Anthropic's Expensive Promise

On October 26 Anthropic made public a registration statement it had first handed to regulators in confidence on June 1.   The document names a ticker, ANT, on the Nasdaq.   It names Goldman Sachs and Morgan Stanley as the banks leading a large group of underwriters.   What it does not name is a price range or a share count, and one account says an amended version that went in on November 2 left them out as well.

A caveat before going further.   I have not read every page of the filing, and the figures below come from coverage of it.   As of November 6 the details are still moving.

The Ticker Is ANT

Start with what is reported.   The roadshow is said to begin in the week of November 9, though that timing comes from the press and not from the company.   First half revenue is reported at about $16.3 billion.   Press accounts have floated a valuation somewhere between $1.5 trillion and $2 trillion, and the filing does not say that.   Treat every one of those numbers as provisional until the next amendment lands.

Those are the numbers everyone is arguing about, and they are the least interesting part.   A filing like this describes a machine.   It tells you who controls it, what it costs to run, and who pays to use it.   The valuation is somebody’s guess about the output.

My claim is simple.   Anthropic is selling a promise.   Its governance is what makes the promise believable, its compute bills are what make the promise expensive, and its new status as a public company is what makes the promise measurable for the first time.   I will take those three pieces in order.

A Board With a Brake

Anthropic is a public benefit corporation, and it has a Long Term Benefit Trust that, according to reporting on the filing, appoints four of the seven board members.   The filing describes the company’s public benefit performance as measured against qualitative standards that the board sets.

In an ordinary growth company, whoever holds the votes decides, and the people holding them want growth.   Here a body that is not paid by growth chooses most of the board.   The point of the design is to settle in advance who is allowed to say no. If safety concerns can delay a launch, or a lucrative customer can be turned away, that is not a cultural preference.   It is a control right.

A promise is worth what it costs to break.   Anyone can say they will be careful.   A promise backed by a board the company cannot easily reshape is different, because the people making it have given up some of their freedom to walk it back.

There is a cost, and it is the cost of every brake.   A company that can choose the slower path on purpose will sometimes lose to one that cannot.   I cannot tell from coverage how the trust’s rights sit alongside those of ordinary shareholders, and I would read that section closely.   The open question is whether a brake stays a brake once public investors, who are buying for growth, are in the room.

What the Filing Forces

A private company tells its story the way that flatters it.   A public one has to tell it in a way that survives an audit, and that changes behavior.

Definitions come first.   When revenue, usage and retention are examined line by line, loose terms stop surviving.   What is an active customer?   What counts as a recurring dollar?   Those sound like accounting questions, but they also decide what the company pays attention to, because people manage whatever gets measured.   Whoever picks the definition is picking the incentive.

Then there is time.   Quarterly reporting drags attention toward whatever can be shown in ninety days.   Research, evaluation tooling and safety work do not produce a clean number each quarter, which makes them the first things squeezed when a quarter looks tight.   Anthropic has a structure meant to push the other way, and I would like to see how it holds when the first soft quarter arrives.

Finally, audience.   The filing is written for investors, but customers read it too.   An enterprise buyer deciding whether to put a model inside a regulated workflow will go through the risk factors with a pen in hand.   The document is a sales document forbidden from selling, an odd and perhaps useful constraint for a company whose pitch is trustworthiness.

Every Answer Has a Price

The reporting on the filing keeps returning to the cost of compute.   Training and serving frontier models takes enormous spending, and that spending sits in the middle of the unit economics.

This is where AI diverges from the software business most of us grew up with.   Traditional software has a gift in its cost structure.   Once the code is written, serving one more customer costs almost nothing, so the game is distribution and scale, and the winner keeps a huge margin on every additional user.

AI bends it.   Every request has a price.   When a customer asks a question, a bill is generated somewhere in a data center, and the answer has to be worth more than the bill.   How long is the default answer?   Which requests go to a large model and which to a small one?   Where do rate limits sit?   In software those are packaging details.   Here they are the business.

It also changes how to read the revenue figure.   About $16.3 billion for a half year says something about demand.   It does not say how much of that demand is profitable, and the filing is where the difference should show up.   I would also want to know who supplies the compute, because whoever sits upstream of a cost that large has leverage over everyone downstream.

What Customers Buy

Anthropic has built much of its identity on safety and careful behavior, and in a filing that identity becomes a set of statements the company is accountable for.   That points at what large customers are actually buying.

A hospital, a bank or a government agency does not first ask which model scores highest on a benchmark.   It asks what the system does when it is wrong, who is told, and how fast it can be switched off.   The answers come from evaluation suites, logging, approval steps and a named person responsible after launch.   When the raw capability of competing models is close, the buyer chooses whoever they believe will behave well under stress.

The three pieces fit together.   Governance is the reason the trust claim can be believed.   Compute is what it costs to deliver the claim at scale, since careful behavior means more checking and sometimes more refusals.   Disclosure turns the claim into something that can be compared against results.   Take away any one and the other two get weaker.

There is also a question of who owns the relationship.   If many customers reach the models through partners and cloud providers, the partner is the one the customer calls when something goes wrong, and the partner is the one building the trust.   Trust rented through someone else is worth less to the company that earned it.

The Next Amendment

I am not going to predict the pricing or the first day of trading.   Several other things will tell me more.

I would read the amended filing for how revenue splits between direct customers and partners, because that shows who owns the customer.   I would look at how it describes dependence on a handful of large customers and large cloud providers, since concentration decides who has leverage.   I would look at what it says about governance when growth and caution disagree.   And once the first earnings call arrives, I would listen for which metrics management chooses to talk about, because that choice is a strategy stated out loud.

The Promise and the Price

Stated plainly, Anthropic is a company whose product is a credible promise of careful behavior, and nearly everything in the filing either backs that promise or pays for it.   The board structure backs it.   The compute bill pays for it.   Public reporting will test it, quarter after quarter, in front of customers who read every word.

That is a harder business than the software companies of the last twenty years, which could count on near zero marginal costs and forgive a lot of mistakes.   It may also be a more durable one, if the promise holds.   I cannot tell you what the stock will do.   I can tell you that the way this company is built, from its board to its cost structure, will shape its next few model releases, and everyone building on those models has a stake in the result.

Sources

Reporting I drew on includes NPR, the SEC filing page, Commons, Tastytrade and BitMEX.   This is commentary and not investment advice, and the opinions are mine.