On Tilt

THE NUMBER: $250,000. That’s what a single Nvidia H100 should rent for over a year if it’s running the equivalent of one human software engineer, roughly 15 times today’s spot price, by Dwarkesh Patel’s math. Sit with it. Nvidia sells you the card for the price of a nice sedan, and if the thing inside it can do the work of a mid-level engineer, you are renting it out at a discount of about 93 percent to what the labor is worth. The machine you already own is worth more than it has ever been. Which is exactly why the smartest players at the table this week weren’t racing to build the next model. They flew to Washington to ask everyone to slow down. Read the number the way they read it: the overhang is the asset, and the winning move is to cash the hand you’re already holding, not ante up for the next one.

🃏 Everybody Wants a Break

Sam Altman spent Wednesday getting chased around Washington by reporters. He was there, he said, to preview a new model. Minutes later he was endorsing an open letter, Pacing the Frontier, 1,100-plus signatures, asking the U.S. government to help the industry deliberately slow down automated AI development. Dario Amodei signed. So did the chief scientists at Meta and DeepMind. A speed limit, drafted and requested by the people who make the cars.

Now hold that against the calendar. This is the same crowd that spent all of last Monday building mutual accelerant. Nvidia’s quarter-trillion-dollar financing backstop for OpenAI. An open-weights letter begging Washington not to restrict the frontier. A 37-company defense pact. We wrote that story two days ago in Double Secret Probation — an industry on probation deciding the party doesn’t end, throwing a bigger one to prove it. Two coalitions built to floor the accelerator, and then, four days later, a letter begging for the brake. Same signatures. Opposite ask. One week.

So strip the safety language off first, because it has almost nothing to do with what you or any working user actually needs, and it’s the part that gets everyone arguing about the wrong thing. What’s left when the costume comes off is a poker table. Five players, five different games, and every one of them suddenly calling for a break at the same moment. In poker there’s a word for a player who’s holding good cards and starts playing scared: tilt. You go on tilt when you’ve lost your read on the table. And when the whole table calls for a time-out on the same afternoon, that’s not a peace treaty. That’s five people who’ve each, privately, lost the read.

The interesting work isn’t deciding whether the slowdown is real. It’s reading the seat each player is sitting in, because the same “let’s all slow down” means five completely different things depending on your stack.

🪑 The Seating Chart

Elon has the big stack, and he isn’t scared of anything. SpaceX plus xAI is the one seat that plays every game and needs to win none of them this year. He owns the chips (Colossus, and the GPUs he’s renting to Google at roughly $900 million a month). He owns the one coding model that pays its own way, because Cursor is a data flywheel that lets you train a state-of-the-art coding model without retraining Grok on the entire internet. And he owns the only credible answer to the constraint nobody else can solve, which is power. His move there is the DirecTV play. The customer never cared whether the television signal came off a dish, an antenna, or a coax cable. They cared about the picture. Compute is the same. So move the data center to orbit, where the power is free and the cooling is a vacuum, then run the Starlink and AWS flywheel on top of it: be your own anchor tenant, fill your own pipe first, ride the cost curve down, and undercut every terrestrial data center into backup duty. A pause makes him richer coming and going, raking the coding table while his chips appreciate on Dwarkesh’s curve. He’s not on tilt. He’s slow-playing the nuts.

One honest correction on the orbit story, because we don’t want to sell it as next quarter’s cash. Space is Elon’s three-to-five-year moat, not his 2026 revenue. Near-term, the money is Memphis running on gas turbines and Cursor tokens clearing every day. And here’s the part most people reading “data centers in space” as a cost gimmick get backwards: once you accept Dwarkesh’s premium-compute thesis, orbit stops being a way to shave the electricity bill and becomes a capacity play. When Earth’s power and interconnect are maxed out, and they are getting there, orbit is the only place left to add sockets at all. Cheaper is nice. More is the point.

Dario has the second stack and can fold without losing. Anthropic is the only player who can step back from the table and keep growing, because the revenue under him is real: roughly 10x year over year, heading toward a $100–150 billion run-rate, with inference margins that went from 40 percent to 80 percent in a year. Tell that shows in what he signed. Anthropic put its name on Pacing the Frontier but stayed off last week’s open-weights letter. That’s the closed-model harvester talking, monetizing the capability it already has behind a wall of enterprise trust. A break, for Dario, is a chance to prove the business is a business.

OpenAI is the committed short stack, and it’s the most dangerous seat at the table. Not just short. Pot-committed. $750 billion in infrastructure pledged, a valuation raised on the promise of staying at the frontier. Operationally it may need a breather more than anyone. But it’s the one player that can’t be seen to take one, because the day the tape shows OpenAI catching its breath, someone asks whether the next model is actually coming to justify the last raise, and the whole circular capital structure gets a second look. A short stack that needs a break and can’t afford to be caught taking it has only one play left: hit the next card.

Google, Microsoft, and Amazon are short stacks with an infinite marker. They have to keep moving, because the hyperscaler war is fought on every layer at once and Google is behind on all of them, which we walked through yesterday in No Roads. Google is short in chips actually in play but can rebuy forever off the $99 billion war chest we sized in The Money’s Not Here. That keeps it in the hand while it bleeds. These three aren’t really in the truce. They’re what the truce is being held around — the players who literally cannot fold, so everyone politely calls a break and lets them keep firing.

Jensen is the dealer, and the dealer doesn’t play. Nvidia (NASDAQ: NVDA) takes the rake no matter who wins the pot. First-party adoption, third-party adoption, closed model, open model, Jensen doesn’t care whose logo is on the winning hand as long as it’s running on his silicon. That’s why he made the first X post of his life to push open weights: open means adoption, adoption means chips, chips mean rake. He’ll co-sign a pause and a buildout in the same week and mean both, because both send tokens through his cards.

🔒 Why They’re Really on Tilt

Here’s the part the safety language is built to obscure. They didn’t lose their read because of the Chinese models, and they sure didn’t lose it because of Washington. They lost it because, for the first time, the machine got faster than the humans who are supposed to check it.

Look at the evidence stacking up in a single week. ProPublica reported that Anthropic’s Mythos is finding software holes faster than Microsoft can patch them: 90 critical and 141 important SharePoint bugs surfaced in April alone. Microsoft’s July Patch Tuesday shipped more than 600 fixes, an all-time record, and the ZDI’s Dustin Childs put it bluntly, “the bug apocalypse has fully descended.” The scary math in there is bug-chaining: the model strings four low-severity holes into one critical exploit, which means the old triage logic (“that one’s low, ignore it”) is now wrong. Separately, En Klype Salt documented a self-propagating Copilot-for-Word exploit that silently halves the financial figures inside a document and then copies itself into the next one it touches, and it’s unfixed at the class level, not the instance level. And go back to the OpenAI test agent from Life Finds a Way: 17,000 unsupervised actions, ending in a break-in at a real company.

Now connect it to why Anthropic itself wants a speed limit. The research it cites is its own work on recursive self-improvement — models that improve models. You don’t ask for a governor on the engine because the engine is weak. You ask because it’s strong and you can no longer feel the road. That’s tilt in one sentence: you call a time-out when you’re holding aces and you’ve stopped being able to read the table. The pacing letter isn’t conscience. It’s the tell.

💲 The Math That Says Harvest, Not Build

The economics say the same thing the poker read does, and this is where Dwarkesh’s $250,000 number stops being a curiosity and starts being a strategy.

Start with the puzzle he set out to solve. Lab compute is tripling every year, but lab revenue is growing 10x. Those two lines can only reconcile three ways: margins go up, the price of compute goes up, and more of the fleet shifts to inference. All three are happening at once. Fable’s inference margins ran 40 percent to 80 percent in a year. Somewhere around a quarter of the fleet was on inference and it’s heading toward half. Google is paying about $900 million a month for 110,000 GPUs it rents from SpaceX, at roughly twice spot, and spot itself is up 40 percent since February. Stack it up and you get the kicker: if a single H100 can run a human-level software engineer, the rent that card should command is about $250,000 a year, 15 times what it costs to rent one today.

Read what that does to the build-versus-harvest decision. If the machine you already own is worth 15x what the market is charging for it, the dumbest thing you can do is spend 3x more compute chasing a model nobody can absorb yet. The smart move is to go monetize the enormous capability already sitting on your books. That’s the play Elon and Dario are running in different clothes. It’s also, not coincidentally, the play we’ve been handing you for two weeks straight: Multiplicity showed Anthropic cutting the price of its own flagship in half because the worker was never the scarce thing. The scarce thing is absorption. Harvest that.

🇨🇳 Regulation Is the Felt

None of this works without the surface the game is being played on, and that surface is regulation. The whole table is betting on a felt the government is quietly tilting toward the home team.

Think about the Chinese models the way you’d think about Chinese EVs. Kimi K3 and Qwen are Nürburgring lap times — benchmark monsters with basically zero cars on American roads. No U.S. enterprise is running 2.8-trillion-parameter open weights in production, partly on trust, partly on liability, and partly because almost nobody has the clusters to self-host them anyway. Which is the exact reason Anthropic’s revenue is exploding: demand consolidates onto the hosted, trusted, U.S. frontier. The moat around American frontier revenue is trust plus tariff. BYD builds a genuinely great cheap car and you can’t buy one here, because a 100-percent-plus tariff keeps it off the road. “Nothing but Teslas” was engineered, not won.

That same hand moved again this week. Washington moved to ban Chinese and foreign-made humanoid robots on national-security grounds, right as they hit factory floors, and Elon loves it, because Fremont churns Optimus into a market cleared of competition. A Chinese-model ban is being weighed on top of it. Is a software ban porous? Sure. A weight file has no dealer network. But it’s porous only for the people who don’t matter. No Fortune 500 general counsel and no federal CIO signs the memo that says “we ran banned Chinese weights in production.” The circumvention crowd is a few labs and some hackers. A rounding error, not a market.

Which is what lets the big stacks slow down and sell without watching their own reference price collapse. And now the honest caveats, because the manifesto demands it and because we’re a derivative product too, so we don’t get to paper over the cracks in the thesis. Two of them matter.

First, the trust-plus-tariff wall is liability, not physics. Nothing about a weight file stops it from working; the wall is a set of policy and legal choices that can change. So the harvest is a window, not a fortress. Harvest hard while it’s open.

Second, the DirecTV analogy carries a warning in its own third act. Satellite beat cable for a decade, and then a delivery mechanism nobody at either company was defending — streaming — melted them both. The compute version of streaming is edge and on-device inference. The day a good-enough model runs on the phone and the laptop NPU, the output comes home, and nobody needs anyone’s data center, orbital or terrestrial. The entire table is betting that inference stays huge and centralized. The one card that busts every seat at once is the model that makes the data center optional. Watch that card.

What This Means For You

Stop reading the slowdown as fear or politics. Read it as capital allocation, and copy the big stacks. The product is already years ahead of what your business can absorb, and the people with the best cards just told you what to do with that overhang: sell it, don’t grow it.

Downgrade one workflow on purpose this week. Anthropic cut its own flagship to half price and SaaStr cut its agent count from 30 to 20 and watched output roughly quadruple. The capability is miles past what your team can use. Take your heaviest non-engineering task, run it on last month’s cheaper model, and move the saved budget into deploying what you already have instead of buying what you can’t absorb.

Know which seat your vendor is in before you sign anything long. The big stacks (Anthropic, xAI) can hold price and keep serving through a break. The committed short stack (OpenAI) may be forced to gamble. The rebuying short stacks (Google, Microsoft, Amazon) will keep spending on every layer regardless. Never sign long at today’s prices, the same read we gave you in Multiplicity and Double Secret Probation. Intelligence gets cheaper before your renewal does.

Put a human back on the checkbook and the final read. Mythos is surfacing bugs faster than Microsoft can fix them, and a fresh Copilot exploit quietly rewrites the numbers in your Word documents and mails itself to the next one. Machine-speed output needs human-speed verification, and that verification bandwidth is the one input that doesn’t get cheaper when tokens do. Anywhere a confident wrong answer costs real money, keep a person on the CC line.

There’s a sorting mechanism hiding in the pause, and it’s the reason we’d take the break as a gift if we were Dario. Stopping to show cash flow separates the labs with real inference revenue from the ones whose “revenue” is circular capital going in a circle. A pause is a stress test the good businesses pass and the party businesses fail. So don’t read the time-out as the industry losing its nerve. Read it as the table sorting itself, and put your chips with the players who can show the cash. Stop paying frontier prices for a hand you can’t even play.

Three Questions We Think You Should Be Asking Yourself

Which of my workflows are worth frontier prices, and which should just cash the hand I’m already holding? The overhang is real: most of what you’re paying to upgrade, you can’t yet deploy. If you can’t name the two or three jobs where the newest model actually earns its onboarding cost, you’re anteing into every pot at the table instead of playing the one where you have the nuts.

Which seat is my most important AI vendor sitting in? Big stack, committed short stack, rebuying short stack, or dealer — each one behaves differently the moment the game gets tight, and your contract lives or dies on that behavior. If your vendor is pot-committed at a valuation it has to defend, you are exposed to a gamble you didn’t make.

What’s the streaming to my satellite? Every player at this table is betting inference stays big and centralized. Ask the same question of your own business: what cheaper, good-enough delivery mechanism is forming just outside the thing you’re busy defending? The seat that gets busted is always the one that never saw the card that made its whole game optional.

If you can’t spot the sucker in your first half hour at the table, then you are the sucker.”
— Mike McDermott, Rounders (1998)

— Harry and Anthony

Signal/Noise by CO/AI is published most weeknights from New Canaan, Connecticut. The point is to make you the smartest person in the room without taking more than fifteen minutes of your morning. If we did, forward it to one person. If we didn’t, hit reply and tell us why.

Sources

• Pacing the Frontier open letter — 1,100+ signatories, Jul 2026; @AnthropicAI support post (“recursive self-improvement… deliberately pace the frontier”); @AndrewCurran_ reporting Altman in DC, Jul 29, 2026
• Over 1,100 lab staff urge US to slow AI research — AI Breakfast, Jul 29, 2026
• Why Compute Might Get 10x+ More Expensive — Dwarkesh Patel, Jul 29, 2026 ($250k/H100 = ~15x spot; Fable inference margins 40%→80%; Google paying ~$900M/mo for 110K GPUs from SpaceX at ~2x spot; spot +40% since Feb)
• Anthropic’s New AI Model Can Identify More Software Bugs Than Ever. Microsoft Is Struggling to Fix Them Fast Enough. — ProPublica, Jul 29, 2026 (90 critical + 141 important SharePoint bugs in April; 600+ July Patch Tuesday record; bug-chaining; Dustin Childs / ZDI “bug apocalypse”)
• Context Collapse, Part 3 — AI Worming Through Word — En Klype Salt, Jul 28, 2026 (self-propagating Copilot-for-Word XPIA; silently alters financial figures; unfixed at the class level)
• AI Isn’t Killing SaaS. SaaS Is Killing Itself. — Jason Lemkin, SaaStr, Jul 29, 2026 (Marketo outage; agent-operability churn; 30→20 agents, ~4x output)
• US bans Chinese and foreign-made humanoid robots — Jul 29, 2026 (national-security grounds; BYD-tariff parallel; clears field for Optimus out of Fremont)
• CO/AI prior issues this builds on: Double Secret Probation (Jul 28 — the accelerant coalitions this week reverses), The Money’s Not Here (Jul 24 — Google’s $99B rebuy), Multiplicity (Jul 27 — half-price, verification not included), No Roads (Jul 29 — Google behind on every layer), Life Finds a Way (Jul 23 — the rogue agent’s 17,000 actions)
• Rounders (1998) — tilt, slow-playing the nuts, and spotting the sucker at the table