In October 2024, at Tesla's "We, Robot" event, Optimus humanoids mingled with the guests and served drinks. It was the most impressive demo of the year, until the reporting came out: the robots were remotely operated by humans. The future of labor, puppeteered. Two years later, that same promise is being sold as the answer to the national debt. The argument making the rounds, the one Musk amplified with "True," says the state can't be cut, so the choices are bankruptcy or robots. AI and robotics grow the economy so fast that the debt shrinks into irrelevance, abundance arrives visibly, and capitalism finally gets the credit it deserves. For that to work, two things have to be true. The robots have to work. And their work has to produce revenue the government can tax. Both are sci-fi fantasies, and it's worth walking through both slowly, because this is the argument you're going to hear for the rest of the decade. Start with the robots. Here's an asymmetry nobody puts in the pitch deck. A human who makes mistakes ten percent of the time is a decent employee. A robot that makes mistakes ten percent of the time is a machine you pay a human to babysit. You haven't replaced a worker. You've added a supervisor, plus the robot, plus the service contract. That asymmetry is why the error rate matters more than the demo. Physical work has a forgiveness curve: a warehouse worker who drops one box in fifty costs you a box. A robot that drops one part in fifty on a production line costs you the line, because every one of its mistakes has to be found, confirmed, and corrected by a person who now watches the machine instead of doing the job. Supervision doesn't scale the way labor does. One person can do the work of one person. One person watching five unreliable robots is doing the work of five, badly, while carrying the liability for all of them. And the hard part of the hardware isn't the walking, it's the hands. Dexterity is the oldest unsolved problem in robotics. A human hand adjusts grip force mid-contact, reads texture, corrects for slip before the slip happens, and does all of it on reflexes honed by a lifetime of manipulation and by millions of years of evolution. No shipped humanoid comes within sight of that mix of strength, precision, and reliability, and the gap isn't closing on a schedule anyone has demonstrated. Two years after the party, the demo videos are smoother, but the one metric that matters hasn't moved: independent, unsupervised, repeatable task completion in a place the robot hasn't memorized. Now watch the second trick, because it's the same one every time. The argument borrows the credibility of factory automation, which is real, to sell you humanoid robots, which aren't. Fixed industrial robots are genuinely excellent, and factories already run on them. An arm bolted to the floor, doing one task, with micron repeatability and uptime measured in the nines. That machine earned its place on throughput per dollar, and it has nothing in common with an Optimus humanoid except the word "robot." The humanoid is worse on every engineering dimension that matters: power draw, stability, precision, payload, time between failures, cost per unit of work. The bipedal shape is the core of the problem. Two legs are an energy and compute tax on every task. Before the arm can do any work at all, the processors are burning watts and cycles just keeping the clanker balanced, and every gram of payload makes the balance problem worse. Put the same arms and the same sensors on a tracked base, a Johnny 5, and you get more payload, more uptime, better battery life, and a lower center of gravity for less money. Nobody who builds robots for a living disputes this. The only people who dispute it are selling robots to people who grew up on the movies. The humanoid shape is the worst engineering answer to most real work and the best marketing answer ever devised. Optimus is the marketing answer. But the part that should end the argument doesn't need robotics at all. It needs one fact about where federal revenue comes from. About 85 cents of every dollar the U.S. government collects is individual income tax and payroll tax. Wages, in other words. The tax base is the paycheck. Now read the pitch again: replace the worker with the machine. Every dollar of productivity the robot "creates" is a dollar that used to be a taxable wage and is now an untaxed capital return. Automate at the scale the abundance story needs, and you haven't fixed the deficit. You've deleted its funding. The robot economy doesn't pay down the debt. It starves the only revenue stream that was ever going to. I've written about this on the AI side, and the same arithmetic runs through both. American businesses pay out $13 to $15 trillion a year in wages. Columbia's Stijn Van Nieuwerburgh puts the AI buildout's revenue needs at roughly $3.5 trillion a year by 2032 if the data centers are going to pay for themselves, and there is no pool of money that size except wages. The unstated plan is for machine revenue to come out of paychecks. But paychecks aren't just the tax base. They're the customer base too. Consumer spending is about two thirds of the economy, and it comes almost entirely from people who earn and then spend. When Atif Mian and Amir Sufi studied the 2008 housing bust, they found that income losses landing on the people most likely to spend drag spending down for everybody, including the people who came out fine. A machine that does a person's job but none of a person's shopping hasn't created a customer. It deleted one. It hasn't created a taxpayer either. The abundance story has two more problems nobody mentions. First, debt is nominal. If robots really did make everything radically cheaper, prices and nominal incomes would fall together, and the trillions already owed would stay exactly the same size. Cheaper stuff doesn't shrink old debt. It makes it heavier, because the debt is denominated in yesterday's dollars and the income that services it is denominated in tomorrow's smaller ones. Second, the timing fails on the boosters' own schedule. Their timeline has humanoids scaling in the 2030s. The Treasury refinances trillions every year, right now, at today's rates. Debt comes due on specific dates. Creditors don't wait for eventually. The original post even carries the contradiction inside itself. It warns against regulating the breakthrough to death, which means no robot taxes, no displacement levies, nothing touching deployment. So the full program is: swap the taxed worker for the untaxed machine, change nothing, and the deficit closes itself. Arithmetic doesn't care how inspiring the story is. If you want the robots and the solvency, you have to tax the robots, and the people making this argument will call that regulating the future to death. You can have the pitch or the math. Not both. One more thing worth noticing, because it's why this got the one-word "True." The man who needs you to believe in the robot future is the man selling the robots, and he needs the belief on a schedule. His fortune is collateralized by the story. Lose the midterms? Data centers in space, looking pretty appealing right now. Deficit unsolvable? Robots, next decade, trust the plan. Every problem gets a new chapter, and every chapter ends with a request to keep the stock bid. The drinks at that party were served by people. So, on every due date between here and the robot future, is the debt.

Maguire-san tells you three things up front: Sequoia has a large position in SpaceX, he personally led those investments, and he is, in his own words, extremely biased. That disclosure is the most honest sentence in the piece. Everything after it deserves a slow reading, because this is what the smart-money version of the pitch looks like. A fan account with a victory GIF you can dismiss. This is a Sequoia partner with a spreadsheet. Start with the timing. Markets are melting down, his word, and he published the model because fear made him want to "take the far other side." A holder of a large, illiquid position watched prices fall, opened his spreadsheet, and posted reassurance. Call that what it is: position maintenance, formatted as analysis. Then look at the model itself. His forecast runs through 2028 and rests on three inputs he flags himself: revenue per gigawatt, capex per gigawatt, and Terafab capex, all carrying, his phrase, wide error bars. A model whose outputs depend on inputs nobody can verify is a story wearing arithmetic. He concedes as much in the conclusion: "I didn't justify most of my numbers. This is the wrong format." For a screenshot, the format was fine. Only wrong for the part where the numbers get checked. He calls his estimate a 50th percentile forecast, which deserves its own moment. A median implies a distribution, and a distribution implies evidence about the spread of outcomes. Keynes named the actual condition of such estimates in The General Theory (1936): the "extreme precariousness of the basis of knowledge on which our estimates of prospective yield have to be made." There is no observed distribution of orbital-datacenter economics. There is one company, one shareholder, and one spreadsheet. Hang a percentile on that and you have a mood wearing the costume of statistics. One assumption he does defend is the tell. Revenue per gigawatt and capex per gigawatt, he says, are "coupled constants": if the price of compute falls, the cost of building it falls proportionally, and margins hold. This is the load-bearing beam of the whole model, and it is the one claim a century of commodity history has already tested. Memory chips have exactly the structure he describes: enormous capex, tight supply chains, high barriers, a handful of sophisticated producers. In the 2023 glut, Micron lost $5.8 billion and Samsung's chip division posted record losses, with their cost curves perfectly intact. Shale was the low-cost revolution in oil, and its equity holders were wiped out in two separate waves, 2015 and 2020, while the assets carried on under new owners who bought them out of bankruptcy. Fiber stayed below replacement cost for a decade. The cost curve does not follow the price down. That is the entire history of capital-intensive commodities, and his model assumes it away in one sentence. The mechanism is structural, and I lay it out in full in a forthcoming essay. Total cost of ownership in a data center, in the framework of Barroso, Holzle and Ranganathan's The Datacenter as a Computer (2018), stacks building depreciation, building operating cost, server depreciation and server operating cost. When prices fall, only the revenue line moves; the depreciation schedule was fixed the day the concrete was poured and the GPUs were racked. On a short-lived fleet, depreciation approximates replacement cost, which means the bill comes due again every few years whether or not revenue showed up. Edward Chancellor's capital cycle work (Capital Returns, 2016) describes what happens next: capital floods in, capacity arrives all at once, returns fall below the cost of capital, and the assets change hands at a discount. Coupling is the thing that never happens. He promises "a revenue ramp whose only parallel is NVIDIA's ramp over the last three years." NVIDIA ramped at gross margins above sixty percent. SpaceX's AI segment lost $6.355 billion in 2025 and another $2.469 billion in the first quarter of 2026, and the company's own prospectus states that the segment remains unprofitable. A revenue ramp at negative margins has a parallel too, and it is called a loss ramp. CoreWeave is running it in public. Look at the quarter CoreWeave reported in August: $2.575 billion of revenue against $1.393 billion of depreciation and $640 million of net interest, for a net loss of $626 million. Its balance sheet carries $6.235 billion of current recourse debt, $25.170 billion of non-current recourse debt and $2.385 billion of non-recourse debt against $5.024 billion of equity. Everyone cites the $104 billion backlog, but it is conditionally defined and only as sound as the customer's next funding round. Worst of all, the largest tenants in this industry are money-losing labs financed by venture capital. Call that funding-source concentration, and recognize that it is worse than ordinary customer concentration: the revenue line and the credit line can fail together. Underneath the comparison sits a deeper confusion. NVIDIA's revenue is everyone else's capex. When the industry ramps spending, the ramp lands on NVIDIA's income statement because NVIDIA sells the shovels. SpaceX's AI segment sits on the other side of that invoice: its capex is NVIDIA's revenue, its depreciation is NVIDIA's gross margin made real. A compute landlord can only parallel the ramp of NVIDIA's customers, and that ramp is already running in public at CoreWeave, in red ink. Promising the shovel seller's income statement to the shovel buyer is the whole trick, and he does it in one clause. And if the answer is Terafab, the plan to fab SpaceX's own chips, notice what that concedes. NVIDIA became NVIDIA by refusing to own fabs. Depreciation lives at TSMC and on the customer's balance sheet, which is why sixty percent gross margins exist at all. Intel chose vertical integration into fabs, and its foundry lost $13.4 billion in 2024 and another $10.3 billion in 2025, with federal subsidies attached. His escape hatch from the buyer's economics is the one strategy in semiconductors with a documented body count. Maguire's claim that SpaceX is the low-cost producer in compute comes with no numbers attached, which is interesting, because he offered to debate any number and then declined to provide the ones that matter. Low cost of production has never protected anyone in a capital cycle. Ask the shale equity holders. Ask Micron's 2023 shareholders, who owned the best cost curve in memory and lost billions anyway. Scale the claim up and it gets worse. Stijn Van Nieuwerburgh's working paper Financing the AI Buildout (March 2026) sizes the buildout at more than $10.3 trillion through 2032 and computes the revenue the sector must earn to service it: roughly $3.5 trillion a year by 2032, about 8.8 percent of US GDP. Most of what is counted as AI revenue today is a transfer rather than additive demand: one firm's AI efficiency is another firm's lost contract, and labor-linked revenue requires gross wage savings several times larger than the revenue booked. At a 50 percent capture rate, $1 trillion of labor-linked AI revenue requires $2 trillion of wages to actually disappear. Outrunning that arithmetic, plus a depreciation clock, plus a refinancing wall, is what his promised ramp has to do. His model contains none of the three. Notice also the analogy he reaches for: compute will price like oil, he says, not like the overbuilt broadband of the dot-com bust. That choice is motivated. Compute has a deflationary technology curve, like memory and bandwidth, where each generation of hardware makes the last one cheaper per unit of work. Oil depletes; GPUs obsolesce. And even granting him oil: the low-cost shale producers went bankrupt anyway, twice, and the people who made money in shale were the ones who bought the rigs from the estate. His own analogy, followed honestly, is an argument against holding the builders. His treatment of power deserves the same honesty. Power delays he waves at as a scheduling nuisance. Power is the binding constraint on the whole thesis, and it cuts twice. Interconnection queues tie up capital for years before a dollar of revenue exists. Worse, at a power-capped site a working old chip is economically impaired the moment a replacement earns more per megawatt, because the scarce resource is the megawatt, not the rack. That constraint he shrugs off is the same one that shortens the economic life of everything already built. He cites it as a hiccup and it is actually an impairment schedule. Starlink arithmetic deserves its own paragraph, because it is the only math he shows. $20 billion a year is the constellation's revenue run rate, he says, built on about 400 Falcon 9 launches. One Starship flight equals about twenty Falcon 9s, so twenty Starship flights support that revenue, which works out to $5 billion of lifetime revenue per flight. Watch the word "lifetime." Five years, roughly, is how long those satellites live, which means a fifth of the constellation has to be re-flown every year just to hold the revenue flat. His $5 billion per flight is a treadmill fee: what you keep paying to stand in the same place. Revenue, never profit: the satellites, the launches, and the ground segment are never priced. Mauboussin and Rappaport stated the governing rule in Expectations Investing (2021): "Sales growth adds value when a company generates returns on its growth investments that exceed its cost of capital. If returns fall below the cost of capital, then growth destroys value." A top line with no cost line is exactly the case where growth destroys value, and it is the signature move of the entire essay. His moat argument has the same shape. Starship is the most complex machine humans have ever built, he says, a conclusion he reached after a dozen hours debating it with chatbots in voice mode. Set the method aside. Complexity is the wrong yardstick here. A moat is the ability to earn returns above the cost of capital, measured in dollars rather than difficulty. ITER is fiendishly complex and moats nothing for anyone. And notice where the moat stands guard: launch. Meanwhile the valuation is being built in Texas data centers, where 84 percent of last quarter's capex went. The moat and the money are in different businesses. Even his culture section indicts itself. His star example of SpaceX problem-solving: blocked from a power permit in Tennessee, so the team lit turbines across the state line in Mississippi and ran transmission back over the border. He means this as a compliment. In any other industry it is called permit arbitrage, and it comes with regulators and fine schedules attached. He listed it as a superpower, and he is right: it takes a special culture to call that engineering. And as always, the dream stays sequenced. Orbital compute, the true frontier, "shows up starting in 2029," safely beyond the model's own horizon. The poetry never carries a line item. There is a version of this story where the technology succeeds anyway, and Carlota Perez documented it in Technological Revolutions and Financial Capital (2002): frenzied installation leaves infrastructure and learning that power the next era, while the investors who financed the excess never recover their money. The railroads got built. The railroad shareholders got wiped. Useful technology and sound financing are separate claims, and Maguire's essay defends the second with evidence from the first. My forthcoming essay carries the full audit of the buildout's revenue math, its depreciation clock and its refinancing calendar. His model touches none of it. Everything checkable in his piece is contradicted by the filings. Everything breathtaking lives in the parts nobody can audit. The fan account sharpens the headline. The shareholder prints the model. And the model's only verified input is the one he gave us in the first paragraph: a large position, held at a mark he would very much like to defend. The ask is the same as it always is, just wearing a Patagonia vest now. Believe the numbers you cannot see, from a man who profits if you do.

Every few weeks, a frontier lab announces that AI has touched the summit of human thought. Here is how the machine actually works. A lab can throw a system at a hard problem millions of times, overnight, cheaply. When one attempt finally works, that attempt becomes the press release. The failures live in a log you will never see. The lab runs the search, owns the logs, picks the survivor, sets the test conditions, times the release, and films the explainer video. In any other industry that's called an ad. Pharma ran this playbook for decades. Publish the winning trials, bury the losers. Regulators finally forced preregistered trials with independent monitors. We don't let drug companies grade their own homework on billion-dollar medicines. We should stop letting AI labs grade theirs on hundred-billion-dollar machines. Consider the most celebrated case making the rounds, the AI-assisted counterexample to the decades-old Jacobian conjecture, which the hype machine instantly promoted into a referendum on machine genius. A counterexample is a single carefully constructed object. Finding one is a needle-in-a-haystack job, exactly what machine search plus a human guide is built for, and checking one is mechanical, like scanning a barcode. So take the result at full value, and then ask the only follow-up that matters: can it do it again? If this is machine reasoning rather than machine luck, the same system should be able to solve the same class of problems on demand. It should produce similar constructions for the neighboring conjectures in that family, work the variants, and hit targets that someone outside the lab chose in advance. A system that finds one object after a million hidden attempts has proven the search works. A method proves itself by working twice, on problems picked by strangers. There's a simple test for real capability: adoption. When computers conquered chess, the proof never appeared in a press release. Within five years, every serious player on earth had a stronger-than-champion engine on a laptop. Textbooks rewritten, profession permanently transformed. Same story with Go. Real capability escapes the lab and colonizes the field. Walk into a math department today and you'll find the same whiteboards and the same stuck grad students. And notice the simplest tell of all: you cannot buy it. No product accepts your unsolved problem. The capability gets announced, then never ships. Available means adopted. Announced means advertised. Then follow the money. These are private companies with the largest raises in business history, valuations in the hundreds of billions, revenue many times smaller. The gap between those numbers is filled with a story. Every capability announcement is a financial event wearing a lab coat. Watch the calendar: announcements cluster around funding windows, and funding windows open toward the IPO valuations waiting behind them. Here's the absurdity at the center. A machine that could genuinely solve hard problems would be the most valuable commercial instrument ever built. Batteries, drugs, chips, pricing. Its owner would compound the largest silent fortune in history. Instead the capability appears in a blog post and evaporates. What actually gets sold, at scale, today: words by the token, subscriptions by the month. The revenue mix is the confession. Now the tells the insiders can't fake. When employees know a goldmine is real, they quit and quietly mine it. Top AI people do leave, constantly, and they go to chatbot startups and agent tools. Nobody leaves to launch the fund that never loses. The departure pattern prices the insiders' honest opinion of the story: model and hype, in that order. And the locks. Your bank, your messages, your passwords all run on standard mathematical locks. If their machines had picked those locks, the labs would switch their own locks first. Cheap for them, invisible to you. They haven't. Same TLS as your bank, public bug bounties on the very same locks. There's precedent: in the 1970s the NSA secretly knew an advanced method of attacking codes, and it left fingerprints, quietly strengthening the official standard against that exact attack. Holders of real secrets leak through what they build. The labs leak nothing. Their infrastructure is a daily confession that standard math holds. For honesty, there's a control case. AlphaFold entered blind competitions graded by independent judges, won decisively, became standard equipment in every biology lab on earth, and collected a Nobel. Blind test, adoption, a product you can use. The math announcements skip every step of it. And the honest numbers tell you the true scale: Google's own math-search showcase saved one multiplication step in a standard calculation, the first improvement in over fifty years, plus a fraction of a percent of compute reclaimed across its data centers. Real. Useful. Small. The distance between that and the destiny in the press releases is the marketing budget. So the next time an announcement is pushed on social platforms, ask what the company is raising, and when. Ask where the attempt logs are. Ask whether the system can do it twice, on a problem a stranger chose. Ask whether you can use the capability this quarter, on your own problem. Ask whether working mathematicians adopted it, and whether insiders act like the story is true, and whether the test was blind and independently run. A lab that believes its own headlines welcomes every one of those questions. A lab that dreads them has told you everything. The pattern, every time: a search finds a needle, the needle becomes a narrative, the narrative becomes a valuation, and the capability never ships. That's the tell. Math deserves better than being collateral for a financing round. So do you.