Three thousand to thirty-six hundred degrees Fahrenheit. That's the operating temperature inside the hottest section of a gas turbine — roughly 800 degrees hotter than the melting point of the metal alloy the blades themselves are made from.
The only reason those blades don't liquefy is that each one is grown as a single, unbroken crystal, slowly, inside a vacuum furnace — no grain boundaries, no microscopic seams where ordinary cast metal would crack under stress. Add internal cooling channels and a thermal-barrier coating, and you get an airfoil that survives conditions that should destroy it.
Exactly four companies in the world can do this at industrial scale. All four are tapped out.
Exactly four companies in the world can grow these single-crystal blades at industrial scale. All four are tapped out. So the fastest way to get one is to become the fifth.
So the fastest way to get one, apparently, is to become the fifth. That is what SpaceX has decided, and it is the most important thing to happen in the AI power trade this year.
On August 30, after The Information found job listings for a "blades and vanes foundry" and confirmed SpaceX had quietly bought roughly 830 acres near its Starlink factory in Bastrop, Texas, Musk confirmed it on X. Natural gas will still be needed to bootstrap solar for years, he wrote, and "the limiting factor for nat gas turbine production is casting the blades & vanes." Doing it in-house, he claimed, accelerates turbines coming online "by up to 18 months."
Sit with the strangeness of that. A rocket company believes it can master single-crystal turbine metallurgy — faster — than the four incumbents who already own the process.
The reason the claim is even plausible is the queue math. GE Vernova, the market leader, is sold out of large-turbine capacity through 2030.
Michael Thomas of Cleanview put a finer point on it: an order placed today gets pushed to the back of the line, meaning delivery somewhere between 2031 and 2033 — and then another one to two years to install. Call it a decade from purchase order to first electron.
That is why "just buy your way in" is not a strategy. The scarce thing isn't turbines you can order; it's the manufacturing capacity that makes them, and no check clears fast enough to conjure a furnace that doesn't exist.
Look at what that does to a project on the ground. SB Energy, the SoftBank affiliate, decided to build a 9.2-gigawatt, $33 billion gas plant to feed OpenAI's data center in Pike County, Ohio.
The financing is not the problem — that's a separate thread. The problem is that a $33 billion plant is worth nothing without turbines, and the turbines are the one thing money can't accelerate through the queue.
So the bottleneck has a shape. It is an airfoil, and it stands between OpenAI's Ohio site and the power it was promised.
This reframes what capital has been doing all summer. The instinct is to read every AI-adjacent bet as a bet on models. It isn't. Increasingly it's a bet on whoever owns the pipes, the wires, and the fabrication.
Take Thiel Macro's $418.7 million reappearance in the 13F filings — a fund that had sat blank for two quarters, then came back with seven of eight positions in power, utility, and oil names. Skip the "contrarian returns" narrative; it rounds to about 1.3% of his balance sheet.
What matters is the map. Four regulated utilities — AEP, DTE, FirstEnergy, CMS — make up about 39% of the sleeve. Those companies don't profit from expensive electricity; they earn a regulated return on the plants and wires they build. It's a bet on construction, on physical buildout, not on power prices.
Now look at Comfort Systems USA — a mechanical and electrical contractor, about as far from an AI narrative as a business gets — up fivefold on data-center demand, blowing past even bullish models. The reason it compounds is a moat almost nobody was watching: surety-bonding capacity, the boring balance-sheet ability to guarantee you'll finish the job.
Utilities, contractors, and now a foundry are the same trade wearing three costumes. Capital has figured out that the constraint moved downstream of the chip, into the physical layer, and it is repricing everyone who can pour, weld, wire, and cast.
Which brings us back to the uncomfortable asymmetry. A company with zero power-equipment track record is credibly "faster" than a four-firm oligopoly — and the reason is that the incumbents optimized for the wrong thing.
An order book sold out through 2030 is a beautiful asset if your goal is pricing power and utilization. It is a liability if the goal is speed. GE Vernova's scarcity is engineered; the queue is the product. Nobody in that oligopoly is incentivized to blow out capacity for a demand spike they suspect is a bubble.
SpaceX is playing a different game because it can. Adam Jonas of Morgan Stanley made the load-bearing point: Raptor turbopump castings and data-center turbine airfoils share the same alloys, the same vacuum furnaces, the same scarce skilled labor.
That's the hedge hiding inside the bet. The furnaces don't have to be justified by AI power alone — they cast rocket-engine components either way. SpaceX can spread the cost across space and AI, which means it can build capacity that would look irrational to a pure-play turbine maker staring at demand it doesn't trust.
So the trade underneath the trade is this: everyone else is writing checks into a queue, and SpaceX is underwriting the risk with rockets. If the AI power boom holds, it owns the scarcest capability in the stack. If it cracks, the furnaces still make Raptors.
None of this is free, and none of it is clean. In Memphis, the NAACP has repeatedly accused SpaceX of running turbines at its Colossus data centers without required permits or pollution controls.
In Virginia's Data Center Alley, a Piedmont Environmental Council study ran the EPA's own COBRA model on a single facility's eight gas turbines and found emissions reaching more than 2.5 million people — an estimated 3.4 to 6.5 additional premature deaths a year, and $53 million to $99 million in annual health damages. Faster turbines mean faster combustion, and the exhaust lands hardest on communities that were already breathing the worst air.
But hold the object in your hand one more time. A blade grown atom by atom as one flawless crystal, in a vacuum, hotter than the metal it's made of — that is now the gate between a hyperscaler and its electrons.
Not the GPU. The casting furnace.
The tell isn't that SpaceX wants turbine blades. The tell is that it thinks vertical manufacturing capability — the thing you can only build, never buy — is worth mastering from zero, and that the same furnaces pay for themselves twice.
Capital is already pricing that in. Watch the foundries, not the checks.