morgan Stanley, Goldman Sachs, and JPMorgan reported to be working on the offering.

What none of that reveals is how much it costs Anthropic to produce a dollar of that revenue — and that is exactly the gap institutional investors are now trying to close before they commit capital at a valuation that could approach or exceed a trillion dollars.

Why per-token and per-gigawatt metrics matter

For a conventional software company, investors lean on gross margin, net revenue retention, and customer acquisition cost. For a frontier AI lab, those metrics are incomplete, because the dominant cost is not sales headcount — it is compute. Every query served and every model trained consumes expensive GPUs and scarce high-bandwidth memory, and the economics of the business live or die on how efficiently that compute is converted into revenue.

Two metrics capture that directly:

• Revenue per token measures how much money the company earns for each unit of model output it generates — a proxy for pricing power and inference efficiency combined.
• Revenue per gigawatt of compute measures how much revenue the company extracts from each unit of power-limited infrastructure — a proxy for how productively its data-center capacity is deployed.

In an industry where data-center economics increasingly determine who wins, and where power itself has become the binding constraint, revenue per gigawatt is arguably the single most revealing number a frontier lab can disclose. It is the AI-era equivalent of same-store sales: a measure of whether the business is getting more productive as it grows, or simply bigger.

What the numbers reportedly show

The early comparisons that have leaked are flattering to Anthropic — with an important caveat.

On revenue per gigawatt, Anthropic is reported to generate roughly $21.4 billion per gigawatt of compute, versus about $12.6 billion per gigawatt for OpenAI — implying Anthropic squeezes something like 70% more revenue out of each unit of compute than its larger rival. On its face, that is a striking efficiency advantage, and it fits Anthropic’s positioning as the disciplined operator focused on high-value enterprise workloads rather than sprawling consumer scale.

The caveat is in the denominator. Anthropic’s revenue per gigawatt is high in part because its total compute footprint is smaller — the company is running closer to the edge of its own capacity than OpenAI is. A high ratio driven by a constrained denominator is a double-edged signal: it demonstrates operational discipline, but it also flags that growth may be capacity-limited, and that Anthropic will have to keep signing enormous infrastructure commitments — the kind that show up as long-term obligations in a prospectus — to keep the numerator climbing.

That tension is the whole ballgame for an IPO buyer. Efficient today does not guarantee efficient at three times the size, especially if the next tranche of capacity comes online at a worse cost basis amid the ongoing memory supercycle squeezing every hardware buyer.

The transparency test

There is a reason investors are having to ask for these numbers: no frontier lab has yet made them standard disclosure, and whatever Anthropic chooses to publish will set a template. If Anthropic breaks out revenue per token and per gigawatt in its S-1, it effectively forces OpenAI and every other lab eyeing the public markets to answer the same questions — or explain why they won’t.

That dynamic cuts both ways. Detailed unit economics let Anthropic tell a differentiated efficiency story that its private-market run-rate growth alone cannot. But the same disclosures hand competitors a precise benchmark and give skeptics a clean line to attack if the ratios deteriorate. The company’s reported $30 trillion addressable-market framing has already drawn scrutiny for optimism; hard operating metrics are harder to spin.

What it means

The AI IPO era is entering its accountability phase. Private investors bought growth; public investors are demanding proof that growth is profitable, or on a credible path to it. The move from “how fast is revenue growing” to “how efficiently is compute being monetized” is the market maturing in real time — and Anthropic, as the first pure-play frontier lab likely to test the public markets, is where that shift gets priced.

Anthropic’s efficiency edge is real but fragile. A 70% advantage in revenue per gigawatt is a genuine selling point, and it reinforces the narrative of a lab that runs lean. But investors will read the constrained-capacity subtext just as clearly, and the durability of the advantage depends on whether Anthropic can add compute without eroding its returns per unit.

Watch what actually lands in the S-1. The most important disclosure decision ahead of this listing may not be the valuation or the share count — it is whether Anthropic publishes the unit economics investors are asking for. If it does, it raises the bar for the entire sector and gives the market its first real yardstick for comparing frontier labs. If it demurs, that silence will itself become the story, and every rival preparing to follow Anthropic to the public markets will be watching how the market punishes or rewards it.