SpaceX turbine foundry: AI compute waits on cast metal

The bottleneck on AI data center capacity is no longer chips. It is castings. Elon Musk confirmed on Saturday that SpaceX is building a turbine blade and vane foundry in Bastrop, Texas, because the parts that make a gas turbine spin are the thing nobody can get.

What actually happened

TechCrunch reported August 30 that Musk confirmed the foundry after The Information surfaced it through SpaceX job listings. SpaceX bought roughly 830 acres near its existing Starlink factory in Bastrop between March and June 2026. Musk's claim: "by doing in-house casting at SpaceX, we can accelerate nat gas turbines coming online by up to 18 months."

The reason that number is plausible is the reason the foundry exists. Per TechCrunch, only four companies worldwide have mastered industrial-scale turbine blade casting, and GE Vernova says it is essentially sold out of production capacity through 2030 — largely because of AI infrastructure demand. Blades and vanes run in the hottest section of a turbine and require single-crystal casting in vacuum furnaces. That is not a line you spin up in a quarter.

The same reporting flags the tradeoff: gas turbines already running at Musk's Memphis data center operation have drawn NAACP litigation over unpermitted operation and emissions.

Why it matters for your business

You are not buying turbines. You are buying the compute that sits behind them, and the delivery date on that compute now runs through a metallurgy supply chain with four credible vendors and a 2030 order book.

"Elastic" compute has a physical floor. The cloud sells you the abstraction of infinite capacity. Underneath it, a specific generator in a specific county has to exist. When a hyperscaler defers a region, the reason is increasingly a turbine, a transformer, or an interconnect queue — not a business decision you can appeal.

Price stability is the thing to watch, not availability. Constrained supply does not usually show up as an outage. It shows up as a rate card. If your unit economics assume today's inference cost holds through 2027, that assumption is resting on a foundry in Bastrop that has not cast a blade yet.

The move is not to panic-reserve capacity. It is to know what your cost model breaks at. Run the numbers at 1.5x your current token spend. If the business still works, carry on. If it does not, you have a design problem to fix now, while fixing it is cheap.

Sources: TechCrunch.