People ask why I still argue with Musk fanboys. Fair question. The facts never seem to land, right? Here's the thing. The facts were never going to. I'm turning 49 in a few days and I'm far from naive. A big chunk of these guys aren't investing in a company. They're investing in a rescue. Musk is the last powerful white man telling them the replacement stops with him, the rocket comes back for them, the future's still theirs. You can debunk a revenue forecast. Good luck debunking salvation. So when you show them the ledger, they don't hear finance. They hear somebody kicking the last prop out from under their whole story. And the story is the point. Watch what happens in my comments. They hold the bag and the faith at the same time, and the tantrums are something to see, walls of text, personal insults, rocket emojis, every one of them convinced he just invented the argument. Losing money on TSLA isn't evidence against the belief. It's tithing. They're too far gone. I'm not a surgeon, I can't shrink an overactive amygdala for them. They're not the audience anyway. I'm documenting it for everyone else. Somewhere out there is a nineteen-year-old with a fresh brokerage account and no idea how the game works, and if the receipts are sitting in public when he goes looking, that's the whole job. The fanboy performs faith in my replies. Somebody has to keep the ledger where the lurkers can find it. Call it civic duty. The church can keep its congregation. I'm just making sure the exit door stays visible.

"He was a hero of the left until he showed conservative political stripes, then he became Satan to you." Notice what this assumes: that mockery of a moon chip fab is politics in costume. Let me answer it properly, because it deserves more than a reply. First, the record. I am not on a team. Not left, not right. To me, the dichotomy is boring and is the least interesting thing about a person. Learning where someone sits on it tells me nothing about their character. It tells me which narratives they grew up inside, and how hard they have worked to climb out of them. Usually the answer is: not very hard. Here is a more interesting question. If you had been born into the opposite environment, with parents and neighbors and friends pushing the other side's story onto you, would you have resisted it? Would you believe, today, the opposite of what you believe, with the same confidence? If the honest answer is yes, then your convictions are weather. They came with the climate. Character is what you would have carried across climates. Same question for race. Same question for country. The person you would have been in the other cradle is closer to who you actually are. Everything else is wardrobe. Robert Anton Wilson built a practice around this. In his book Prometheus Rising (1983), and later at his Maybe Logic Academy, he assigned what he called "paradigm shifting": take the position you despise and defend it honestly, at full strength, until you could pass for a believer. The point was never conversion. The point was to show you that your convictions are mostly software, installed by whatever room you grew up in, and that other software runs on the same hardware. Most people cannot finish the exercise. The ones who can never argue the same way again. Now back to the accusation. The post that earned me a political filing was about silicon. I quoted his own words, "next level of the video game," and noted that lunar soil being 20% silicon means about as much for chipmaking as a beach does. Nothing in the post about his stripes, conservative or otherwise. Nothing in any of my posts about his politics at all: capex lines, debt raises, deadlines with dates on them. He read a joke about moon rocks and saw a political enemy, which tells you exactly which lens he looks through. The man who hears regolith as a tribal attack is telling you how he hears everything. "But I'm the one with the bias." Correct. Just not in the direction he meant.

Elon wants chip fabs on the Moon now. Lunar soil is 20% silicon, he says, so you just mine it, refine it, stamp out solar cells and chips. Next level of the video game. His words. He's a walking 1954 Popular Science cover. "Moon Factories by 1975!" You can picture him red-faced at some poor SpaceX engineer, jabbing a finger at a render: "I saw this in a magazine! Why can't you guys do this?" Because the magazine was fiction, Elon. That's why. Boring question one: water. A single leading-edge fab burns through millions of gallons of ultrapure water every day, the kind filtered and polished until it turns chemically aggressive and rips ions out of anything it touches. Making it takes an industrial plant of its own. On the Moon you'd be chipping ice out of permanently shadowed craters at minus 230 C and feeding it through a purification chain that doesn't exist, powered by energy that also doesn't exist. Boring question two: shaking. Lithography machines sit on tuned vibration isolation, springs and masses and counterweights engineered around Earth's gravity down to the millimeter. Lunar gravity is one sixth of that. Every spring rate and tuned mass in the stack assumes 1 g, and the Moon has its own moonquakes nobody has fully mapped. So forget shipping a fab up there. You'd have to redesign the entire physics package from scratch, in a hailstorm of charged abrasive dust that eats seals and bearings for breakfast. And here's the part that makes it funny instead of just dumb. NASA studied all of this while Elon was still in college. Ignatiev and Freundlich published the actual paper in 1991: waste silicon from regolith processing, vacuum-deposited thin-film cells laid right on the surface. Serious work, fifteen years of it, and by 2005 the best anyone could promise was 5-8% efficient cells sprayed on melted moon glass, economical only at megawatt scale. Chips? Never in the plan. Not once, in thirty years of study. Then a 2006 follow-up walked back even the crude version, because melted regolith is too lumpy and dirty to make a decent substrate. The 2025 research gave up on purifying lunar silicon entirely and just ships the semiconductor layer from Earth. Thirty years of papers saying "hard pass" and this guy shows up with the crayon version of a 1991 NASA study, calls it visionary, and the fanboys clap like he invented the Moon. He isn't a genius ahead of his time. He's a man who skips the reading, and it shows every single time he opens his mouth just to pump his cult stocks.

No one's doubting cognition is useful. The question is whether it's $10 trillion of buildout useful, $3.5 trillion a year useful. Price comes from marginal cost and competition, not usefulness, which is why the most useful technologies of the last thirty years, bandwidth, storage, compute itself, got cheaper every single year. Useful and deflationary are the normal pairing for a commodity.

Every earnings call, every keynote, the ask is identical. More chips, more power plants, more data centers. Always more time. Ask how much more it takes to reach AGI, the human-level machine intelligence they keep promising, and you do get answers. Plenty of them. $500 billion here, $7 trillion there. But no two answers agree, none comes with a date that has ever held, and every missed target gets repriced upward. A promise that keeps moving has a name: a subscription. Here's the silliness at the center of it all. Almost nobody outside the field has ever heard the industry's founding bet stated plainly. You teach a program to guess the next word in a sentence. You feed it essentially the whole internet and score how often it guesses wrong. Their wager: as you spend more on computers, that error score falls smoothly, and somewhere along the way, real intelligence shows up. A lower error score does mean the model does better on quizzes and tests. Trouble starts there. Worse, those tests get contaminated (answers leak into the training data) and they top out, so the score gets less trustworthy exactly where the spending gets biggest. And doing better on quizzes was never proof of the thing being sold: a machine that can actually do valuable work on its own. A program that got better at guessing the next word got better at guessing the next word. That leap, from error scores to "thinking", is faith, not measurement. And everything in this essay, the data centers, the power plants, the loans, is a bet that it arrives. If it never does, the spending has no story left. Meanwhile the money math runs the wrong way. Each small improvement costs roughly ten times more computing power than the one before it. The progress crawls. The invoice sprints. And the money can't buy its way off the curve. Ten trillion dollars buys a few more steps along it: better quiz scores, smoother chatbots. What it cannot buy is a different curve. The leap is not for sale, and the whole buildout is priced as if it were. That's why the ask never ends. And the industry's newer trick, rewarding models for getting math and coding problems right, has the same flaw in miniature: train for the test and you get a very expensive test-taker. So what has the bet actually bought, four years and a trillion and a half dollars later? One genuinely good business: AI tools that help programmers write code, about $13 billion a year, and roughly half of that is the industry selling to itself. One big convenience business: chatbots. Then the other side of the ledger. MIT went and counted: 95% of company AI projects produce no effect anyone can find on the bottom line. S&P Global found that the share of companies abandoning most of their AI projects jumped from 17% to 42% in a single year. PwC asked CEOs directly, and 56% said AI has had no financial impact at all. McKinsey puts the share of companies getting real profits out of AI at about 6%. And the handful of projects that do work? Matching invoices. Processing expense reports. Useful, boring, small. Zero new drugs approved. Zero scientific breakthroughs from chatbots. And the crowning achievement of the biggest construction boom in American history? Programmers type less boilerplate, and people like me, plus every student with a deadline, get their essays ghostwritten. Full honesty here: I used these tools to help write what you're reading. If that makes you want to dismiss this essay as slop, go right ahead. You just supported my point. Even the one success story turns cannibal. A few weeks ago a single developer, Brandon Thomas, used AI coding agents to rebuild Adobe's entire creative suite from scratch, app by app, free and open source, working with Adobe's own file formats. Days later he did the same thing to Word, Excel and PowerPoint. Sure, the clones are rough; the project itself says they're maybe half ready for daily professional work. That isn't the point. One person, with an AI bill you could put on a credit card, just reproduced decades of engineering at two of the most valuable software companies on earth. Now watch where the value goes. Adobe and Microsoft lose the $100 a year. The AI company collects a few hundred dollars of usage. You get free software. Nobody in the AI stack gets paid. The productivity is real, and it works by destroying the exact software revenue the buildout was supposed to capture. As business models go, that's a bonfire. Here the defense reaches for history. A famous economist joked in 1987 that computers were everywhere except in the productivity numbers, and ten years later, there they were. Maybe AI just needs its decade. Here's the problem with it: electricity and personal computers got a generation to spread through the economy. This boom gets about two years, because the debt payments start in 2027 whether the revolution has shown up or not. That history lesson assumes time the loans don't allow. The other time excuse is shorter: we're early. Early for what, exactly? Ask which features we're supposed to be adopting and the answers turn to fog. Agents, automation, transformation. Words for a slide deck, not a product. And the analogy doesn't survive contact. Email over a screeching dialup line beat the fax machine on day one, and the first spreadsheet paid for itself the afternoon it was installed. What took twenty years was shipping machines and laying cable, not finding a use. This needed no shipping and no cable. It's a browser tab, free, already in front of hundreds of millions of people. PCs and the internet had proven value waiting on slow delivery. This has instant delivery waiting on proven value. And the numbers describe quitting, not waiting: companies are now abandoning their AI projects at two and a half times the rate of a year ago. Early means people haven't tried it yet. Abandonment at that scale is what it looks like after they have. Hold the two stories side by side and they don't fit. The spending is sized for a revolution that's already certain. The excuse says it hasn't started. What shipped is fluency. The models write beautifully. What the valuation needs is reliability, and reliability is a different animal. Five years of scaling made them far more fluent. It didn't make them reliable. A worker who's right 80% of the time can't hold a job that needs 99.9%, and there's no chart anywhere showing that gap closing. Being early has a price. Early on the PC cost you a machine that paid for itself in weeks. Early on this means investors front trillions for data centers, to be paid back by revenue the industry's own boosters place ten to twenty years out. There's a word for being that early: wrong. Be honest about the size of it, too. Call it $100 billion a year in real AI revenue. Call it $200 billion if you let the tech giants count generously. That's a fine industry, about the size of video games. Nobody hates Mario Kart. But the bills coming due in 2027 alone require about $1.3 trillion, and even the optimists' own spreadsheets get you to a third of that. The comparison doesn't matter. The gap does. Point any of this out and defenders retreat to AlphaFold, the AI that cracked protein folding and won a Nobel Prize. Let's grant the strongest version: a tech giant's profits paid for it, the way Bell Labs' old phone monopoly paid for the transistor. Now look at the difference. Bell's research came out of profits. This research comes out of borrowing. AlphaFold was built in 2020, before the spending wave, on computers so modest it can fold a protein in minutes on a single machine, and then its results were given away free. None of the giant data centers being financed today is for more AlphaFolds. They're for chatbots. If science were the point, you could fund a hundred AlphaFolds for the price of one of these campuses. The casino funds the opera sometimes, fine. Stop selling the casino as the opera. And then the money itself. This buildout is heading toward $10 trillion, and the loans behind it need $3.5 trillion a year in revenue by 2032 to pay off. That's almost 9% of the entire US economy. All the dollars American businesses spend on all software and IT combined today add up to 5%. A single technology is supposed to become nearly double the whole software industry in six years. Our one clean window into the real numbers is CoreWeave, a company that does nothing but rent out AI computing. It charges almost exactly the price the loan math requires, holds $100 billion in signed contracts, and still lost $600 million last quarter. The value is real. The unit economics are a donation. Don't take my word for any of this. Ten days ago Gartner, the biggest research firm in tech, ran the same math and published it. They figure the data centers need $8.5 trillion in revenue by 2030 to cover the spending, the interest, and a normal investor return. Their own revenue forecast tops out at $3.6 trillion. So they've said it in writing: more than $3.5 trillion of this buildout is at risk of write-downs by 2030, and four out of five GPU rental companies won't make it as independent businesses. The bills don't stop when the construction does. Start with electricity. Ten trillion dollars of buildout works out to something like two or three hundred gigawatts of computing capacity, and at full buildout the power alone runs about $120 billion a year, roughly a quarter of what America's entire electricity market takes in today. Then the bigger treadmill. About 60% of this buildout is chips, and chips are not bridges. They fail constantly (one failure every three hours on Meta's sixteen-thousand-GPU training run) and they get outclassed: NVIDIA ships a new generation every year or two, so the silicon is obsolete in three to five years no matter how the accountants stretch the schedule. Replacing $6 trillion of chips on that clock costs $1.2 to $2 trillion a year. Every year, forever. Construction projects end. This one just replaces itself, annually, at ten times the industry's entire current revenue. The veterans of the last hardware treadmill already voted. Core Scientific went bankrupt mining Bitcoin and now lives as CoreWeave's landlord. IREN, Cipher and Hut 8 made the same pivot. The survivors of that business rent out power and buildings. They won't own the chips. Now the deepest cut. What all of this machinery actually produces is tokens, the little chunks of text a chatbot spits out. At full buildout a token costs almost nothing to make. The electricity per token is too small to meter, and even the silicon wear comes to pennies per million on the commodity tiers. But the loan math needs tokens to sell for about 35 cents per million, at a hundred times today's volume, forever. And the price of any given capability has been falling tenfold a year for four straight years, the steepest deflation any industry has ever recorded. Supply multiplying a hundredfold, marginal cost rounding to zero, and a price that must not fall. In every market in history, that gap closes toward cost. Bitcoin mining at least had an escape hatch: the coin could pump. Inference only deflates. There's no halving-day rally for a server rack. This one has to be the exception, or the loans don't get paid. Sit with one more number. Of the roughly $1.5 trillion spent on this buildout so far, about $536 billion has flowed to a single company: NVIDIA. It designs the chips and builds nothing physical (its own capital spending last year was $6 billion), it keeps about 71 cents of gross profit on every dollar of sales, and it posted $120 billion of net income in one year. That is more profit than the entire AI application layer has ever produced. Half its sales go to a handful of tech giants paying with money raised against the promise that AI pays off later. In a trillion and a half dollar gold rush, exactly one layer is wildly profitable, and it's the layer selling shovels. Renters lose money. Model makers lose money. Buyers see nothing. The whole stack's economics rest on the hope that the people buying the chips eventually earn a return that today exists only at the company selling them. The last defense is always the railroads. Yes, the investors got wiped out, but society kept the tracks, and it was worth building anyway. Notice what that argument admits. If the best case for the economics is that shareholders get destroyed but the asset turns out useful, you haven't rebutted the argument. You've repeated it. Somebody pays for every mile of track. This time the somebody is pension funds, insurance companies, and every 401(k) holding an index fund. And this time there are no tracks to keep. The rails lasted a century; the chips last three to five years. Society keeps the empty shells, the power lines, and the debt. So when does AI deliver on the promises? What's actually needed? More data centers? More time? The industry gives out dates all the time. It just never gives the same date twice. The debt has dates. The promises keep moving. That asymmetry is the whole story.

Let me say something nice about AI, because it's true. I use it every day. It helps me write essays and it helps me do research, and it is genuinely, almost embarrassingly good at both. It reads everything, it never sleeps, it never gets tired of my fourth draft. As a sparring partner for thinking, there has never been anything like it. But a few years of daily use teaches you something the demos never show: it works best when you know exactly what it can and cannot do. It is brilliant at fluency and terrible at knowing when it is wrong. It will cite a source that doesn't exist with total confidence. So I check everything, and because I check, I get enormous value out of it. The productivity is real. It is also operator-dependent. It is a power tool, and a power tool in skilled hands is a wonderful thing. Now notice the irony. The people who use it most, who know its limits down to the bone, price it at $20 a month. The people who price it at trillions have never hit its walls. The valuation depends on not knowing what daily users know. The tool is good. I mean that. The story about the tool is where the con lives. A $20 subscription that saves me hours is a real business. A $26.5 trillion story that requires the tool to hold down jobs unsupervised is a fantasy. The difference between those two things is the entire bubble.

Honest question: early for what? Name the feature we're waiting for. Run through the options. Chatbots and drafting? Already here, already priced. Twenty bucks a month. Fine little market, but nobody needs to be "early" for something that exists. Coding? Also here, and it genuinely works, though it mostly works by eating the software companies that pay for it. So when you say early, you must mean the big one: AI that can actually hold down a job. No supervision, owns the outcome. That's the only version worth trillions. And it doesn't exist. Here's where your PC comparison breaks. The PC worked on day one. VisiCalc did your books the day you plugged it in. Email delivered the day you got a modem. Early buyers of those things bought a product that already worked. They were early on adoption, not early on the invention. AI is the reverse. What shipped is fluency. It writes beautifully. What the valuation needs is reliability, and reliability is a different animal. Five years of scaling made it far more fluent. It didn't make it reliable. A worker who's right 80% of the time can't keep a job that needs 99.9%, and there isn't a chart anywhere showing that gap closing. Last thing, on the cost of being early. Early on the PC cost you the price of a machine that paid for itself in weeks. Early on this means investors front a trillion dollars in data centers, paid back by revenue your own side places 10 to 20 years out. There's a word for being that early: wrong.

This is it. Everyone is excited about AI. I'm not. AI is the biggest commoditizer of human capability we have ever built. Doctors earn more than cleaners because expertise takes a decade to acquire and judgment to apply. AI is going to make most of that expertise available to anyone, at marginal cost. The gap doesn't go to zero. But it shrinks. And the same thing happens in almost every knowledge profession. Programmers. Lawyers. Analysts. Traders. Designers. The taxi driver replaced by a robotaxi is the easy case. The harder ones are people who spent twenty years training to do something a model can now do in a second. The one corner of human life I don't see this hitting is sport. Nobody pays to watch a robot beat a human. The whole appeal is humans competing against each other. That's why nobody watches Deep Blue play chess anymore. From an investing perspective, the next decade plays out in three phases. Phase 1 is the buildout. Hundreds of billions in capex on chips, data centers, and energy. This is what's driving today's market. Phase 2 is the deployment. Productivity rises sharply, then deflation sets in as the cost of knowledge work collapses. Jobs disappear faster than the economy can replace them. Phase 3 is the social adjustment. Capital becomes more powerful than labor. The biggest winners are likely to sit where AI cannot reach. Real assets. Energy. Commodities. Precious metals. Scarce land. And there's something deeper than the economics. The Japanese have a word for it: ikigai. The reason you get up in the morning. Your vocation. Your purpose. The work that makes you feel needed. AI is coming for that too. Nobody has yet figured out what a healthy society does when most of its members no longer feel needed by anything. I might be wrong. I hope I am. If I'm not, the world is about to change faster than most people are ready for.

Horrible analogy there Mr. 'finance' guy. 'Early' is a claim about technology adoption and 'bubble' is a claim about capital allocation. They're answers to different questions, so there's no contradiction and historically the correlation runs the other way. The technologies that were most "early" are exactly the ones that attracted the most premature capital: railways in the 1840s, electrification, fiber in 1999. The fiber was real and got used. The investors who paid 1999 prices for it mostly got wiped out anyway.

"I'm up 680% on Tesla" is supposed to end arguments. Reality check time! Let's grade it! First, the date. Up 680% from today's price means an entry around $48 a share, split-adjusted. That is early 2020: after the Model 3 ramp, after profitability, after the whole market was already reading the story. The IPO believer, in at $17 in 2010, adjusted to $1.13, is up 33,000%. The 680% flex is not early faith. It is a mania entry with good manners. Second, the scoreboard. Same window, early 2020 to today: Bitcoin returned about 800%, no CEO required. Ethereum over 1,000%, no earnings calls. Nvidia over 3,000%. Solana went from a dollar to over a hundred. The 680% is the weakest number on the speculative shelf. In 2020 the Fed flooded the zone and everything with a story ripped. Holding the ticker with the best marketing is a liquidity position, not an insight. Third, the shelf life. Tesla has been roughly flat for five years: about $414 at the late-2021 peak, $372 today. Over the same five years the S&P gained two-thirds and Nvidia multiplied nine times. The 680% exists only because of one exact entry window, which means it documents the tide, not the boat. And here is the part that should bother the believers most. The best performer on that list is the shovel seller. Nvidia sells the chips that the AI landlords rent at a loss. It beat the vision stock four to one, the same way Bitmain won while the miners went bankrupt. The market keeps paying the people who sell the hardware and stiffing the people who buy the story. A great return from a 2020 entry is a receipt. Frame it if you like. Just stop submitting it as evidence.

The more people make these silly analogies the more difficult it is for them to defend their positions.

You've done more due diligence in one sentence than the entire bull case. SpaceX spent $16.44 billion last quarter, 84% of all capex, building AI data centers. The revenue story is that the world will pay hundreds of billions a year for AI output. You just encountered text you suspect is AI output and refused to read it at a price of zero. If the product isn't worth your attention when it's free, the $500 billion ARR projection has a problem, and the problem is you, multiplied by everyone. 😂🤣😂🤣

Every so often someone lands what they think is the zinger: "If you hate Elon so much, why do you pay for the blue check?" Because the check isn't a donation. It's range. This is the platform where the myth gets sold, which makes it the platform where the myth has to be answered. The earnings-call hedges get printed here as headlines. The fan accounts live here. The people being courted as exit liquidity scroll here. If your argument is that a $1.77 trillion story is contradicted by its own checkbook, you make that argument where the shareholders are. As for the money: eight dollars a month. His paper wealth moves more than that every time he walks near a microphone. What he gets from me is a rounding error on a rounding error. What I get is a long-form slot in the middle of his congregation, and I fill it with his own prospectus, his own transcript, and his own capex lines. Call it a civic duty with a cover charge. If one person reads the numbers and decides not to buy the top, the $8 was the best money I ever spent. You're welcome.

The AI compute business has already been run. It was called Bitcoin mining. Same hardware treadmill: a better chip every year, revenue per machine in permanent decline, margins competed down to the cost of electricity. In mining, the winners turned out to be the chip maker and the cheap-power landlords. The fleet operators in the middle ate the depreciation and, at cycle top, the bankruptcy. Riot's own filing shows the shape: cash profit per coin, GAAP loss once you count the machines wearing out. The miners themselves tell you how it ends. Core Scientific went bankrupt mining Bitcoin and now lives as CoreWeave's landlord. IREN, Cipher, Hut 8, the same pivot. The survivors of the last treadmill looked at this one and chose to rent out the power, never the chips. And mining had one escape hatch this business lacks: the coin could pump. Inference only deflates. Every model generation cuts the price of what a GPU produces. There is no halving-day rally for a server rack. SpaceX is spending $16 billion a quarter to stand in the exact spot the miners walked away from.

Fine, grant them the generous case. Say the chips earn for five years, not three. The H200s really are still for rent. Does the math survive? No, because depreciation was never the binding constraint. The revenue curve is. A GPU doesn't have to die to stop paying for itself. It just has to be out-earned, and three clocks run faster than the five-year schedule. First, model efficiency. Every model generation delivers the same capability on a fraction of the compute. DeepSeek repriced the entire stack in a single day in January 2025, and inference cost per unit of capability keeps falling by roughly an order of magnitude a year. Buyers pay for capability, not for your silicon. A chip whose output sells at a tenth of last year's price is spent, whatever the books say. Second, your own supplier. Nvidia ships a new generation about every twelve months, and each one turns last year's fleet into the discount option. You are straight-lining five years of value on an asset that competes with a better version of itself annually. Third, Chinese competition. DeepSeek and Qwen deliver near-frontier capability at a fraction of the inference cost, and Huawei's Ascend line is building domestic supply behind the export wall. Every hardware market China has entered, margins left: solar, batteries, EVs. You'll be too naive not to think compute is next in line. Now look at who pays the rent. The anchor tenants of the entire AI compute market, Anthropic and OpenAI, are both loss-making and both funded by venture rounds. The landlord's revenue is the tenant's expense, and the tenant pays it out of the next round, not out of earnings. Roughly 64% of AI revenue traces to those two labs. Bloomberg keeps a standing tracker of the circular deals: the chipmaker invests in the lab, the lab commits the money to compute, and each side books the other's check as growth. Zitron put it in one sentence: your customers are unprofitable AI startups, and your revenue source for paying off the debt is another company that cannot pay its bills. And if the business were profitable, why are the tenants broke? Because the rent is the loss. Their red is the landlord's green, and the whole building runs on the next funding round. Meanwhile the debt payments are fixed, signed at 8.3%, with the chips themselves as collateral. Revenue per chip decays every quarter. "Still rented" and "still earning back its cost" are different sentences. Chanos already ran the generous math: grant a seven-year hardware life and CoreWeave still earns roughly zero on invested capital. This has run before. In 2000 the fiber builders had backlog too, with vendor-financed customers to fill it. When the funding stopped, the revenue turned out to have been the funding all along. Grant the five years. The books get prettier. The cash still doesn't come back.

Want to know what Musk believes in this week? Same test as always. Check the checkbook. Two wires just crossed. Musk reportedly put $1 billion of his own money into David Ellison's Skydance, the media group now running CBS, CNN, HBO and Warner Bros. And SpaceX is finalizing a $40 billion debt raise, led by Apollo and Goldman, at around 8.3% on the riskiest slice, more than a point above where its existing bonds trade. The collateral on the safer slice: AI chips. He has personal cash for a media conglomerate. The moon city gets other people's money at 8.3%, secured by server hardware with a three-year honest life. If the vision were worth believing, the chief believer would be its investor. Instead the vision pays a premium rate to strangers and the Ellisons get the cash. Forty billion at these rates is over $3 billion a year of interest. That is the price of conviction purchased with someone else's money. Money only goes where belief is. His went to Hollywood.

This is what Elon said in January 2025: "Weʼre making good progress in Optimus. We expect to have thousands of of Optimus robots working in Tesla factories by the end of this year [2025], doing this more. And we expect to scale Optimus up faster than any product." This is what Elon said yesterday: "People don’t understand new technology that well, or volume production of complex new hardware that well, it sort of seems like, well, can’t you just make a whole bunch of these things immediately and send them out" Amazing, that it's Elon who lectures the listeners on production while it was him ridiculously overestimating Tesla's ability to bring Optimus to production. "I do expect these issues to be solved around the end of next year. [2027]" Every sceptical voice is immediately shot down with "ooh, what do you know about manufacturing. Haha, smarter-than-Elon syndrome" or some other nonsense. Meanwhile, Elon makes one awful prediction after the other. It's ridiculous. We really do live in a Dilbert comic.

The most common reply to this post has been some version of: the AI data centers will fund the moon city. I want to thank the business geniuses in my mentions for cracking that one open. Their plan has three steps. Build AI compute. Make money. Build the moon. The only weak link is step two. SpaceX's own prospectus shows the AI segment losing $6.355 billion in 2025 and another $2.469 billion in Q1 2026, and it says in plain language the business stays unprofitable. But sure, geniuses. A funding source. Want a preview of step two? Meet CoreWeave. It does nothing but AI hosting, at scale, with a $99 billion backlog of real contracts, cheap capital, and the fastest deployment story in the industry. It loses money every quarter. Cash losses projected through 2028. Chanos ran the return on invested capital at roughly zero, once the GPUs are depreciated on the schedule the hardware actually lives on, two to three years, not the five to six the books pretend. Even the market did the math once. The stock fell 15% when the company announced more capex. So the business genius pipeline looks like this. A money-losing segment funds a program with zero line items, using profits that don't exist yet, generated by buyers who can't pay their bills without the chip vendor's help. That's the plan. And the best part is the delivery: all identical, all confident, all from self-described CEOs whose companies, I notice, have never had to fund a moon base either. 😂🤣😂 The moon city was always going to be funded by something. Now we know what. Vibes, with a depreciation schedule.