Parallax is gambling with fuel efficiency to feed AI’s power hunger.
The wait for a gas turbine is now longer than the wait for a custom-built supercar. Multiyear backlogs. Multi-thousand-part assemblies. Aerospace-grade metallurgy. A handful of giants — Caterpillar, Cummins, and the legacy manufacturers — sitting on order books so fat that lead time has become less a competitive variable than a weather condition.
That is the supply shock the AI buildout has created, and it is precisely the supply shock a tiny California startup called Parallax intends to break.
Founded earlier this year by Stanford graduate Carl Schoeller, who previously co-founded a defence-hardware company, Parallax has raised $117 million to do something audacious: print, rather than assemble, the turbines that will power the artificial-intelligence economy. The company is betting that the multiyear turbine backlog is not merely a supply-chain problem. It is a design problem.
The distinction matters. If the industry cannot deliver a conventional turbine core with more than 2,000 parts, Parallax wants to redesign the core around a handful of printed pieces, lower combustion temperatures, and material streams that already feed aerospace manufacturing. The resulting machine would generate roughly 10 megawatts, fit inside a shipping container, and trade some fuel efficiency for the one thing the AI buildout is starving for: speed.
That is the whole trick.
Schoeller put the customer distinction plainly: “We don’t have to design for a utility. We don’t have to design for aircraft. We get to design for this customer.” In a market that has spent decades optimising machines for the grid and the aircraft industry, that is not a minor change in specification. It is a change in what counts as success.
A utility wants thermal efficiency over a long operating life. An aircraft manufacturer wants a tightly controlled, highly certified system. A data-centre developer wants power delivered before the revenue opportunity disappears. The machine that arrives in 2028 can be more valuable than the machine that arrives in 2031, even if the later machine burns less fuel per kilowatt-hour.
The thesis is not that smaller is better.
The thesis is that shippable is better.
Each Parallax unit is designed to produce about 10 megawatts and ship in volume. A typical AI data centre requires hundreds to thousands of megawatts. The arithmetic is not subtle: dozens to hundreds of these units per site, depending on the load. A machine that can be manufactured and delivered in parallel attacks the problem at the level where the problem now exists — not inside the turbine laboratory, but on the construction schedule.
Eclipse founder Lior Susan, whose firm led the round, described the power market as “starved” and compared Parallax’s parts-consolidation strategy to SpaceX’s approach to rocket engines. The comparison is less about rockets than about industrial method. Stop treating the inherited machine as sacred. Remove parts. Remove steps. Replace a specialised assembly chain with a design that can be manufactured faster from materials the broader industrial economy can actually provide.
That is why Lux Capital, Founders Fund, Greylock Partners, and General Catalyst joined the financing. This is not a bet on an incremental improvement to a mature product. It is a $117 million vote that the next industrial supply chain for AI-era power will be built, at least in part, by companies that treat manufacturing as a design problem.
The risk is equally plain. A lower-temperature, common-materials machine that sacrifices efficiency does not abolish the cost of power. It moves the trade-off. More fuel per unit of electricity means more operating expense, more emissions, and a larger claim on the gas infrastructure feeding the site. Three-dimensional printing does not make thermodynamics negotiable. It makes a different constraint negotiable: the time required to manufacture the machine.
To be fair, that may be enough. The customer is not choosing between a Parallax turbine and a perfect zero-cost alternative. The customer is choosing between a turbine that arrives in time to power an operating data centre and a more efficient turbine that remains trapped in a multiyear queue. When months of delay can erase billions in expected AI revenue, fuel efficiency becomes a secondary variable.
Speed is the product.
The surrounding market is already signalling that the constraint is real. Generac has locked in $2.4 billion in initial deliveries to Amazon for backup generators across 2027 and 2028. Caterpillar and Cummins are pivoting toward data-centre power equipment. Elon Musk has purchased a turbine manufacturer and announced that SpaceX will make its own blades and vanes for natural-gas-fired power turbines.
SpaceX is also planning a natural-gas plant to power a Texas AI-chip facility, while Tesla and SpaceX are stacking infrastructure projects across Texas and Louisiana. Those projects are not identical to Parallax’s, but they point in the same direction: the AI race is becoming a race to build power faster than the old industrial system can supply it.
The legacy suppliers are not standing still. They have balance sheets, customer relationships, certification experience, and regulatory muscle. But they are still optimising inside the inherited constraint — long lead times, specialised materials, complex assemblies — while Parallax is attempting to delete that constraint from the problem.
Parallax’s timeline says what kind of company it is trying to become. Prototype by year-end. Testing in 2027. Customer deliveries in 2028. Three years from founding to revenue, in an industry where established manufacturers quote multiyear backlogs.
The timeline is aggressive. The physics is indifferent. The supply-chain problem is not solved because a venture-backed company has described a better architecture. The printed parts must survive combustion, vibration, heat, maintenance cycles, certification, and the unromantic punishment of continuous operation. A turbine is not a slide deck with a rotating shaft.
But the opportunity is enormous because the constraint is real. Data-centre developers are locking down turbines, reciprocating engines, and fuel cells with the urgency of a wartime procurement cycle. Some sites need hundreds or thousands of megawatts — orders of magnitude beyond what any single legacy machine can provide. The AI buildout will not wait for the old grid to finish arriving.
If Schoeller delivers a working prototype by year-end and ships in 2028, the $117 million round will look like the cheapest seed ever planted in the AI infrastructure stack. If he does not, it will look like a cautionary tale about how difficult gas turbines remain when the manufacturing problem is reduced to a sentence and the machine is required to work.
Either way, the bet is on the table. The AI power boom is not waiting for the legacy supply chain to catch up, and a new generation of builders is racing to print, not forge, the machines that will run it.