Anthropic prices its IPO at $70–78 a share as "ANT". Its ">4x compute efficiency" is revenue per compute dollar, and it fell last quarter

Anthropic has put a price on itself. The second amendment to its registration statement, the S-1/A filed with the SEC on November 10, sets an expected range of $70.00 to $78.00 a share for 825,000,000 shares of Class A common stock, to be listed on the Nasdaq Global Select Market under the ticker ANT. The roadshow began on November 10; according to Anthropic's IPO site, the price is expected to be set on November 18 and trading to begin on November 19 (both US Eastern time). The company announced the offering on X on November 12.

We have read this filing twice before: the first S-1 in October, and the November 2 amendment that filled in the size of its compute commitments. The third version adds the deal terms, a slightly bigger compute bill to come, a new vocabulary, and one marketing number worth taking apart.

The deal

Of the 825 million shares, 801,351,354 are new shares sold by Anthropic and 23,648,646 are sold by existing stockholders; the underwriters, led by Goldman Sachs and Morgan Stanley, can buy up to 123,750,000 more. Alongside the IPO, seven companies have agreed to buy $26.0 billion of non-voting Class N stock at the IPO price: AMD, Broadcom, Google (as part of investment it had already committed), Micron, Microsoft, NVIDIA and TSMC Partners. That is most of the hardware and cloud stack Anthropic runs on, buying shares in its customer. At the $74 midpoint, Anthropic estimates net proceeds from the offering and the private placement at about $84.6 billion.

Control stays where it was designed to stay. A single share of Class F stock, held by an LLC whose members are Dario Amodei, Daniela Amodei and the five other co-founders, carries 50.1% of the votes on ordinary stockholder matters. The Long-Term Benefit Trust, through Class T stock, elects a majority of the board. Employees get a Class E share class that votes only on limited matters. Public buyers of Class A get one vote per share in a company that calls itself a "controlled company" under Nasdaq rules.

"Compute efficiency" is revenue per compute dollar

The IPO site's list of third-quarter highlights includes a ">4x compute efficiency gain since Q1'25". The footnote defines it: revenue divided by compute and infrastructure expenses in Q3 2026, divided by the same ratio in Q1 2025. It is not a measure of FLOPs per token, or of how cheaply a model is served. It is a revenue ratio, and the denominator includes all of Anthropic's compute, training and research included.

The prospectus gives the quarterly figures, so the ratio can be recomputed for every quarter (the arithmetic is ours):

The ">4x" checks out: 4.3 times. But the series peaked a quarter earlier, at 5.2 times, and fell by 17% in the third quarter. The filing says so itself, in the MD&A: "underlying compute efficiency improvements were partially offset by the delivery of additional compute, including $5.0 billion of compute under short-term commitments, which currently costs significantly more than comparable compute procured under our multi-year commitments. This reduced revenue per dollar of compute compared to the preceding quarter." The highlight picks the base quarter that makes the number look like an uninterrupted climb.

Because the measure is revenue-based, it moves with things that have nothing to do with efficient hardware use: pricing, the mix of customers, and how much compute goes to research rather than to serving. The S-1 defines compute efficiency as "maximizing the intelligence produced and revenue generated per dollar of compute invested" and names three drivers: training efficiency, inference efficiency, and running Claude on four chip families (AWS Trainium, Google TPUs, NVIDIA GPUs and AMD GPUs) through its own orchestration layer. Only the revenue half of that definition comes with a number. As our earlier reading showed, at most about a quarter of the compute bill goes to serving paying customers; the rest sits in research and development, which is why a big training push can drag this ratio down while doing nothing to make inference less efficient.

$5.5 billion more in eight days

The compute commitments, signed as of and after September 30, now total approximately $900.9 billion, up from $895.4 billion in the November 2 version. The part due in the 12 months after September 30 rose from $69.8 billion to $70.7 billion, against third-quarter compute expenses of $9.9 billion. The company also says it intends to "secure further compute beyond our existing commitments" and expand its own data-center footprint.

AI is now "super intelligence", by definition

The most visible edit is lexical. The November 10 prospectus adds a glossary entry: "'super intelligence' or 'SI' refers to a machine-based system that learns or is trained from data and, for explicit or implicit objectives and with varying levels of autonomy, processes that data to infer patterns, generate outputs—such as predictions, content, recommendations, and decisions—and act upon physical or virtual environments." The entry for "artificial intelligence" in the same glossary uses exactly the same words. The filing explains why: Executive Order 14434, issued on September 29, 2026, "directs federal executive departments and agencies to use the terms 'Super Intelligence' and 'SI' in place of 'Artificial Intelligence' and 'AI'", and Anthropic says it uses the terms interchangeably "to avoid confusion".

The effect on the text is measurable. In the November 2 version, "super intelligence" appeared once, inside the name of a government body; in the November 10 version it appears 23 times, and "SI model" 14 times. Anthropic also leans on a term of its own: "frontier intelligence" went from 2 mentions to 59, while "AI model" fell from 103 to 74. Nothing about the technology changed between the two filings. A chatbot that summarizes a PDF is, for the purposes of this prospectus, super intelligence.

What investors are told about Claude's limits

One risk factor will interest Claude subscribers more than investors. "We have also changed, and may in the future change, usage limits, rate limits, or model and product capabilities," it says, adding that "some of these changes have resulted, and could in the future result, in claims that our disclosures regarding model access or capabilities were inadequate or misleading." That sentence is not new: it has been in every public version since October 26. Subscriptions brought in $6.9 billion of the $33.2 billion in revenue for the nine months to September 30, about 21%; the rest is billed by tokens consumed.