SIGNAL / NOISE

By the Hour

Anthropic priced last night at $83 a share, above the $70 to $78 range, and sold 980 million shares when it had planned to sell 825 million. That's $81.3 billion, the biggest IPO ever, and about $1.5 trillion for a company that opens this morning as ANT. About $79 billion of it goes to the company.

Here's the number I'd look at before buying a share. Brad Gerstner put up a slide Tuesday: 26% of Anthropic's compute spending this year went to serving paying customers. The other 74% went to training, research, internal use and the free tier. That's Altimeter's estimate, and the arithmetic other people have run on the filing lands in the same place.

So who is the 74% for? Not the customers, as far as I can tell. Outside of coding, I haven't seen the job this year's model can do that last year's can't. Show me the contract. What I see is a preference, and a CFO can fix a preference. Companies pay for intelligence by the token, which is like paying your best employee by the hour. Sooner or later somebody in finance reads the bill and asks whether the cheap model would have done. Microsoft stood up Tuesday and said it for them: "Anything you do in Claude or ChatGPT, you can do in Copilot for less." OpenAI matched Anthropic's list price eight days after Fable 5.2 shipped.

(A tweet I can't find again made the other half of the point. When a scientist has to commit 45 days of compute to a question, he rations his questions. At 45 minutes they come hard and fast. If that was yours, step up and take the credit. So the bill may not fall. The price of each job will.)

That leaves the labs one way to pay for all this. A friend who runs money at a tech hedge fund puts next year's AI financing need at $545 billion, against about $200 billion this year. You don't cover that selling tokens against three rivals with the same product. You cover it the way Harvey Keitel does in Pulp Fiction. "I'm Winston Wolfe. I solve problems." Nobody asks the Wolf his hourly rate. Sell the finished job and you pick the model, so you keep the difference between the frontier price and the good-enough cost.

And once you sell the job, you're in the customer's business. First the labs solve problems for more and more companies. Then they can see, from the token flow, which of those businesses is ripest. Amazon ran the marketplace, watched what sold, and made the towels itself. Cabs lost to Uber, Uber is fighting Waymo, and Waymo is looking at a Cybercab built by the man who owns the car factory. Every margin in that chain ends up in one lap.

Jensen Huang told the Milken conference last year that you won't lose your job to AI, you'll lose it to someone who uses AI. I'd go one further. After that, the AI company takes his job too, by going after the company he works for. That's what the 74% buys, and as of this morning the public is paying for it.

Price the job before they do.

At COAI today: the full Signal/Noise, with the compute math from the filing, the pricing, and the list of businesses the labs have already walked into, is live at getcoai.com.

If you have kids and worry about how AI will affect them, turn what you read here every day into a weekly lesson planner. Subscribe to Office Hours, the newest newsletter from the CO/AI team.

ONE — A NUMBER THAT SUMMARIZES THE DAY

74%. That's the share of Anthropic's compute spending this year that did not go to serving a paying customer, by Altimeter's estimate. The customers already pay for the Claude they use. The 74% pays for the next one, and outside of coding nobody has shown me the customer who asked for it. That spending only earns its keep if the lab stops selling tokens and starts selling the finished job. Winston Wolfe never billed by the hour.

THREE — ACTIONS TO TAKE TODAY

For the investor — count the businesses, not the benchmarks. Before the open, write down what you think ANT is. If it's a token business with a fat gross margin, you own a product a rival matched on list price in eight days. If it's a company selling finished work in a growing list of industries, track the list. Claude Sheets and Whiteboards started rolling out Nov 10. The score is which incumbent lost revenue.

For the business owner — run last year's model on this week's work. Pull the tasks your team sent to the most expensive model and run them through the cheaper tier today. OpenAI says its second-tier model lands within half a point of its flagship on a coding test at about a fifth of the cost. If your work passes, bring that to the renewal. Then ask which of your jobs a vendor could sell back to you finished.

For the parent — teach them to own the problem. Ask your kid tonight who gets paid more, the person paid by the hour or the person paid to make the problem go away. The kid who learns to use AI is fine for a while. The one who can find the problem, own the outcome and stand behind it is a lot harder to route around.

FIVE — STORIES TO KEEP YOU INFORMED

Thursday, November 19

  • Anthropic raised the price and added stock. $83 a share, 980 million shares, $81.3 billion, past SpaceX's $75 billion. The seven founders sold $1.96 billion of their own. Orders ran near seven times supply. Last week I called raising the price and adding shares the Facebook 2012 move. Today we start finding out. (Full analysis above.)

  • OpenAI says Astra beats Fable at 5% of the cost. Read the footnote. The comparison drops nine tasks, and one developer says Astra spent seven days making his compiler 56% slower. Cheap per token and cheap per finished job are two different prices. (Full analysis above.)

  • Microsoft built a meter for the meter. Agent 365 Business Analytics ties token spend to work actually completed. Once a CFO can see what a million dollars of tokens bought, paying by the job is one contract away. (Full analysis above.)

  • A who's who called Anthropic unsafe on pricing day. Steven Pinker, Judea Pearl, Laurence Tribe and Nobel economist Oliver Hart signed a statement saying the company races to build the thing it warns about. It went out as a paid press release. Timing a letter to the print is a trade too.

  • Mubadala's CEO says winter is coming, and it'll be mild. Khaldoon Al Mubarak told Semafor's conference a slowdown in the buildout "doesn't mean minus 30 degrees." When one of the biggest checkbooks in AI starts describing the weather, I'd check what he's doing with the checkbook.

— Harry and Anthony

Sources:

Reply

Avatar

or to participate