Two Things That Can’t Be True At The Same Time

THE NUMBER: $40 trillion. That’s the US federal debt, which crossed the mark this month, months earlier than the forecasters penciled in. Hold it next to another number making the rounds: $30 trillion, the total addressable market Anthropic plans to wave at IPO investors, nearly the entire output of the country. Two of the biggest numbers in the world, and almost nobody has noticed they’re pointed straight at each other. What stands behind the $40 trillion is a tax base, and the load-bearing brick in that base is the withholding that comes off a human paycheck. What the $30 trillion promises is a machine that does the human’s job and pays no withholding at all. Sit with that, because once you see that the two numbers can’t both come true, you’ve seen the whole trade.

---

Two opposed ideas

F. Scott Fitzgerald said the test of a first-rate intelligence is the ability to hold two opposed ideas in the mind at the same time and still function. The market has taken him a little too literally. Right now it is long the AI trade and long the United States Treasury at the same time, and it calls the combination a diversified portfolio. It isn’t diversification. Those aren’t two opposed ideas you hold in balance until the facts come in. They’re mutually exclusive outcomes. One is false, and the market will find out which the hard way.

Let me say the thing the bears keep forgetting to say: AI is real. This is not a bubble-denial piece and it is not a Luddite one. The technology works, it’s already changing how real companies operate, and ten years out the world looks materially different because of it. Accept that as settled. The argument here isn’t with the destination. It’s with the passage — the years between the bills coming due and the payoff showing up — and whether the country and the companies can get through it without wrecking on one rock or the other.

I got to this the way I get to most things now — an email. Wednesday I forwarded our issue on Anthropic’s $30 trillion pitch to my former partner, Brendan Maher, a sharp guy who has spent a career around markets. He wrote back fast. First a question: “Long-dated puts? The black swan been spotted?” Then, a few exchanges later, the line that became this whole issue. I’ll quote him directly because I can’t say it better: “the real loser is the US Treasury and entitlement programs. Agents, robots, etc. don’t pay withholding taxes, and therefore we probably expedite the time to US default with every advance AI makes.” That’s it. That’s the piece. Everything below is just the arithmetic underneath Brendan’s two sentences.

The withholding the machine doesn’t pay

Start with how the American promise actually gets funded, because it’s less abstract than the debate makes it sound. Social Security, Medicare, most of what the government owes its own citizens — that’s paid for by payroll taxes. Money skimmed off a wage, every two weeks, before the worker ever sees it. The whole apparatus assumes a country full of people earning wages the government can tax on the way through, employment doing double duty: producing the output, distributing the purchasing power, and funding the state that backstops both.

Now introduce a machine whose entire selling proposition is that it does the work without the wage. An agent doesn’t get a paycheck, so there’s nothing to withhold. A robot doesn’t pay into Social Security. You don’t pay it at all. This isn’t a bug the labs are hiding; it’s the product. When Anthropic tells investors its market is $30 trillion, the pitch is that AI captures a huge slice of what human labor is paid today. Every dollar captured that way is a dollar of wages that stops existing, and with it the withholding that wage threw off. The engine we’re all told will grow the economy pays nothing into the system its promises stand on.

One of the sharper voices I read on X put the structural version cleanly this month: “the economy can become post-labor before society becomes post-wage.” Output rises, profits rise, asset values rise — and the mechanism that hands ordinary people a claim on that output weakens, because that mechanism was always the wage. The value doesn’t vanish; it concentrates around whoever owns the compute, the energy, the data, the models. Brendan called AI “the ultimate value-concentration model,” and he’s right. A value-concentration model is a tax-base problem in a growth-story costume. You can run an economy where the machines do the producing. You cannot, on the current plumbing, run one where the machines produce and the humans still fund the government out of wages they no longer earn.

And don’t comfort yourself that your job is physical. The richest man on earth is building two factories right now. One, in Texas, makes Teslas batteries and soon will be adding Optimus robots and semiconductors. The other, in Fremont, is being turned into a humanoid-robot plant aiming at something on the order of a million Optimus units a year. So if the software agent doesn’t come for the desk job because your work is hands-on, the robot comes for the hands-on job, and it does it without a lunch break, a pension, or a bad Monday. Here’s the part that matters for this argument: the robot doesn’t get paid either. White-collar work routes to the agent, blue-collar work routes to the machine, and neither one has a dollar withheld from a paycheck it never collects. There is no wing of the labor force the tax base can retreat into.

The teaser period, again

Here’s where my finance brain starts itching, because I’ve seen the shape of this before and so have you.

A piece went around this week called “The Teaser Period,” and it nails the mechanism. The take-or-pay compute contracts underneath the entire AI build-out are structured like the 2/28 subprime mortgages of 2006. Cheap while the thing is being built, then a payment that steps up hard on a date fixed in advance. A lab signs a multi-year capacity commitment today, but the money doesn’t start flowing when the ink dries — it starts on delivery, when the data center is energized, twenty-four to thirty-six months out. During that window everybody looks brilliant. The cloud provider books the backlog, the lab books access to compute, the market prices both as growth. Nobody’s paying yet.

Then the meter turns on, on a schedule already written, whether the revenue showed up or not. There’s roughly $2.1 trillion of these obligations sitting off balance sheets right now. For scale, the entire subprime mortgage market at its 2007 peak was $1.3 trillion. Take-or-pay is debt — rating agencies have treated it that way for thirty years — it’s just not booked as debt here. And the payments commence in 2027 and 2028, which means the 2025-26 signing boom has already guaranteed a 2027-28 reset wall. The bill isn’t a risk. It’s a date.

And the revenue to cover it? Still a rumor. Somebody ran three years of earnings calls from the sixty largest US financial firms looking for a single realized dollar return from AI. They found one. One firm, about $19 million. Meanwhile the honest price of the product is hidden from all of us: a $200-a-month ChatGPT Pro plan can absorb $14,000 worth of tokens at list prices, and the provider starts losing money the moment a heavy user burns through a ninth of what they’re allowed. Every impression you have of how good and how cheap these tools are was formed while somebody else covered the bill.

The rats and the bagholders

Now the part that moved me from cautious to bearish, and it happened while I was writing this. Brendan sent it over with his cynical hat on, and he earned the cynicism — he had a front-row seat to the mortgage machine at Fidelity and after.

Underneath the splashy headline — Nvidia and the biggest private-equity shops on earth, Apollo, BlackRock, Blackstone, Brookfield, Goldman, KKR, unveiling a push to mobilize more than $500 billion in third-party capital for compute — sat the quieter item that actually matters. The SEC just cleared the way for data-center asset-backed bonds. Staff guidance determined that data-center ABS “are not asset-backed securities” in the traditional sense, which exempts them from the disclosure and, critically, the risk-retention rules that apply to ordinary securitizations. Issuance has already gone from $2.4 billion in 2020 to $15.5 billion in 2025, with records expected this year. Jensen Huang said the quiet part into a microphone: this is “the first time that technology chips have become an investable asset class.”

Read that the way Brendan read it. Risk retention is the skin-in-the-game rule — the requirement that whoever originates a securitization keeps a slice, so they can’t offload all the risk and walk. It was bolted onto the system by Dodd-Frank after 2008, precisely to kill originate-and-distribute. They just took it off. For data centers. So the most sophisticated capital in the world, the PE shops that got in early and cheap, is now busy packaging this exposure into bonds it can sell to everyone else. That is not a courtesy to democratize access for Joe Public. These are not generous people. That is a signal that the asset has gotten too big and too risky to hold, and the smart money wants the risk spread onto somebody who isn’t them. In Canada they have a word for who ends up holding it: the bagholder. Brendan, who lived through no-doc and low-doc at Countrywide and Washington Mutual, had the flashback the second he read the press release. If Michael Burry is looking for a catalyst, that release might have been his bat signal.

And it tells you who gets hurt. Not the shops selling the paper. The people buying in now — the late money, the retail allocation, the pension reaching for yield — are the ones standing where the music stops. The rats are already easing toward the gangplank. They’re just being polite about it.

Short of cheap balance sheet

Widen the lens, because the reset wall doesn’t sit alone. It sits in a world running out of the one thing both it and the Treasury need: cheap, long-term money.

Watch the bond market and you can see the collision get priced. The US 30-year has pushed above 5.3%, and yields are at multi-year highs in Japan, Germany and France at the same time, four countries with four local stories and one signal underneath. As one analyst I follow put it, governments, the AI build-out, and the rest of the economy are now competing for the same pool of long-duration capital, and “the world is running short of cheap balance sheet.” And this isn’t only the bears talking. Brett Winton, the chief futurist at ARK — Cathie Wood’s shop, about as bullish an address as exists — spelled out the same mechanism from the other side: because Nvidia chips hold their value, operators can collateralize debt against them, which lowers the cost of finance for even balance-sheet-tenuous builders, and compute then gets presented as an “investment-grade-y” asset class competing with ordinary corporate borrowers for funding. When the AI optimists and the AI skeptics independently describe the same crowding-out, it’s not a mood. It’s the plumbing.

Paul Tudor Jones drew the loop all the way through this week, and he’s exactly the voice Brendan was reaching for when he wondered aloud what a Druckenmiller or a Gurley would say. Jones: the US market is at 252% of GDP, against 65% in 1929 and 170% in 2000. Mean-revert to a normal multiple and you take a 30-35% decline. Then the part that connects to everything above — “10% of tax revenues are capital gains, they go to zero,” the deficit blows up, the bond market gets smoked, and the whole thing turns into a negative self-reinforcing loop. His other number is the one that rhymes with the securitization story: private equity was about 7% of institutional portfolios in 2007-08 and it’s roughly 16% now. We are far more illiquid, far more concentrated, and far more levered to the same trade than we were the last time this movie played.

Scylla and Charybdis

Here’s the frame that pulls it together, and I owe it to Brendan, who pictured a sailor trying to thread between the two monsters of the myth. Odysseus had to pass a strait with Scylla on one cliff and Charybdis on the other. You can’t avoid both. Circe’s advice was brutal and correct: hug Scylla, lose a few sailors, because the alternative — Charybdis, the whirlpool — takes the whole ship. That is the American position, and it is why I keep calling this a timing mismatch rather than a bubble. There is no course that dodges both rocks. There’s only the narrow water between them, and how much you lose getting through.

Scylla is the case where AI wins. Anthropic starts making a real dent in that $30 trillion, the machines take the wages, and every word Brendan wrote and every loop Paul Tudor Jones drew comes true — the withholding base cracks, capital-gains receipts follow the market down, the deficit blows out, and we get a genuine debt crisis. Success is the monster on that side.

Charybdis is the case where AI stalls. And this is where I part ways with the smartest bear in my inbox, Phil McAlister, who argues the dominant force isn’t inflation at all but disinflation — that demographics and a rising debt-to-GDP are crowding out growth, that breakevens are pricing barely 2% inflation, that the whole doom-by-inflation story is backwards. He may be right about the mechanism. It doesn’t get you off the ship. Because his disinflation world is Charybdis: low growth, an aging population, debt suffocating private investment, and a country whose only real growth lever left is the very AI spending that just stopped paying off. In that world the labs can’t fund themselves — only four or five entities on earth can write the checks a frontier model now needs, Poolside just missed a raise and lost its cluster, and the mainframe gets eaten by the PC. Which is why the most interesting hardware launch of the week wasn’t a lab’s at all. It was Apple, shipping an M6 Mac mini and an M5 Ultra Mac Studio built to run AI locally and to cluster, the desktop coming for the timeshare exactly as it did in the eighties. AI failing doesn’t hand you a soft landing. It hands you a popped bubble on top of an unsustainable balance sheet with the growth lever broken.

So McAlister isn’t a rebuttal to the thesis. He’s the other rock. If the bulls are right, we hit Scylla. If he’s right, we drift into Charybdis. There is one narrow channel between the rocks, and it has a name: yield curve control. The Fed caps interest rates by buying whatever it takes, inflation is allowed to run hot above those pinned rates, and the debt quietly melts in real terms while savers pay the difference without ever seeing a tax line. This isn’t theory. It’s the WWII playbook — the last time US debt-to-GDP looked like this, financial repression ground it from wartime extremes down to about 30% by 1980, the very room that later let Volcker break inflation. And the plan looks staffed. One analyst I read this weekend laid it out: Bessent at Treasury, Warsh at the Fed, both partial to exactly this, and the tell is that Trump would never replace Powell with a bigger hawk after a year of calling him too tight. Inflating the debt away was always the plan.

You don’t even have to imagine the repression tax, because you’re already paying it. The house that cost $180,000 in 2012 costs $434,000 today, and for a new graduate housing has gone from a quarter of a starting salary to two-thirds of it. That is not a separate affordability story. It is this one. A generation of quietly inflating the debt away has already repriced every hard asset out of reach, and the bill has been landing on the public the whole time.

And it rhymes with the bonds. The Fed’s version of the move sends the bill to savers through inflation; Wall Street’s version, the data-center paper, sends it to whoever buys the bonds. Different desks, one instruction: socialize the risk of the AI-and-debt bet onto the public. Whether the machines deliver or not, the plan is for the citizen to hold the bag.

But read the WWII escape all the way to the end, because it breaks in the one spot that proves the point. That playbook worked because the country had a baby boom behind it, a wave of new workers whose wages you could tax and repress against for thirty years. Today the workforce is shrinking, and the engine you’re counting on to grow out of the debt is the very thing deleting the wages the playbook runs on. The escape hatch and the trap are the same lever. It could work. Sure. And a camel can get through the eye of a needle too. The Chinese and Soviet planners, with far more control than we have, never once timed a transition this fine, and I wouldn’t bet a dysfunctional Congress does better.

We’ve seen this movie, and we’re still not calling the end

Give the bull his fair hearing, because I’ve been early on enough doom calls to respect him. The honest rebuttal: AI raises productivity, productivity raises GDP, and a bigger economy is a bigger tax base even if the composition shifts — you stop taxing the wage and start taxing the owner, corporate profits, capital gains, maybe an automation tax on the machines themselves. And we’ve heard the labor-apocalypse story before; the tractor was going to end farm work, the ATM was going to end the teller, and instead new work appeared that nobody could have named in advance. All true, all fair. If the fiscal system could pivot cleanly from taxing labor to taxing capital, and the new work showed up on schedule, the strait opens into calm water.

Our claim is narrower than doom, and harder to wave off. It’s the timing. The debt is long, the data centers are long, the take-or-pay contracts are long, and the productivity payoff that’s supposed to justify all of it is late, uneven, and unproven. Something gives in the gap, and there’s no dial to turn that makes the tax code shift from labor to capital as fast as the machines shift the work. The politicians are already picking sides the clumsy way — a broad chip tariff under discussion would delete a fifth of planned data-center projects, and a wildly profitable, job-displacing machine is the most obvious thing left to tax once the boom is real. We’re not calling the end of the world. We’re saying there’s a dangerous passage to get through, and pretending you can be long everything through it is how people end up as the bagholder.

The golden age has an invoice

Give the boosters their best case, the whole thing. Maybe they’re right. Maybe the machines make everything so cheap and so abundant that we cross into something like a golden age — food, medicine, software, energy, most of the drudgery of life done for us at almost no cost. Set aside the harder question they never ask, of what a species built on struggle does with itself when the struggle becomes optional, and how a civilization rebuilds its sense of who it is once work is no longer the spine of it. Grant the entire utopia.

It still arrives with an invoice, because abundance produced by machines that earn no wage and pay no tax is abundance no government can fund itself on. The richer the golden age, the poorer the Treasury that’s supposed to underwrite it. That is the line nobody on the boosters’ side reads out loud: the better AI’s best case gets, the faster it bankrupts the states of the world. Even paradise, on this plumbing, comes due.

What it means for you

Three moves, and they touch you whether or not you own a single AI name.

Find out which side of the bet your revenue is on. Pull the report this week. What share of your revenue comes from selling human hours a machine can now do cheaper, and what share from something a machine can’t do, own, or verify? Long the wage base or long the machines — most businesses are quietly long both and have never asked. If you’re on the wrong side, the repricing is coming, and you’d rather meet it in a spreadsheet than in a renewal call.

Don’t buy the paper the smart money is selling. When Apollo and Blackstone start packaging data-center debt into bonds for everyone else, the interesting question isn’t the yield, it’s why they’re so eager to share. Before you or your pension reaches for a compute-backed instrument, ask who’s keeping the risk. After this SEC change, the answer is increasingly nobody — which means you. Late money in this cycle is bagholder money. Size accordingly.

Assume the devaluation, the tax, and the tariff — don’t hope them away. Build your plan for a world with a cheaper dollar, a higher cost of compute, and a government that has decided AI is a thing to tax rather than subsidize. If you’re holding cash, know what inflation does to it. If you’re modeling AI savings, model them above today’s cost of compute. The political price of this boom is going up, not down, and the people who plan for that keep more of the upside than the people it surprises.

Neither rock

So let me answer the question this whole thing has been circling, because a fair reader is entitled to ask it: whose side am I on? If the AI trade and the American balance sheet can’t both be true, and I wouldn’t be long Anthropic at two trillion, then the frame says I should want to be long the US government. I don’t. Not a little, and I say it as a citizen who would like nothing better than to be wrong. I think there’s a real chance of a debt crisis, and I’m all but certain the savings I’ve worked for get devalued. I do not want to be a holder of dollar-denominated debt.

That’s what “pick a side” gets wrong. Both sides of this trade are paper — a share of a company priced for a future that may end up belonging to its bondholders, or an IOU from a government that has already decided to inflate its way out. The move isn’t Scylla or Charybdis. It’s refusing to be made of the stuff either monster eats. You get off the paper. You stop holding promises of cash flow a long way out, and you start holding things.

Here’s the reconciliation, because I know how it sounds to be bearish on AI equity in the same breath I say I’d never bet against Elon. Bill Gurley — who sat on the same First Boston equity-research desk Brendan and I did, a lifetime ago — said it better than I can from the stage this year. “AI is a bubble precisely because it works,” he put it, “and working technology attracts money, speculators and charlatans.” I believe in AI completely; it is the next leg of efficiency in knowledge work and physical work both, and it arrives whether today’s equity holders own it or their creditors do. Believing in the technology and paying up for the security are two different acts. On the pricing, Gurley names the whole game: “You price to gain market share. You don’t price for profitability. You simply can’t.” Price always mattered. It still does. And competence matters more every year AI speeds the board up.

So if you made me put down names, I’d own the fortresses, not the frontier. Apple, with a balance sheet like a vault, the best way to own open models running on hardware you already control, a category of one. Nvidia, because compute itself has become an asset, a fortress balance sheet with no real peer. SpaceX, the finest manufacturing mind of the age, going full-stack down to the vanes on his own turbines, a category of one in market after market. And I’d leave the model companies alone; as Gurley says, funding a competitor to Anthropic, or handing Anthropic more, probably isn’t a good return. Fortress balance sheets, market leadership, real hard assets. That’s the list.

For the cash that has to sit somewhere, I’d rather it sat in a currency run by a government that can still count — the Norwegian krone, the Swiss franc, the Singapore dollar — than in the one drawing up plans to debase. I’d hold some gold and some bitcoin as insurance against exactly the devaluation this piece is about. And then the rare assets that grow and pay you while they do it: real estate, timber, water rights. Things. Not promises. Things.

None of that is advice. It’s just where a man who has watched a few of these cycles is choosing to stand.

Early is not wrong

I keep coming back to Michael Burry. Not the caricature — the actual trade. He read the subprime data in 2005, saw a reset wall that was mathematically certain, and bought the puts. Then he sat there. Through all of 2006 the thing he’d bet against kept climbing, his own investors demanded their money back, and one of them told him flatly he’d be broke before he was right. The defaults didn’t tick up until the second quarter of 2007. He wasn’t wrong. He was early. There’s a difference, even if it doesn’t feel like one while you’re waiting — and the securitization just started, which in the last movie was the late chapter, not the first.

That’s roughly where I sit. The AI trade and the American balance sheet are two things that can’t both be true. The reset walls — corporate in 2027, sovereign right behind — are already on the calendar. Scylla is success, Charybdis is failure, and the ship has to thread the water between. I’d bet on the government before I’d bet on Anthropic at two trillion. Maybe that’s early. It usually is. But early is not the same as wrong.

You can’t be long both. Pick a side.

---

The email edition — THE NUMBER, the three moves, and the day’s five stories — is in your inbox. If figuring out which side of this bet your own business is standing on is the question keeping you up, that’s the conversation we run at Outsider Labs.