The AI Spice Must Flow

The robot walked. Now find the magnet.

This week a humanoid robot named IRON walked off the world’s first automotive-grade humanoid production line, in Guangzhou, on its own two feet. Seventy-six joints, 2,250 TOPS of compute in its chest, built by XPENG — a car company, which is the tell we’ll come back to. Elon says he’ll build ten billion of them. Figure, the American one, is uploading thirty-five minutes of training data every second and just committed to a hundred thousand GPUs. Every forecaster who matters agrees the humanoid boom is coming; they only argue about whether it’s a million robots a year or a billion.

And the whole industry is still arguing about the wrong thing. It’s arguing about whether the robots are smart enough. That question is basically settled and getting cheaper by the week — the intelligence is downloadable, and China’s open models are distilled copies of ours at a tenth of the price. What actually moves a robot isn’t the model. It’s the motor. Each of IRON’s seventy-six joints needs a high-torque electric motor, and a high-torque motor needs a permanent magnet made of neodymium, iron and boron, doped with a pinch of a heavy rare earth — dysprosium or terbium — so it keeps its magnetism when the joint heats up under load. No heavy rare earth, no heat tolerance. No heat tolerance, no actuator. No actuator, no robot. A humanoid carries somewhere between two and four and a half kilos of that magnet and thirty-five to eighty grams of dysprosium. Multiply by a few billion robots and you can see the shape of the problem before you’ve done any of the math.

Read it as a ladder

The single number is 94%, but the honest number is a ladder, and the ladder is the whole argument. China mines about 60% of the world’s magnet rare earths. It refines 91%. It manufactures 94% of the finished magnets. The grip tightens the farther you get from the dirt — which is the opposite of how most people picture it. When you hear “China controls rare earths,” you picture a country sitting on the only deposits, the way Saudi Arabia sits on oil. That picture is wrong.

We have the rocks. The United States has them at Mountain Pass in California. Australia has them, Brazil has them, India has them, and Canada — a country we are currently, insanely, treating like a hostile trading partner — is loaded with them. Deposits are not the scarce thing. Reserves are not the scarce thing. The scarce thing is the midstream: the separation chemistry that pulls one rare earth away from the fifteen others it’s chemically almost identical to, and the sintering that turns refined oxide into a finished magnet. That’s the part that’s hard, dirty, capital-hungry and low-margin. That’s the part China built while the West decided it was beneath us. And it’s worse than the averages look, because the heavy rare earths that give a magnet its heat tolerance — the dysprosium and terbium the actuator actually needs — are far more concentrated in the Chinese sphere than the light ones. Today’s electric-vehicle fleet already eats roughly a third of the world’s heavy-rare-earth supply. The EV industry alone, right now, is the heavy-magnet demand of about twenty-six million humanoid robots. The robots haven’t even shown up yet and the pipe is already most of the way full.

We shut it ourselves

Here’s the part that should sting, because it’s not a story about Chinese cunning. It’s a story about American choice. Mountain Pass — the California mine that has the rocks — went bankrupt last decade. Not because it ran out of ore. Because China flooded the market, drove the price of the finished product below the cost of Western production, and waited. We let the mine die, and for years afterward the ore we did dig came out of California, got put on a boat to China to be refined, and came back as magnets with a Chinese markup. We had the mine and still couldn’t make the magnet, because we’d let the middle of the chain rot.

That’s the Dune move, and almost nobody clocked it while it was happening. In Herbert’s books the spice isn’t valuable because it’s rare in the ground; it’s valuable because it’s the one substance that makes interstellar travel possible, and one power controls the supply. No spice, and you’re stuck on your homeworld, no matter how advanced you are. That’s exactly the trap. Bits that only flow to the screen in your hand keep you on the homeworld — a better chatbot doesn’t move society an inch. The things that actually move it off the homeworld are physical: cars that drive themselves, rockets that land and fly again, robots that take a shift of brutal labor off a human back. Every one of those is an atoms problem, not a bits problem, and the atoms run through a refinery in Inner Mongolia.

So widen the lens, because the magnet is just the sharpest edge of it. We won the bits — American labs are ahead, and the Chinese models are knockoffs of our work. But the bits were the easy part, because software scales for free and copies itself. Turning that intelligence into anything that moves is an atoms problem: it needs power, and a grid that can carry it, and data centers, and refineries, and magnets, and the mines under all of it. We are, at this moment, world-class at the one input that’s nearly free to reproduce and dead last at the ones that aren’t. That’s not a technology gap. It’s a will gap.

The clock

And the will gap has a date on it now, which is what turns this from a think-piece into a trade. Last October, China put dysprosium, terbium and the magnets made from them under export controls. Then, as part of a broader thaw, it suspended the U.S.-facing piece of those controls — a pause, not a repeal, that runs out on November 27, 2026. Dysprosium and terbium never came off China’s dual-use control list; the architecture is fully intact and switched to standby. Analysts who read this for a living expect the screws to turn again the moment the diplomacy sours.

Read what that pause actually is. It’s Beijing demonstrating, in the politest possible way, that it can throttle the raw material of the entire Western robotics and defense industrial base with a press release, and choosing — for now — not to. In Dune, the lesson Paul learns is that the power to destroy a thing is the power to control it. The export-control list is that power written into customs law. The pause is the reminder. And it expires right about the time every forecast has the robot build ramping into real volume. You do not want to discover on November 28th that your entire actuator supply was a Chinese policy decision the whole time.

The answer is a build order

So what do you actually do about it, beyond worry? You build the midstream. And the good news — the genuinely American news — is that this is finally starting, because the demand got big enough to overpower twenty years of “let someone else make the dirty stuff.”

We called the shape of this a week ago, in Field of Dreams: coal mining isn’t coming back to West Virginia, but data mining might, and the roughly trillion-dollar data-center buildout is the reindustrialization of America whether we call it that or not. The magnet is the next rung, and it’s a heavier one. The data center was phase one — server halls, power, fiber. Phase two is the real thing: mines, separation plants, magnet lines, the parts of an industrial economy we spent a generation offshoring. This is Anduril’s move, generalized. Anduril didn’t ask permission to join the defense-primes club; a new-guard, vertically integrated company walked in and rebuilt how the business works. Someone has to do that to American mining and refining — treat it like a technology company’s problem instead of a legacy industry’s — and the demand from a few billion robots is finally large enough to underwrite the reopening.

Somebody’s writing the checks. The Pentagon, in July, made itself the largest shareholder of MP Materials — a $400 million preferred stake — set a price floor of $110 a kilo on the neodymium-praseodymium it produces, and guaranteed to buy 100% of the magnets from a new $1.25 billion plant in Texas that will add ten thousand tonnes a year of capacity when it comes online in 2028. Lynas is building a heavy-rare-earth refinery in Texas. USA Rare Earth is breaking into the separation business China has owned outright. The IEA puts the total bill to build a real magnet supply chain outside China at about $60 billion over the decade, roughly half of it in refining. And this is the same thing Sequoia was pointing at in August when it made the biggest bet in its fifty-four-year history — ten billion dollars into “AI plus reindustrialization,” which is to say robots, defense, power and mines, and pointedly not another model. The smartest money already moved off the bits and onto the atoms.

Notice what none of that is. None of it is a free market. You do not floor a price and guarantee a purchase for something that pencils out on its own. The government is subsidizing a strategic industry back into existence because the alternative is being unable to build a missile — or a robot — without asking a rival for permission. And after COVID, that argument shouldn’t need making. We learned, at real cost, that “the cheapest supplier” and “the supplier you can count on in a crisis” are not the same company, and that finding out the difference during the emergency is the most expensive way to learn it. Masks, then chips, now magnets. This is not a partisan point, and the piece falls apart the second it becomes one. It’s the rare fight a populist and a union Democrat should both want to win, and mostly neither has bothered to. The rocks became a jobs-and-security question, and we outsourced it on a spreadsheet because the spreadsheet said it was cheaper.

The honest part

Now the counterweight, because I don’t want to sell you a clean story — clean stories are how you lose money. Reopening the midstream is the easy sentence and the hard decade. Separation chemistry and magnet sintering are genuinely difficult, filthy, and slow to stand up; a mine permit is nothing next to a refinery that actually holds environmental muster and hits spec. China has price-warred Western supply into bankruptcy before — that’s literally how Mountain Pass died the first time — and it can do it again the moment the reshored plants need real prices to survive. That’s the whole reason the DoD had to floor the price: private capital won’t underwrite a ten-year, capital-brutal build on a political promise that a trade détente or the next administration can gut. Which is the deepest tension in the entire story. Political will is exactly what’s required, and political will is the least durable input on the list. The November 27 pause is the proof — the whole posture can flip on a handshake.

And keep one skeptic in the room on the demand side too. Ben Evans made the point this week that we’re terrible at judging robot progress: a machine that does a backflip is easy, a machine that can pour a cup of coffee is hard, and a machine that can walk into a strange kitchen and figure out how to make the coffee is a bet on real general intelligence we haven’t won yet. If he’s right and the humanoid ramp is slower than the hype, the magnet crunch is slower too. That cuts both ways. It buys us time to build the midstream — or it deflates the urgency that’s the only thing funding the build in the first place. I lean toward building anyway, because the defense demand is real regardless of whether your house ever has a robot in it, and because being early on strategic capacity is a rounding error next to being late. But I’m telling you where the argument is soft, because anyone selling you the hard version is selling you something.

The spice must flow

Here’s where it lands. For two years the entire conversation has been about intelligence — whose model is smartest, whose benchmark is highest, when AGI arrives. And while we argued about the mind, the thing that turns the mind into muscle quietly became a single-country dependency with a policy switch on it. The robot walked off the line in Guangzhou this week. The magnet inside it, and inside the American one, and inside the missile meant to deter both, comes from the same refineries. That’s not a technology problem. We solved the technology. It’s a will problem, and will is a choice.

So do the three things that are actually yours to do. Trace your own exposure to the magnet — if your product or your fleet or your roadmap touches a motor, find out today where the magnets come from, before November 27 finds out for you. Price the midstream, not the miners, because the ore was never scarce and the refining always was; follow the guaranteed buyer. And stop booking resilience as a cost line — the input you outsourced purely because it was cheaper is a hostage position wearing a margin’s clothing, and COVID already sent you that invoice once.

In Dune the line is that the spice must flow, and the terror underneath it is that someone else decides whether it does. We spent a generation handing that switch to a rival because the switch was boring and the margins were thin. The robots are here to collect on that decision.

You can’t build a robot — or a rocket, or a missile — on your rival’s permission slip.

Sources:

• IEA — Rare Earth Elements: executive summary (China 60% mining / 91% refining / 94% magnets, up from ~50% in 2005; ~$60B to diversify, ~half refining)
• The Robot Boom Has a Magnet Problem — Foundation for American Innovation (Sept 8, 2026) (2–4.5 kg NdFeB and 35–79 g Dy per robot; EVs already ~⅓ of heavy-REE supply = ~26M robots)
• Clark Hill — China hits pause on rare-earth export controls (Oct 9 2025 controls; US-facing suspension expires Nov 27 2026; Dy/Tb remain on the dual-use list)
• Manufacturing Dive — MP Materials’ Texas magnet campus + DoD deal ($1.25B plant, +10,000 t/yr, DoD $400M/largest shareholder, $110/kg NdPr floor, 100% offtake, online 2028)
• XPENG press release — IRON humanoid production line commissioned (Sept 8, 2026) (76 DOF, 2,250 TOPS, mass production by year-end; robotics arm raised ~$900M at ~$6.3B in Aug)
• Brett Adcock / Figure (X, Sept 8, 2026) (35 min of training data/second; 100k GPUs on Vera Rubin; “you can’t buy the hardware engineering with unlimited money”)
• China Is Turning Its Humanoid Robot Boom Toward the Battlefield — Reuters review (Sept 8, 2026) (~95% of global humanoid shipments; PLA Daily pushing robot “combatants”)
• Benedict Evans, weekly newsletter No. 659 (Sept 8, 2026) (the humanoid-progress skepticism: backflip vs. pouring coffee vs. making coffee in a strange kitchen)
• CO/AI, Field of Dreams (Sept 1, 2026) — getcoai.com/news-letter/field-of-dreams (the ~$1.1T data-center capex cycle as US reindustrialization; “data mining, not coal mining”)
• CO/AI, Know Thyself (Aug 7, 2026) (Sequoia’s $10B bet on “AI + reindustrialization” — robots, defense, power, mines)
• Energy Intelligence — The Monroe Doctrine, Reimagined by Trump (Western-hemisphere critical-minerals realignment; South America lithium/rare earths)