Why SpaceX Is Learning How Metal Freezes

Picture a data center that is ready to go. The building is finished. The computers are installed. Outside sits the gas turbine meant to power them. It is complete except for the parts at its hottest end, and those parts have not been made yet.

Nothing runs until they are. The building and the computers sit idle, waiting on curved pieces of metal that a foundry has not finished pouring.

Farzad calls this the last part problem. A nearly finished system still waits on its last essential part, and that part decides when the whole investment starts working. That is why a rocket company now cares a great deal about how metal freezes.

Why a rocket company needs turbines

SpaceX's CFO says the company is on track for about $100 billion in annual recurring revenue at the end of 2026. Elon's projection is somewhere between $500 billion and $1 trillion at the end of 2027.

Computing is a big part of that story. SpaceX runs huge AI data centers called Colossus in Memphis and Southaven, and gas turbines on site help power them. It had 1.7 gigawatts of computing capacity at the end of September 2026, and Elon is targeting 13 gigawatts by the end of 2027. He says the limit on that growth is more power, cooling and infrastructure than chips.

More computers means more electricity, and more electricity means more turbines. Inside those turbines are the parts Elon says are holding up delivery.

Blades, not whole turbines

SpaceX still buys turbines. Elon's plan is to make the blades and vanes that go inside them by casting, which means pouring molten metal into a mold and controlling how it becomes solid. A vane directs the hot gas. A blade takes energy from that gas to turn a shaft.

His reason comes from the suppliers. "When we talk to the turbine suppliers, they tell us that the blades and vanes are 12 to 18 months behind the rest of the turbine," Elon said. "Casting is this really difficult bottleneck, and if anything, it seems to be getting worse. So that means that we can bring power online 12 to 18 months faster than anyone else."

If the rest of a turbine is ready before its hottest parts, ordering the whole machine still ties you to a smaller factory's schedule. SpaceX is betting that its own casting team can move the date an entire power project becomes useful. That only pays if its parts are good enough for a turbine to run on and arrive before the supplier parts they replace.

What a blade has to survive

A gas turbine compresses air, mixes it with fuel and burns it. The hot gas flows past rows of stationary vanes that aim it onto rotating blades, the way moving water pushes a water wheel. The blades turn a shaft. Some of that work runs the compressor, and the leftover power can drive a generator. The blade sits exactly where burning fuel becomes mechanical work.

It also has to keep its shape. A turbine blade is an airfoil: a carefully shaped surface that controls the flow around it, similar in idea to an aircraft wing. If that surface slowly stretches or distorts, the passage the gas flows through changes, and so do the loads on the part. A blade can be ruined while it is still completely solid.

That slow deformation under heat and load is called creep. Leave a weight on something long enough and the material gradually gives way. On top of creep, every start and stop heats and cools the metal, which adds repeated stress, and hot gas attacks the surface chemically. The blade takes all of it at once.

So melting point tells you very little. Metal can lose the strength to hold its shape long before it becomes a puddle. The Department of Energy says gas in a typical power turbine can reach about 2,300°F, while some critical metals can withstand only about 1,500 to 1,700°F. Even at the top of that range, the gas runs about 600 degrees hotter. This only works because the design keeps the metal cooler than the gas around it, every second the machine runs.

Freezing metal the right way

These parts often use nickel-based superalloys. Nickel is the foundation, and other elements are mixed in so the metal holds its strength under extreme heat. The recipe only gets you partway. How the metal freezes also decides how well it resists deformation.

Ordinary cast metal is made of many separate crystals called grains. Where neighboring grains meet there is a boundary, and in a hot turbine those boundaries can become the places where deformation starts. Directional solidification freezes the metal in a set direction so the grains grow the way you want. A single crystal goes further: the whole casting grows as one continuous crystal, with no boundaries between grains. That makes the blade more resistant to that weakness. It does not make the blade immune to damage.

Not every blade needs it. Howmet, a major supplier, describes industrial turbine blades as predominantly conventional grain or directionally solidified, with single crystal a smaller portion. Single crystal is simply where you can see most clearly how hard it is to control the metal.

Cooling from the inside out

Even the best alloy needs help. A blade can have internal passages that carry air taken from the compressor. That air is hot from compression but cooler than the combustion gas outside, so it pulls heat from the metal as it flows through. Some of it then escapes through small holes and spreads across the outer surface as a protective layer. This is called film cooling, and it only works if the holes are in the right places. Miss a spot and that spot runs hotter.

A ceramic coating on the outside slows the heat on its way to the metal. The coating has to stay attached as the part heats and cools, while the metal under it still carries the load.

Cooling air also costs something. The compressor already spent energy on that air, so diverting more of it for cooling takes efficiency from the turbine. The engineer has to keep the metal alive without giving away the work the machine exists to do.

How you pour a hollow blade

To get a hollow space inside a metal blade, you have to put something in that space before you pour.

The foundry makes a ceramic core shaped like the internal cooling passages, then forms a wax pattern around it in the outside shape of the part. The wax is coated to build a hard ceramic shell and then removed, leaving a cavity with the core still inside. Molten metal fills the gap between shell and core.

So before a single blade exists, the foundry has already made tooling for both its outer surface and its hidden air passages. If the core shifts during the pour, the metal wall comes out wrong. If it breaks, so does the passage. A blade with the right outside and the wrong inside still fails.

The mold is heated, and the alloy is melted and poured under controlled conditions, with a vacuum keeping reactive ingredients away from the air. For directional solidification, the bottom freezes first and the mold moves through a temperature boundary so solid metal grows upward into the liquid. Change the speed and you change how the crystals form. For a single crystal, a narrow selector under the useful part makes the growing grains compete. In one design it is a twisting passage: grains that hit the walls get blocked, and one survivor grows up into the blade.

Afterward the shell comes off and the core is removed from inside. Both are destroyed to make the part, so the foundry has to keep making accurate replacements. Then comes heat treatment, which has to stay within limits because too much heat damages the structure it is meant to improve. After that come finishing, the features that connect the part to the machine, precisely drilled cooling holes and the coating. Each step has to preserve the work done before it.

Busy is not the same as delivering

That makes a defect found late brutal. If a problem shows up after finishing, the mold work and furnace time are already spent. A tiny unwanted grain is a problem in a part meant to be a single crystal. A small pore can become the place where a crack starts.

As Farzad puts it, a busy factory and a factory delivering accepted parts are very different measurements. The parts have to come out right again and again, pass inspection, and earn the customer's confidence that they will do the job in its turbine. Building that confidence through evidence and testing is called qualification. A good sample is encouraging. A dependable flow of accepted parts is what changes a delivery schedule.

Why the line is so long

These foundries also supply turbines that are already running. Blades and vanes wear out, and those machines need replacements to keep making electricity. Siemens Energy says about 70% of its blades and vanes go into service, meaning roughly seven of every 10 go to machines already installed. New data center demand does not cancel that obligation.

GE Vernova says hitting its end of 2026 bookings target would leave it mostly sold out through 2030. That is well past the end of 2027 revenue goal. GE Vernova has also described funding additional supplier furnaces, with metal arriving before the later increase in finished turbine output. That order is exactly what the manufacturing story predicts. A funded furnace still has to become part of a process that reliably produces accepted parts.

So why not just buy from GE or Siemens? SpaceX does buy the machines. That leaves it stuck behind the same scarce parts, and the shortage is why Elon wants another source of blades and vanes.

What SpaceX actually brings

There are real reasons to take the attempt seriously. According to Elon's public statements, SpaceX has had an in-house superalloy foundry since at least 2019, long before this power push. Elon has described a SpaceX superalloy called SX500, made for the hot, oxygen-rich environment of Raptor's turbopump. A turbopump uses a turbine to drive the pumps that feed propellant into the engine.

Farzad is careful about what that proves. SX500's chemistry has not been publicly disclosed, and there is no verified evidence that it is the alloy for these power turbine parts. SpaceX's Bastrop job postings describe the actual target: power generation components made through the wax and ceramic process, including single crystal work, nickel superalloy castings and an industrial gas turbine blades and vanes line.

There is a Tesla thread too. Foundry Management and Technology reported that one person was recruited to coordinate alloy development for both Tesla and SpaceX. But Tesla's large car body castings are aluminum, made by high pressure die casting into a reusable steel die. The blade process uses controlled solidification and a ceramic shell that gets destroyed. The materials expertise carries over. The equipment does not. The more direct link is the Megapack batteries installed at Colossus, which store electricity and release it when needed.

Taken together, Farzad's view is that Elon's industrial background makes his companies the best positioned to tackle the physical constraints on AI.

The part nobody has proven

Now the other side. Jefferies says its industry checks put SpaceX turbine parts about four years away, and sector analysts cited by the Wall Street Journal say a new casting factory takes at least four years. That is at least four times the one-year window between the $100 billion and $1 trillion marks. No SpaceX-cast industrial turbine blade has been publicly shown or verified as shipped, and the public hiring record says nothing about accepted production volumes.

Elon is claiming a head start of 12 to 18 months. If SpaceX gets it, Farzad says, power arriving that much earlier matters enormously to a business trying to grow its revenue pace. For now, with the analyst estimate on one side and no demonstrated blade on the other, the advantage is unproven.

Partners versus the whole stack

Other AI companies are taking a different route. The big cloud operators and the neoclouds focused on supplying AI computing need power too, and their documented projects go through power companies and equipment suppliers. Meta's Hyperion project has Entergy building combined cycle plants, where a gas turbine makes electricity and its exhaust heat makes steam for more. Crusoe has ordered gas turbines from GE Vernova, and equipment suppliers like GE also offer service agreements covering parts and maintenance planning.

Farzad reads those deals as buyers getting power through organizations that already build and service the equipment. He is explicit that the deals do not show those companies are incapable of making a blade.

His distinction: securing capacity through partners is different from controlling the entire physical cost stack. Buy a partner's output and the partner controls the process. Bring the process inside and you own the rejected parts and the unfinished qualification. You also get the chance to send a cooling passage defect straight back to the people adjusting the casting. In Farzad's words, ownership of the factory makes the attempt possible. Dependable output makes it useful.

Bigger than one blade

Farzad sees the foundry as one constraint inside a much larger energy build-out. In his book Abundance or Collapse, he writes, "My first principles take, in the long run, the sun wins." Solar panels have their own manufacturing requirements and do not need hot section airfoils, so they can add power while the turbine industry works through its constraints. Farzad also thinks batteries and AI inference placed where the grid has idle capacity can put that capacity to work. He calls the grid constraint political and regulatory as well as physical.

Go back to the finished data center waiting on its blade, and follow the electricity out the other side. Picture a small business owner who spends the end of every day on work they cannot afford to hand to anyone else. An affordable AI tool that takes some of that work gives them their evening back. That tool needs working computers, and those computers need electricity. More power alone does not guarantee it, since the service still has to work and be priced so that person can pay. But the chain leads back to a blade that holds its shape in gas hotter than its metal can stand. The person getting their evening back will never need to know how it was made.