{
  "$type": "site.standard.document",
  "content": {
    "$type": "blog.pckt.content",
    "blob": {
      "$type": "blob",
      "ref": {
        "$link": "bafkreiesu374zqcmdbhszrg4edxydbt4wbwjzr6zr7oqr2h4hyy6uinxzy"
      },
      "mimeType": "text/plain",
      "size": 29282
    },
    "references": [
      {
        "$type": "blob",
        "ref": {
          "$link": "bafkreichwcfir5gh6f3ndaticpespuvv44iwxx3xo3n57rzcictexwx6xa"
        },
        "mimeType": "image/jpeg",
        "size": 48514
      }
    ]
  },
  "coverImage": {
    "$type": "blob",
    "ref": {
      "$link": "bafkreichwcfir5gh6f3ndaticpespuvv44iwxx3xo3n57rzcictexwx6xa"
    },
    "mimeType": "image/jpeg",
    "size": 48514
  },
  "description": "I need to talk about SpaceX’s IPO, the shaky foundations it’s built on, the fraud it’s being used to commit, and the consequences for you and your grandparents. Two cases to make: the engineering can’t be built on schedule, and the economics double-counts itself. Let me start with “orbital data centers,” a piece of their S-1 filing, alongside Anthropic’s recent announcements. Critics usually attack orbital data centers on launch physics. That’s the weakest argument. The energy to lift a data cen...",
  "path": "/elon-musk-is-stealing-your-retirement-sb3bz85",
  "publishedAt": "2026-06-04T13:42:55+00:00",
  "site": "at://did:plc:z3wf4rvs3j5eglznzbnolnom/site.standard.publication/3mjtvix2naghm",
  "tags": [],
  "textContent": "I need to talk about SpaceX’s IPO, the shaky foundations it’s built on, the fraud it’s being used to commit, and the consequences for you and your grandparents. Two cases to make: the engineering can’t be built on schedule, and the economics double-counts itself. Let me start with “orbital data centers,” a piece of their S-1 filing, alongside Anthropic’s recent announcements.\nCritics usually attack orbital data centers on launch physics. That’s the weakest argument. The energy to lift a data center into orbit is trivial. Low Earth orbit costs about 33 megajoules per kilogram, and a multi-gigawatt facility pays that back against its own output in days. If you wanted to kill this idea with thermodynamics, launch energy is the first place you’d reach, and it’s the one place where the idea survives.\nThe real problem is worse, and it’s the kind physics won’t save you from. Every hard part of this needs the same thing, and there isn’t enough of it.\nThe heat has nowhere to go\nA data center is a machine for turning electricity into heat. On the ground you blow air over it or run water through it. In space there's no air and no water. You can only radiate heat away, and radiation is feeble. A surface no hotter than your chips can tolerate sheds a few hundred watts per square meter. To dump a gigawatt of waste heat, you need around four million square meters of radiator.\nYou could shrink that by running the whole thing hotter, but hotter means chips that tolerate temperatures normal chips can’t. So now your radiator breakthrough is also a chip breakthrough, and they have to arrive together. A hot radiator bolted to a chip that quits at ninety degrees is scrap.\nThat's the first time two problems turn out to be one problem. It won't be the last.\nOrbit cooks the chips\nAnother thing the renderings leave out: chase the one real reason to go to space, sunlight all the time, no night, no clouds, and you climb into the radiation belts, which slowly kill your hardware. A lower dawn-dusk orbit gets you most of the sun without the worst of the dose, but you still pay in cosmic rays, solar storms, and the South Atlantic Anomaly. The free lunch isn’t there.\nRadiation breaks chips three ways, and optimists only talk about the gentlest one:\nA particle hits a chip and flips a bit. The math comes out wrong, but the chip is fine. That’s recoverable; software has handled flaky hardware for decades. It’s the convenient one to handwave with.\nThey’re not gentle at all. Radiation builds up in the chip over months until it drifts out of spec and dies, on a clock you can’t stop. And sometimes a single particle triggers a short that draws enough current to kill the chip outright. Software cannot resurrect a dead chip.\nYou can armor a chip against this. Every spacecraft has had computers in it, after all. The problem is that you do it with bigger, older transistors that hold more charge and shrug off more dose. Which means a radiation-hardened chip is years behind the best chip you can buy. Again, this is fine for the computer flying a spacecraft, but useless for a machine built to run the most advanced AI on Earth. You cannot want the best chip made and accept one five years old.\nModern AI chips are the worst possible shape for this. They’re huge, packed with tiny features, and run incredibly hot. Everything that makes them good at the job makes them fragile in the place you want to put them.\nRepair is a launch program\nThe answer to all this is supposed to be maintenance. Hardware dies, you swap it, the way every data center on the ground already does. Pull the dead board, slot in a new one.\nTwo problems: nobody can do it, and even if they could, the numbers don't work.\nThe \"nobody can do it\" part: every time anyone has serviced something in orbit, it's been a tug nudging a satellite or topping off its fuel. Docking and a shove, a little love-kiss. Opening up a packed compute rack in vacuum, pulling a dead board out of a running system, and seating a live one, with no human anywhere near it, has essentially never been done. It's a research problem. You can't buy it and you can't assume it.\nThe numbers don’t care whether the robot shows up.\nA data center on the ground loses a few percent of its hardware a year. Put that hardware in a high-radiation orbit and the rate climbs, because now the dose is killing parts on a schedule and the random shorts are picking off the rest. Be generous and call it five percent a year.\nA multi-gigawatt orbital data center is plausibly around seventy thousand tonnes. That’s an assumption, not a published number, but it’s in the ballpark of what Starcloud’s own materials imply. Five percent of seventy thousand tonnes is thirty-five hundred tonnes of dead hardware every year. At about a hundred tonnes a launch, that's thirty-five launches a year just to replace what died. Just to stand still.\nThe current FAA cap for Starbase is twenty-five launches a year.\nRead that twice. The repair flights alone, before a single new server, before Starlink, before anything else SpaceX actually flies for, blow past the entire licensed launch capacity of its main site. And five percent is the kind number. Nobody has flown a dense pack of modern chips in a harsh orbit to measure the real rate. If it’s ten or twenty percent, the resupply alone is a national launch program, running forever, with no end, and that doesn’t even consider the lifespan of a chip. You’re replacing them non-stop until the whole data center has to be replaced.\nThere's a clever way out, and it's the best card the other side holds, so let’s engage with it seriously: don’t repair anything. Treat every chip as disposable, the way the big cloud companies already treat cheap servers. When one dies, route around it in software and leave the corpse in the rack. No robot needed. The system degrades gracefully and you never fly a repair mission.\nThis works right up until it doesn't. Routing around dead chips is free until enough of them are dead that you're running a graveyard with a few lights still on. Then you have to add capacity, and adding capacity means launching it. This doesn’t avoid the resupply program; it just renames it. “Repair flight” becomes “expansion flight,” and the rocket goes up either way.\nWhich is where this has been heading the whole time.\nIt all comes down to launch\nUnder all three problems is the same rocket. Every tonne of radiator, every replacement chip, every \"expansion\" flight or repair or whatever you want to call it, all of it goes up the same way, on the same vehicle, against the same hard limit. On a slide the problems might look separate, but they aren’t. They're three withdrawals from one account, and the account is launch.\nThat's why the physics objection misses. Physics is fine with this. Launch energy is cheap. What’s not fine: you need three or four separate tenfold breakthroughs that have to land at the same time and not undercut each other when they do.\nAnd here's the part that should end the argument:\nSay you get them all: Cheaper launch, lighter radiators, hot-running radiation-proof chips, robots that fix things in vacuum. Every one of those makes a data center in the Nevada desert better too. Cheaper chips are cheaper on the ground. Better cooling is better on the ground. Hardware that runs hot and lasts a decade untouched is exactly what you want in a building you can walk into.\nThe space version needs all of it just to get to the starting line. The desert version is already there. And the desert version has air, water, gravity, and a door.\nThe money runs out before the miracles arrive\nEverything above assumes the rockets keep flying. Not at today’s pace, but faster, and for decades, a lot of it on repair flights that earn nothing, because replacing dead hardware is a cost, not a sale.\nAll of this rests on one company. There's no competitor that can move this kind of mass. This is the only supplier of the whole chain, not just a market wobble that competitors route around.\nThere’s good news, though: the breakthroughs don’t all die together. The split is where it gets interesting. The ones that pay for themselves today are safe: reuse, cheaper launches, Starlink with its ten million paying subscribers. Those are real businesses; they survive any restructuring, because they’re the assets everyone wants to keep. The launch that serves Starlink keeps flying.\nWhat dies is the part revenue never paid for. The market is meant to fund the jump to hundreds or thousands of launches a year out of the valuation, off its willingness to keep feeding the promise of what comes next. That's the first thing cut when belief runs out, because it was the most speculative thing on the books. The repair robots, barely funded now, go first. The orbital compute buildout goes with the segment that promised it.\nAnd the one breakthrough that keeps moving no matter what, better chips, was always driven by demand on the ground, not by anything in orbit. The point repeats itself quietly: chips would have improved for the desert with or without anyone in space.\nSo a collapse doesn’t reset the tech to zero; it freezes the curve at whatever was already paid for. The part the space data center actually needs is the part that stops.\nTwo years for three breakthroughs\nHere's where the filing stops being vague.\nStarship starts carrying real payloads in the second half of 2026. That’s right now (or at least, it is at the time of writing). The next-generation Starlink satellites follow. And solar-powered data centers go into orbit as early as 2028.\nSo the gap between \"rocket carries its first real payload\" and \"data centers operating in space\" is about two years on paper. Two years for three breakthroughs.\nYou don’t need to assume my skeptical position to question this; just look at SpaceX’s own track record. Last year SpaceX set a target of twenty-five Starship flights and flew five. Missed it by five times over. Every new vehicle they’ve built arrived late: the booster catch by months, the first orbit by nearly two years, the heavy-lift rocket before it by five. One analyst's rule of thumb for reading any Musk hardware date is to multiply it by two and a half. That's not a fact, but it is drawn from a long list of slipped dates, and not much else has predicted them better. Run the 2028 data center through it and you land somewhere past 2033.\nAnd the hardware they need doesn’t exist yet. The filing leans on chips that don’t exist at performance the filing concedes is years away. By any honest read of the silicon roadmap, the gap between today’s best chip and the chip the plan requires is four to six process generations. Closed by 2028. While the rockets that carry it are still in test flights and currently grounded.\nThe filing calls all of this uncertain and unproven. I’ll be clear: I call that delusional. Magical thinking.\nWhat they're actually selling.\nLook at what they actually told the market it’s selling.\nSpaceX says its total addressable market (TAM) is $28.5 trillion. Roughly the annual output of the United States. Space is $370 billion of that. Connectivity, the Starlink business that funds everything, is $1.6 trillion. The rest, about 93 percent, is AI. Inside the AI number, the biggest single piece, $22.7 trillion, is labeled “enterprise applications.”\nThat last number isn’t a guess at a software market; it’s someone’s estimate of the entire global digital economy, adopted whole. To \"address\" it, the AI agents have to be doing the work that economy currently runs on. The vehicles are named in the filing: an agentic platform built with Tesla, and the enterprise version of Grok. Strip the language and the pitch is that software agents take over a large share of white-collar work, and the company collects the value that work used to produce.\nHere's the problem. The compute that's supposed to do all that white-collar work is, right now, mostly earning its living as a rental. xAI's biggest AI customer isn't an enterprise buying agents. It’s a competitor, renting raw capacity, more than four times what the whole AI division earned the year before.\nYou'll hear that the training cluster simply moved on, that they rented the old one because the new one took over. That's true and it's not the point. The old cluster isn't idle. It's running inference, at scale, profitably. Just not theirs. They rented it to someone who had the demand to fill it, because their own product didn't. By their own numbers, they used about a tenth of the chips they owned. The rest sat waiting for customers who weren't coming fast enough, so they leased the building to a tenant who had a line out the door.\nNow pick either reason that's true, because both of them break the same number.\nIf they rented because they couldn't fill that compute with their own agents, the enterprise pitch is already not working, at the exact moment they're valuing it at twenty-two trillion dollars.\nIf they rented because compute is just floor space and the real money is in whoever's models the enterprises actually pick, then the twenty-two trillion belongs to whoever owns those models, and nothing in the filing shows that's them.\nEither the product isn't landing, or the value lives one floor up from where they're standing. There's no third door. They're selling the destination. The revenue is a landlord's.\nThat’s not wrong or bad per se, but a landlord's market isn't twenty-two trillion dollars. Renting computers is a real business. It is not the global economy. The moment you admit the money is in the rent, you've admitted the number on the cover is the wrong number, off by at least an order of magnitude.\nThat's the bet the price is built on. Not rockets. Not broadband. A wager that AI does the office work, and the orbital data centers are the story for where the compute to run it comes from. Every problem in this essay is a problem with that wager. The physics is a fun distraction for nerds like me. A little treat.\nSo, you have at least two problems, stacked.\nThe first is that the thing can't be built. Not on this schedule, not at this cadence, not for these losses. The rockets don't fly often enough, the orbit cooks the hardware, and the plan needs four miracles to show up at once and in order.\nThe second is worse, because it's true even if the first one isn't.\nSay they build it. Say the compute shows up and the agents work and they do the office work. The pitch says that's a $22.7 trillion market. But that number is the size of the digital economy, and the digital economy is people buying things. Agents don't buy things. The people whose jobs the agents took were the ones buying things. You can sell the labor savings or you can keep the customers, but you can't do both with the same dollar. The paycheck you delete was somebody's grocery money.\nThe filing books the savings and keeps the customers. It counts the dollar twice.\nThe usual answer is that this always sorts itself out. Machines took the old jobs, new jobs appeared, the people moved. Sometimes that's true (although at a steep cost). Whether it is this time turns on three things: whether new jobs appear as fast as the old ones go, whether the savings flow back out as wages instead of pooling at the top, and whether the new work pays enough to put the spending back.\nThere's a real argument that all three land badly when the machine comes for thinking and not just lifting, because there's less higher ground to run to. That's contested. I lean one way on it. I'm not going to pretend it's a settled debate though.\nBut the filing doesn't even hold the debate. It books the upside of replacing the workers and leaves the cost of replacing the workers off the page entirely. A $22.7 trillion market with no line for \"what happens to the people who used to earn that money\" is something very simple: a plan to extract from a thing while hoping the thing sits still and lets you.\nThe bill goes to you.\nHere’s where it all comes crashing down.\nA company this size doesn't get bought the way a normal stock does. At nearly two trillion dollars, it’s instantly one of the biggest things in the market, which means every index fund has to hold it. Not because anyone at those funds judged it worthy, but because the index contains it and the fund’s job is to mirror the index. The buying happens by rule.\nThat forced buying is a door, and the people who understand the company best are standing closest to it. The filing carves out a friends-and-family block exempt from the usual lockup, so the people closest to the engineering can walk before anyone else. And the dual-class share structure means that whenever insiders do sell, control stays where it was. So the people who know the most can walk out through the door the index holds open, while the position lands in passive funds that aren't allowed to have an opinion and in the accounts of people who were never shown the engineering.\nFollow what's being handed off. The bet is that agents do the office work. The infrastructure story under it can't be built on schedule. The market math counts the worker's wage as profit and the worker's spending as intact, at the same time. And the mechanics of an IPO this size pour that bet into the boring retirement account of a teacher who’s never even heard of it.\nIf the bet is what I think it is, the people left holding it are the same people the bet was built to replace. They lose the job. Then they lose the savings.\nI don't know that the price falls. Nobody does; it hasn't traded. Starlink is real and the government needs the rockets, and that might hold the thing up longer and higher than it should. That part is my read, and I'll own it as just that: a read. But the shape of the trade doesn't depend on timing. The downside, whenever it comes, lands on whoever was holding when the rule-bound buying stopped. That’s a wealth transfer, straight up. From the people who understood the risk to the people who structurally couldn't.\nA solar farm in the desert was never going to sit in your pension as a two-trillion-dollar bet on dumping heat into vacuum and deleting your own customers. This is.\nThat's the whole case. Not that it breaks the laws of physics. It doesn't. It's that it needs a stack of miracles to arrive together, resting on a math that cancels itself out, sold at a price only its size makes possible, and that size is exactly what forces it into the hands of people who never got a vote.\nYou can believe all of that lines up. Just don't call it a plan. The filing has a word for it. It's in the risk factors. They put it there for a reason.",
  "title": "Elon Musk is stealing your retirement.",
  "updatedAt": "2026-06-04T14:29:29+00:00"
}