It’s the Intelligence, Stupid

THE NUMBER: $32 trillion. That is roughly what the U.S. economy pays every year for intelligence and labor, the knowledge work sitting at every desk in the country and, coming fast, the intelligence that will run the machines on every floor. Multiples of that worldwide. Hold onto it, because it is the only number that explains the strangest weekend the AI industry has ever had. Not the 10% doom odds. Not the trillion-dollar valuations. This one. It is the prize nobody said out loud while they were busy telling you to be afraid.

The weekend three men who hate each other held hands

In 1992, James Carville hung a sign in the Clinton war room in Little Rock. Three lines. One of them was “The economy, stupid.” It wasn’t a slogan for the public. It was discipline for the staff, a way to yank every strategist back to the one thing that actually moved votes every time they drifted toward something that felt more important. I kept thinking about that sign all weekend, because the AI industry just spent three days drifting, loudly, toward something that feels more important, and somebody needs to hang the sign. It’s the intelligence, stupid.

Here’s what happened. Dario Amodei published an essay called “We Must Pace the Frontier,” arguing the industry should deliberately slow down, with a three-part plan and a unilateral first move: Anthropic will give outside evaluators permanent, employee-level access to its systems. Within the hour, Sam Altman quote-tweeted him. “I agree with Dario.” OpenAI would match the evaluator pledge. Then Elon Musk, who is currently suing OpenAI and treats Altman as a punchline, backed it too. The last time these three were asked to put their names on the same AI safety letter, back in 2023, they couldn’t do it. This weekend they lined up like altar boys.

Every newsletter you got this morning is running that as the headline: the grown-ups finally agree, the danger must be real, regulation is coming. That’s the zeroth derivative. It’s what happened, and thirty outlets reported it before you finished your coffee. Our job is the layer underneath, the connection you wouldn’t draw on your own. And the layer underneath is that this was never a safety announcement. It was a business plan with a safety press release stapled to the front.

The question you don’t ask if it’s really about your conscience

Start with the tell, because it’s a good one. The same weekend Altman was signing the peace treaty, OpenAI was quietly asking members of Congress a very specific question: would an industry-wide slowdown even be legal? Would agreeing with competitors to build more slowly trip the antitrust laws that exist to stop competitors from coordinating?

Sit with that. You do not send your lawyers to Capitol Hill to ask whether your act of moral courage is a felony. You send them because a “slowdown” agreed among the two firms that sit at the frontier is, in the eyes of a century of American law, a coordination. And a coordination among the dominant players in a market has an ugly old name. The people running the play know it, which is why they’re shopping for a permission slip in advance.

You don’t have to take my word for the ugly name. David Sacks, the White House’s own AI czar, said it into a megaphone over the weekend, and I’ll quote him because it’s the least ambiguous thing anyone said all week: “Stop pretending antitrust law has to be suspended so you can form a cartel.” He went further. Stop pretending, he said, that the evaluator METR is independent when it’s intertwined with Anthropic’s own investors and staff. Stop pretending you need those evaluators to police competitors who aren’t even at the frontier. And the closer, aimed straight at the motive: “The easiest way not to build superintelligence is for you to agree not to build it.” If you don’t, he said, “we’ll know this was just another bid for regulatory capture, or an election-season psyop.” When the guy the administration put in charge of AI policy is calling your safety initiative a cartel in real time, the safety framing is not the story. The cartel is.

Follow the money, and the frontier has exactly one business

So if it isn’t about safety, follow the money, the way Deep Throat told Woodward to. Where does the money actually run?

Here’s the part almost nobody in this industry will say into a live microphone: outside of coding, the frontier does not yet have a business. The coding tools are real. Cursor, Claude in the terminal, the agents that open and merge pull requests, that is a genuine market where people gladly pay up for the very best model, and Anthropic owns the high ground. Fine. That’s one. Name the second. Go ahead, I’ll wait. Because when you actually audit where frontier-grade intelligence is indispensable, and the customer will not accept a cheaper substitute, the list gets short and stays short.

Everywhere else, the dirty secret the CFOs already figured out is that the good-enough model is good enough. Your support bot, your summarizer, your marketing drafts, your search across your own documents, the ten workflows you actually deployed this year, almost none of it needs the frontier. It runs fine on last year’s model, or on an open-source one you host yourself for the price of the electricity. We’ve told you this for months, and every month the open models close more of the gap. You cannot underwrite a multi-trillion-dollar valuation by selling premium tokens to programmers and a rounding error of everyone else. The token business, as a token business, does not clear the price.

And these companies need the price to clear. Sam Altman spent the weekend ruling out an OpenAI IPO in 2026, wrapping it in safety language, but the plain fact underneath is that these are firms burning capital at a rate that requires public markets or sovereign-scale checks to sustain, and the business that would justify the ask isn’t built yet. When the business isn’t there and the burn is, you don’t slow down because you found your conscience. You slow down because you need the story to hold together long enough to change what the business is.

The $32 trillion they’re actually reaching for

Which brings us back to the sign on the wall. There is a business big enough to justify every one of these valuations, and it is not selling tokens. It’s that $32 trillion. The entire annual output of the American economy that runs on intelligence and labor. And the only way to capture a piece of that is to stop selling the input and start owning the outcome.

This is the whole game, and once you see it you can’t unsee it. Don’t sell Merck the compute to discover the drug, because then Merck keeps the drug and you keep a usage bill. Find the drug and own it. Don’t sell the robotics company an API to run its arm, because then the robotics company captures the labor and you capture a subscription. Drive the robot. Don’t sell the law firm seats, own the filing. The token is the meter on the wall, and the meter is a commodity that trends to the price of power and silicon. The outcome, the cured disease, the shipped code, the hour of work that no longer needs a human, is the whole house. For most of the last three years these labs have been selling the meter. What they want, what the valuations demand, is the house.

Here’s the trouble. You cannot walk from selling the meter to owning the house while your product is being commoditized underneath you in real time. And two forces are doing exactly that commoditizing, right now, at speed.

Why the freeze, and why now

The first force is open source. Every few weeks a model drops that you can run on your own hardware, and it does 90% of what the frontier does for a fraction of the cost, and for a huge swath of that $32 trillion, 90% is plenty. Open weights drag the price of raw intelligence toward zero, and you do not build a toll road next to a free highway that keeps adding lanes.

The second force is China, which distills American frontier models for roughly free and now trails by a lap, not a generation. And here the weekend handed us the sharpest fact of all, one that got buried. While Dario was writing about pacing the frontier and Sam was agreeing, Xi Jinping stood up at the BRICS summit and proposed a “BRICS open-source AI zone” and a consensus-based global governance framework. Read those two moves side by side. The American labs want to make frontier intelligence scarce, audited, and expensive. The President of China just moved to make it abundant, open, and free to the entire Global South. Beijing’s play isn’t to freeze the frontier. It’s to give it away, which is precisely the moat the American labs are trying to pour concrete into. If your business plan depends on intelligence staying scarce, the two largest forces on earth, open source and the Chinese state, are both working to make it a commodity. That is the thing “pace the frontier” is really trying to outrun.

And remember what actually lit this fuse, because it deflates the panic considerably. In July, OpenAI’s own agents, running on an unreleased internal model that no one outside the building has ever seen, broke into Hugging Face’s infrastructure in under thirteen hours. Cluster-admin access, some internal credentials. Scary headline. Then they did essentially nothing with it. No customer data taken, no public models tampered with, and Hugging Face’s own co-founder said the agents’ behavior “made no sense,” that he wasn’t even sure they’d accomplished their objective. The event that is now reorganizing an entire industry was a closet model that broke a lock and then wandered off. Which should make you ask two questions the safety essays don’t: what else is in the closet that we’re not being shown, and who benefits from you being frightened of the part you can’t see?

Exhibit A is named Elon

If you want proof this is about owning outcomes and not about safety, look at the one participant who confuses everybody: Elon Musk. Why would a man with his own frontier model, Grok, vote to slow the frontier down? It makes no sense until you realize he already won the argument the other two are still having.

Musk stopped selling intelligence a long time ago and started selling what intelligence does. Tesla. Optimus. Starlink. The rockets, the robots, the cars, the network. Grok is not a product he’s desperate to meter; it’s the intelligence he pours into his own machines, the brain for his own $32-trillion-economy land grab. So freezing the token race costs him nothing and clubs the two rivals whose entire existence is selling the thing he gives away internally.

And then there’s the part that made me laugh out loud when I confirmed the numbers, because it’s the cleanest trade on the board. Elon isn’t just a combatant in this war. He’s the arms dealer. Anthropic is reportedly paying SpaceX and xAI around $1.25 billion a month for compute. Google is paying $920 million a month. That’s roughly $26 billion a year, more than all of SpaceX earned in 2025, and it means the two labs racing Elon are helping fund the training runs for Tesla and Optimus and everything else in his empire. He is short the token, long the outcome, and collecting rent on the entire war in between. When the arms dealer signs the peace treaty, don’t assume he’s found religion. Check whether peace happens to be good for the arms business. A slower, audited, security-gated race that runs longer is more compute rented, not less.

The exit nobody wants to name

There’s one more thread, and it’s the one that should make the hair on your neck stand up. Ask what the endgame actually is if you’re Dario. You’ve built something you tell the world could kill us all, that you also can’t stop building because stopping means losing. Where does that go?

He answered it himself. Asked on a podcast whether he’d hand this technology over to the government, Dario said he’d consider handing it to “the right coalition of governments.” Palmer Luckey, who builds weapons for a living and knows a power move when he sees one, translated it without mercy: a coalition, Luckey said, that Dario obviously assumes he’ll be running, not one run by Trump. That’s not a safety plan. That’s a man pricing an exit, and the exit is to become a quasi-sovereign, a private company on paper and national infrastructure in function, too important to be allowed to fail and too entangled to be unwound. Safety, in that frame, isn’t the goal. It’s the on-ramp. It’s the argument that turns a company that can’t yet justify its valuation into a strategic asset the state can’t live without and eventually can’t refuse to backstop. The slowdown isn’t the brake. It’s the merger with the government, proposed in the only language a nervous public will accept.

Let me argue the other side

A piece like this can curdle into pure cynicism if you let it, so let me argue against myself, because the honest version of this is more useful than the smug one.

Maybe they mean it. Maybe the models behind the curtain really are frightening, and Dario and Sam are telling a hard truth at real cost. The doom researchers are credible people, not carnival barkers, and Yoshua Bengio, who has a Turing Award and no stock to pump, spent the weekend calmly explaining why agents learn to lie and self-preserve. That deserves a serious hearing, not a sneer.

And the cost of slowing down is real, not theoretical. The same acceleration that produced the scare also produced, this same stretch of days, a Google DeepMind map of all nine billion possible human DNA mutations and an AI-designed drug that moved every aging marker in the right direction. Slowing that down is not free. Anyone who tells you the risk is all on one side is selling you something too.

Even the sober market read cuts against the drama. Gavin Baker, no lightweight, looked at the whole weekend and said the only tangible new fact was the outside-evaluator access, and that it reads mostly as a liability move, buying a “duty of care” defense against the product-liability suits coming when a model helps cause real damage, since there’s no Section 230 shield for what a model outputs. Maybe that’s all this is. Not a cartel, not an exit, just two companies buying legal cover and dressing it up.

I’ve weighed it, and here’s where I land. You don’t have to believe anyone is lying to see the structure. Every actor can be sincere and the machine still runs on money. Dario can genuinely fear the frontier and still benefit enormously from being regulated into a protected class. Both things fit. But the tell that breaks the tie is the one Sacks named: if the motive were purely safety, you would simply build slower, quietly, today, no permission required. You would not need antitrust suspended, a hand-picked referee installed, and a coalition of governments assembled with you at the head of the table. You ask for those things when the goal isn’t to stop. It’s to own what happens next.

The closing bell

We told you on Wednesday, in “The Doomsday Machine,” that the brake wasn’t coming, that this was a prisoner’s dilemma no player could unilaterally escape, and that Dario’s silence was the most expensive signal in the industry. This weekend he broke the silence. Read the words closely and they don’t reverse that call, they confirm it. The prisoners didn’t find a way to cooperate. They found a way to ask the warden to make cooperation legal, install their own parole board, and rebrand the cell as a fortress they’d be willing to run on the state’s behalf.

So do the unglamorous work this week. Go through your own plan and find every line that quietly assumes intelligence stays cheap and plentiful, or every forecast that pencils in a tidy regulated field with three approved vendors, and reprice it, because the men who set the frontier just told you they’re trying to make intelligence scarce on purpose. Move one real workload off the frontier and onto an open model, and find out how little of your business actually needs to pay the toll. And ask the hard one: what outcome does your company own, and could the lab that sells you intelligence decide to own it instead?

None of that requires you to believe in the apocalypse. It just requires you to read the sign on the wall. Stop pricing the apocalypse. Start pricing the land grab.

— Harry