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

$40,000,000,000,000. The US just crossed forty trillion dollars in debt. What's supposed to pay it down is the tax base, and the load-bearing piece is the withholding off a human paycheck. AI's entire pitch, a $30 trillion market, is to swap those paychecks for machines that pay no withholding at all. Every dollar the machines win by deleting a wage is a dollar pulled out from under the debt. That's the trade nobody's pricing: the US Treasury is the short side of the AI boom.

No skim tonight. This one is the whole argument, because there's nothing more important I could put in front of you to start the week.

F. Scott Fitzgerald said a first-rate intelligence can hold two opposed ideas at once and still function. The market has taken him too literally. It is long the AI trade and long the US Treasury at the same time and calling that a diversified portfolio. It isn't. Those aren't two opposed ideas you hold in balance until the facts arrive. They're mutually exclusive outcomes. One of them is false, and the market is going to find out which the hard way.

Start with the number nobody wants to sit with. The US just crossed $40 trillion in debt. What services that debt is the tax base, and the load-bearing brick in that base is the withholding that comes off a human paycheck — Social Security, Medicare, the whole promise runs on payroll taxes skimmed from wages. Now read Anthropic's IPO pitch: a $30 trillion market, nearly the whole US economy, built on AI doing the work people used to do. My former partner Brendan Maher put it better than anyone on TV: "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." The machine built to grow the economy pays nothing into the system its promises stand on. So the more AI works, the faster the wage base erodes under $40 trillion of debt. That isn't a hole in the bull case. It IS the bull case, read from the Treasury's side of the table.

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 trucks batteries and soon: robots and chips. The other, in Fremont, is being turned into a humanoid-robot plant aiming at something like 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 the robot doesn't get paid either. White-collar routes to the agent, blue-collar 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 take-or-pay compute contracts under the whole build are structured like the 2/28 subprime mortgages of 2006 — cheap while the data center gets built, brutal the day it switches on. Some $2.1 trillion of these obligations sit off balance sheets, more than the entire 2007 subprime market, and they start billing on delivery in 2027 and 2028, on a schedule already set. The revenue to cover them is still a rumor: across three years of earnings calls from the sixty biggest financial firms, one named a real dollar return from AI. About $19 million. And the honest price is hidden from all of us — a $200 ChatGPT Pro plan absorbs about $14,000 of tokens at list prices, and the provider loses money the moment a heavy user burns an eighth of what they're allowed. Every impression you have of how cheap these tools are was formed while somebody else paid the bill.

Then the part that moved me from cautious to bearish. Underneath the splashy headline — Nvidia and the biggest PE shops on earth, Apollo, BlackRock, Blackstone, Brookfield, Goldman, KKR, mobilizing more than $500 billion for compute — the SEC quietly cleared the way for data-center asset-backed bonds, exempting them from the risk-retention rule Dodd-Frank added after 2008 to stop exactly this. Risk retention is the skin-in-the-game rule. They just took it off. So the smartest, earliest money is busy packaging this exposure into paper it can sell to everyone else. That is not a courtesy to democratize access. It's a signal that the asset has gotten too big and too risky to hold, and the smart money wants it spread onto somebody who isn't them. In Canada they have a word for who's left holding it: the bagholder. If Michael Burry is looking for a catalyst, that press release may have been the bat signal.

You can see the strain in the bond market already, the 30-year above 5.3% and yields at multi-year highs in Japan, Germany and France at once, because governments, the AI build and everyone else are bidding for the same long-term money. Paul Tudor Jones drew the loop all the way through this week: the market is at 252% of GDP, and "10% of tax revenues are capital gains, they go to zero" in a real drawdown, the deficit blows out, the bond market gets smoked, and it feeds on itself. Two former equity-research analysts from the same First Boston desk I sat on, decades apart from the news, arriving at the same place. That's not a mood. That's the tape.

Here's the frame that holds it together, and I owe it to Brendan, who pictured a sailor threading the two monsters of the myth. Odysseus had to pass a strait with Scylla on one cliff and Charybdis on the other, and he could not avoid both. That is the American position, and it's why I keep calling this a timing mismatch instead of a bubble. There's no course that dodges both rocks — only the narrow water between them, and how much you lose getting through.

Scylla is the case where AI wins: Anthropic dents that $30 trillion, the machines take the wages, and every word Brendan wrote and every loop Jones drew comes true — the withholding base cracks, capital-gains receipts follow the market down, and we get a real debt crisis. Charybdis is the case where AI stalls: low growth, an aging population, debt suffocating investment, and a country whose only real growth lever left is the AI spending that just stopped paying. In that world the labs can't fund themselves (only four or five entities on earth can write the checks a frontier model needs), the AI-levered market unwinds, and the $40 trillion is still there.

There's one narrow channel between the rocks, and it has a name: yield curve control. The Fed caps rates by buying whatever it takes, inflation runs hot above those pinned rates, and the debt melts in real terms while savers pay the difference with no tax line to point at. It's the WWII playbook — last time debt-to-GDP looked like this, financial repression ground it down to about 30% by 1980. One analyst laid out this weekend that the plan looks staffed: Bessent at Treasury, Warsh at the Fed, both partial to exactly this, and Trump was never going to replace Powell with a bigger hawk. 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 graduate housing has gone from a quarter of a starting salary to two-thirds. A generation of quietly inflating the debt away has already priced hard assets out of reach.

And it rhymes with the bonds. The Fed's version 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. But read the WWII escape to the end, because it breaks in the spot that proves the point. That playbook worked because the country had a baby boom behind it, a wave of new workers to 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.

Give the boosters their best case, the whole thing. Maybe the machines make everything so cheap and so abundant that we cross into a golden age — food, medicine, software, energy, 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. 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. Even paradise, on this plumbing, comes due.

So whose side am I on? If the two can't both be true and I wouldn't be long Anthropic at two trillion, 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'd 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 priced for a future that may end up belonging to the bondholders, or an IOU from a government already planning to inflate its way out. The move isn't Scylla or Charybdis. It's refusing to be made of what 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'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 this year: "AI is a bubble precisely because it works, and working technology attracts money, speculators and charlatans." I believe in AI completely; it's the next leg of efficiency in knowledge 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 price, 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.

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 is now an asset, a fortress balance sheet with no 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. 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 against exactly the devaluation this whole letter 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.

I keep coming back to Michael Burry. He read the subprime data in 2005, saw a reset wall that was mathematically certain, bought the puts, and then sat there for a year while the thing he'd bet against kept climbing and his own investors begged him to fold. The defaults didn't tick up until the second quarter of 2007. He wasn't wrong. He was early. And the securitization here is only just starting, which last time was the late chapter, not the first.

The AI trade and the American balance sheet are two things that can't both be true. The reset walls, corporate in 2027 and sovereign right behind, are already on the calendar. I'd bet on the government before I'd bet on Anthropic at two trillion, and I wouldn't much want to own either. Maybe that's early. It usually is. But early is not the same as wrong.

You can't be long both. Get off the paper.

None of the above is investment advice — it's one guy who's watched a few of these cycles telling you where he's choosing to stand. The full Signal/Noise, with far deeper insight and analysis is live at getcoai.com.

— Harry and Anthony

Sources: