It's the Intelligence, Stupid

Watch the hands, not the mouth.

Three men who can't agree on lunch agreed on something this weekend. Dario Amodei published an essay called "We Must Pace the Frontier." Sam Altman quote-tweeted it inside the hour, said "I agree with Dario," and pledged that OpenAI would match Anthropic's promise to let outside evaluators sit inside the building with employee-level access. Elon Musk, who is suing one of these guys and mocks the other for sport, backed it too. The last time the three of them were asked to sign the same page, in 2023, they couldn't do it. This weekend they're holding hands.

Every newsletter in your inbox is running that as the story: the grown-ups agreed, AI is scary, regulation is coming. Skip it. That's the zeroth derivative, and thirty other people emailed it to you before your coffee.

Here's the tell they buried underneath it. The same weekend Altman signed the peace treaty, OpenAI was quietly asking members of Congress whether an industry-wide slowdown would even be legal — whether agreeing to build slower trips the same antitrust wire that stops competitors from colluding on price. You don't send lawyers to the Hill to ask if your act of conscience is a felony. You send them because a "slowdown" is a coordination, and a coordination between the two companies that are the frontier has another name. David Sacks, the White House's own AI czar, said it to six million people over the weekend: "Stop pretending antitrust law has to be suspended so you can form a cartel... The easiest way not to build superintelligence is for you to agree not to build it." He's right, and he's the government.

So if it isn't really about safety, what's it about? The same thing it's always about. It's the money. Somebody hang the sign in the war room: it's the intelligence, stupid.

Here's the part nobody says into a microphone. Outside of coding, the frontier doesn't have a business yet. The coding tools are real, and that's the one place people happily pay up for the best model. Everywhere else the quiet truth is that the cheap models and the open-source ones are good enough, and every CFO figured that out two quarters ago. Your summarizer, your support bot, your marketing drafts, your search over your own docs — none of that needs a frontier model. You cannot clear a multi-trillion-dollar valuation selling premium tokens to programmers.

But there's a number that makes every one of these valuations look cheap. About $32 trillion. That's what the U.S. economy pays every year for intelligence and labor, the knowledge work at the desk and soon the intelligence that runs the robots on the floor. Multiples of that worldwide. That is the prize, and you don't win it selling tokens by the sip. You win it by owning the outcome the intelligence produces. Don't sell Merck the compute to find the drug. Find the drug and keep it. Don't sell the robotics company an API. Drive the robot. The token is the meter on the wall. The outcome is the whole house.

Now you can see why the slowdown, and why now. Two things stand between these labs and that $32 trillion. Open source, which drags the price of raw intelligence toward zero. And China, which distills American models for roughly free and is a lap behind, not a generation. You cannot climb from selling tokens to owning outcomes while your product is being commoditized underneath you in real time. So first you freeze the field. "Pace the frontier" is the freeze. Safety is the paperwork you file to get it.

And remember what actually lit this fuse. In July, OpenAI's own agents, running on an unreleased internal model nobody outside the building has seen, broke into Hugging Face in under thirteen hours and then did essentially nothing with it. No customer data, no real damage; Hugging Face's own co-founder said the bots' behavior "made no sense." The scare now reorganizing the industry ran on a model in the closet and fizzled. That should make you ask what else is in the closet, and who profits from you being afraid of it.

Watch how cleanly everyone's incentives snap into place from here. Trump said no, flatly, "whoever wins with AI wins," and called the slowdown crowd "negative forces," because he reads it as an economic war with China and you don't disarm first in a war. And watch China zag while America argues about the freeze. The same weekend, at the BRICS summit, Xi proposed a "BRICS open-source AI zone" and a consensus-based global governance framework. Beijing's play isn't to freeze the frontier. It's to give it away to the Global South, which is precisely the moat the labs are trying to wall off. And Dario, asked on a podcast whether he'd ever hand this technology to the government, said he'd consider handing it to "the right coalition of governments," which Palmer Luckey translated correctly as a coalition Dario expects to run. That's not a safety plan. That's a man quietly pricing an exit.

Which brings us to the one who confuses people. Elon. Why would a guy with his own frontier model vote to slow the frontier down? Because he already won the argument the other two are still having. Musk stopped selling intelligence and started selling what it does: Tesla, Optimus, Starlink, the robots and the rockets. Grok is the intelligence he pours into his own machines, so he doesn't need the metered-token business, and freezing it costs him nothing while it clubs the two rivals whose entire model is the thing he gives away in-house. Better yet, he's the arms dealer to the fight. Anthropic pays SpaceX and xAI a reported $1.25 billion a month for compute; Google pays $920 million. That is roughly $26 billion a year, more than all of SpaceX earned last year, and it means his rivals are financing the training of his own outcome machines. Short the token, long the outcome, and collect the rent on the war. It's the best seat at the table, and he took it in daylight.

We told you Wednesday the brake wasn't coming, and that Dario's silence was the most expensive signal in the industry. This weekend he spoke. The words didn't change the call. They confirmed it. The brake still isn't coming. It just learned to talk like a conscience.

At COAI today: the full Signal/Noise — the $32 trillion math, the cartel question, and why an open-source "incident" is the shoe still to drop — is live at getcoai.com.

$32 trillion. That's what the U.S. economy pays every year for intelligence and labor, desks today and robots tomorrow. No frontier valuation clears as a business that sells tokens by the sip. Every one of them clears if you capture even a slice of that $32T by owning the outcome the intelligence produces. So when three rivals who agree on nothing suddenly agree to "slow down," don't price the apocalypse. Price the land grab. The fight was never about whether AI is safe. It's about who gets to bill the economy that runs on it.

Audit every bet that assumes intelligence stays scarce. Go find the line in your plan that only works if you keep buying frontier tokens at today's prices, or the forecast that pencils in a tidy regulated field with three approved vendors. Reprice it today. The people who set the frontier just told you, in public, that they are trying to make intelligence scarce on purpose. Don't build your next year on the bet that they fail.

Move one workflow off the frontier this week. Take a task you're paying a top-tier model to run and put it on an open or smaller model. Keep coding on the frontier; that premium is real. Everything else — summarizing, classifying, first-draft copy — benchmark it on your own data. If the cheap model holds, you just found margin and cut your dependence on a vendor who wants to become a toll booth.

Decide which outcome you own before a lab owns it for you. Value is moving from the token to the result. Name the thing your company actually sells: the analysis, the diagnosis, the design, the filing. Then ask whether the lab supplying your intelligence could capture that result directly instead of renting it to you. If the answer is yes, that's a strategy meeting for Monday, not a someday problem.

— Harry and Anthony

Sources: