Musk is now casting his own gas turbine blades to keep SpaceX's AI compute running

The binding constraint on AI compute has moved from chips to data centers to the inside of a gas turbine. Musk is opening a metal foundry in Bastrop, Texas to pull the casting step in-house.

SpaceX has posted job listings for a "Blades and Vanes Foundry" in Bastrop, near Austin, that will cast the rotating airfoil-shaped parts inside natural-gas-fired generators. Musk framed the move in two X posts this week as a response to a power-constraint gap he says could leave roughly 15 gigawatts of 2027 AI compute unable to start up, about the output of fifteen large nuclear reactors running flat out (QbitAI summary of the Musk posts; Musk on the 2027 power gap).

The bottleneck Musk is naming is not the turbine itself. It is the casting.

In a gas turbine, the spinning blades and the stationary vanes that direct airflow sit in the hottest, most stressful part of the engine, which is why they are made from nickel-based superalloys poured into ceramic molds in a process called investment casting. Musk's pitch is that SpaceX can deliver those castings in-house about 18 months earlier than waiting on an outside supplier, and that every other major turbine component can arrive 12 to 18 months ahead of the castings anyway (Musk on the 18-month in-house lead time). For a company that already builds Raptor rocket-engine turbomachinery, including castings for extreme-environment rotating parts, the foundry is a continuation of an existing capability, not a leap into a new industry.

The 15-gigawatt-at-risk figure is Musk's own estimate, not an industry consensus. The company has not shown that the Bastrop foundry is in production; the listings describe a facility being stood up, not a running line. Treating the foundry as a working answer to the 2027 power gap would overstate what is currently public.

GE Vernova, the dominant Western gas-turbine maker, told investors in its most recent quarterly disclosure that gas-turbine orders and capacity reservations hit 116 gigawatts and are expected to reach at least 125 gigawatts by year-end, with capacity already booked through 2031. Its chief executive has said 2030 capacity is likely to sell out by the end of 2026. Against that backlog, Musk's pitch to SpaceX is that an in-house casting line can sit in front of the queue rather than behind it, because the castings are the slow step, not the turbine assembly.

Musk's public targets for SpaceX's own AI compute are more than 2 gigawatts online by the end of 2026 and roughly 10 gigawatts by the end of 2027, supported by 20 gigawatts of planned power, cooling, and electrical infrastructure. If the foundry's casting lead times are real and the 18-month pull-in holds, it could close a meaningful slice of the gap between those targets and the 15-gigawatt shortfall Musk has warned about for 2027. If they are not, the foundry will look more like an expensive hedge against a queue SpaceX could not otherwise jump.

TechCrunch's reporting on the same week adds a counterweight the foundry math alone cannot answer. Gas turbines are faster to deploy than the new solar and wind capacity Musk also champions, but they raise emissions per megawatt-hour in the meantime. A faster path to AI compute is not a free lunch, and any honest read of the foundry has to carry both numbers at once.

The next dated milestone that will move this story is whether SpaceX's Bastrop listings convert into a running foundry line before the end of 2026. The castings are the slow step, but they are also the step Musk has now put inside his own factory walls.