Gas-Turbine Startup Parallax Raises $117M, Exits Stealth to Solve AI’s Power Problem

The real bottleneck for AI data centers isn’t chips — it’s electricity. Grid connections take years, and the gas turbines that could bridge the gap are themselves backlogged into the 2030s. SemiAnalysis estimates a gap of more than 65 gigawatts between compute already committed and power the grid can actually deliver by decade’s end. The machines exist. You just can’t get one in time.

Parallax said it has come out of stealth with $117 million raised to date. Eclipse co-led the $42 million seed with Lux Capital, Founders Fund, and Valor; Greylock Partners and General Catalyst led the $75 million Series A. Diffusion also participated.

Parallax is building a gas turbine designed from scratch for data centers — roughly 10 megawatts, small enough to ship in a container. Incumbent manufacturers chase efficiency through exotic materials: single-crystal blades cast by just two suppliers worldwide, and a rhenium superalloy with about 60 tons of global annual supply. Parallax runs its turbine cooler instead, trading away some efficiency in its first generation to design both chokepoints out of the machine entirely. A core that normally runs more than 2,000 parts drops to a handful through 3D printing. As founder Carl Schoeller told The Wall Street Journal, “We don’t have to design for a utility. We don’t have to design for aircraft. We get to design for this customer.” The plan is a prototype by year-end, testing next year, and customer deliveries in 2028.

Schoeller studied mechanical engineering at Stanford and, before Parallax, designed around 2,000 unique parts — some now in mass production on the Tesla Cybertruck, some in an Apple product. He co-founded Theseus, which builds GPS-free drone navigation systems and continues to operate today in partnership with U.S. Special Operations Command. Parallax is his second company.

Eclipse founder and CEO Lior Susan told the Journal that neither chipmakers nor the power industry are adding capacity fast enough to meet AI demand, leaving the market “starved.” “Building a gas turbine is freaking hard, but if you can do it well, naturally it can be a gigantic business,” he said, comparing Parallax’s parts consolidation to SpaceX’s approach to rocket engines.

Parallax is chasing a market with far more players than turbine makers alone. The field breaks into four distinct bets.

The first is designing new hardware from scratch to route around the bottleneck, same as Parallax. Elon Musk quietly bought mobile turbine maker APR Energy for roughly $1 billion in May, and is now building a dedicated turbine-blade-and-vane foundry in Bastrop, Texas.

The second is repurposing proven engines to deploy power right now. GE Vernova‘s “aeroderivative” turbines — adapted from commercial jet engines — are already powering Musk’s xAI Colossus and a Microsoft data center in Texas. ProEnergy takes retired CF6 engine cores from Boeing 747s and turns them into its PE6000 turbines, promising to compress a normally three-to-five-year development cycle down to 2027. Caterpillar and Cummins have redirected existing generator businesses toward AI data centers, and Generac disclosed a long-term deal with Amazon worth up to $8 billion for backup generators, with initial 2027–2028 deliveries alone valued at $2.4 billion.

The third skips new hardware altogether and builds full power plants on-site, bypassing the grid entirely. VoltaGrid is installing natural-gas microgrids directly at data center sites to supply Oracle with 2.3 gigawatts, while Joulent is partnering with Chevron to build a 2.67-gigawatt natural-gas plant, Project Kilby, for a Microsoft data center.

The fourth bets on nuclear instead of gas. Google has a deal with Kairos Power for up to 500 megawatts of small modular reactor capacity, Amazon invested $700 million in X-energy, and Meta lined up up to 6.6 gigawatts across Oklo, TerraPower, and Vistra. The catch: commercial SMRs aren’t expected online before 2030 at the earliest — later than even Parallax’s 2028 target.