Space vs. AI: Who Gets the Electrical Gear?

The U.S. space industry wants to scale. The supply chain won't let it.

That's the conclusion of a March 2026 report from the Aerospace Industries Association and PricewaterhouseCoopers, which found that demand across the U.S. space sector is outpacing supplier capacity across critical components. The report identifies nine categories of specialized parts — optical intersatellite links, actuators, rocket motor nozzles, composite overwrapped pressure vessels, connectors, switchgear and transformers, valves, field-programmable gate arrays, and post-processing services — where supply is running well behind what the industry needs to meet current launch schedules.

"Demand growth across the U.S. space sector is outpacing supplier capacity," Doug Anderson, a partner in PwC's operations and supply chain practice, said in the report. "Now is the time to reassess supply strategies, invest in scalable capacity, and build the operational resilience needed to compete in a rapidly expanding market."

The report landed in a space industry already feeling the squeeze from an unlikely competitor: AI data centers. Switchgear, transformers, and electrical infrastructure that space companies need for launch pads and ground stations are the same equipment that hyperscalers are buying to power the current wave of AI compute buildout. The U.S. added roughly 20 gigawatts of new data center capacity in 2025, and the queue for utility interconnection equipment is long enough that space companies can end up at the back of it, with lead times stretching years.

That's the visible problem. The structural one runs deeper — and the most consequential constraint for the semiconductor content of those nine component categories isn't a part on any one satellite, but a machine that none of them can be built without.

ASML's extreme ultraviolet lithography systems — the machines required to fabricate the most advanced chips, from the processors in Starlink satellites to the FPGAs managing spacecraft avionics — are produced at a rate that is nowhere near what current demand profiles require. ASML shipped 48 EUV systems in 2025. Analyst projections put 2026 at roughly 67 tools and 2027–2028 in the 80-to-85 range per year. The physical ramp in tool count is measured in single digits per year. Dylan Patel's estimate that ASML can make about 70 tools annually is a manufacturing capacity figure, not a shipment rate — and not a number the company itself publishes. The gap between what ASML's factory can theoretically produce and what it actually ships is meaningful for anyone doing supply-side math. The AI labs know this. They're already queuing for position.

Standard EUV lithography systems cost roughly $200 million per unit. ASML's high-NA system, the Twinscan EXE, runs approximately $380 million. Either way, you are not going to fab more of them next quarter because a satellite program needs them.

Patel has been tracking the EUV tool allocation math closely. His calculation, as he laid it out on a March 2026 episode of the Dwarkesh Patel Podcast: a data center consuming one gigawatt of power requires approximately 3.5 EUV tools to produce the advanced chips that run it. The four major hyperscalers — Amazon, Google, Meta, and Microsoft — are collectively spending $600 billion in capital expenditures this year. Sam Altman has said $20 billion in infrastructure per gigawatt of compute per Axios. At that run rate, 50 gigawatts of AI compute represents a $1 trillion infrastructure commitment. At 3.5 tools per gigawatt, that's 175 EUV tools demanded just for AI data centers — against a global production ceiling that won't reliably clear 90 per year this decade. Space satellites consume fewer advanced chips than hyperscale data centers running inference at scale, but the fab capacity that produces the advanced silicon in several of the nine AIA-PwC component categories — connectors, FPGAs, and avionics processors among them — draws from the same queue. And the queue is getting longer and better-funded on the AI side every quarter.

"ASML will be the number one constraint for AI compute scaling by 2030," Patel said on that same episode. The argument is structural: you cannot train frontier models or run inference at the scale the labs are projecting without the wafer capacity that EUV enables, and you cannot get more wafer capacity without more EUV tools. There is no substitute for ASML in the advanced lithography market. The Dutch company's monopoly on the equipment needed to print the most advanced logic is not under immediate threat from any competitor.

The AIA-PwC report's nine shortage categories are independent bottlenecks in the strict engineering sense — a connector shortage and a motor-nozzle shortage do not have the same root cause. What they share in practice is competition: the same capital, the same qualified supplier base, and in some cases the same advanced silicon, all being pulled simultaneously by defense programs, satellite manufacturers, and the largest semiconductor buildout in history. The EUV constraint is one thread running through several of those nine categories, not a grand unifying explanation for all of them. Whether you call that a common root or a compounding coincidence is a framing choice. The effect is the same either way: space companies that need specialized components are queuing behind buyers who can pay more, wait longer, and make longer-term volume commitments.

ASML's CEO, on the company's most recent earnings call, pushed back directly on the bottleneck framing: "I sense concern we may be the bottleneck," he said. "This is not the case, certainly not this year and again for next year." He has a point for the near term — 48 tools is a real number, and the AI compute ceiling the labs are projecting is a multi-year construction problem. But the physical ramp in tool production is measured in single digits per year, and the demand trajectory is not. The CEO's confidence about the next two years doesn't resolve what happens at 100 tools a year when the hyperscalers are collectively demanding allocations for multi-gigawatt campuses.

There is also no shortcut around the physical limits of the machine. ASML and Zeiss announced on February 23, 2026, a prototype 1-kilowatt EUV light source, targeting 330 wafers per hour throughput by 2030 — up from the roughly 220 wafers per hour the NXE:3800E achieves today. The improvement is real. It is also a component-level R&D milestone, not a production line. The gap between a 1kW prototype announcement and a tool that fabs can actually run at scale is measured in years of integration, yield validation, and process development.

TSMC is currently the dominant EUV customer, absorbing a substantial share of available capacity. Within the U.S., space companies compete for what remains — and the competition includes hyperscalers with purchasing power that satellite manufacturers cannot match.

The implications are immediate for satellite manufacturers. The AIA-PwC report notes that testing infrastructure is a compounding bottleneck: satellites and their components must undergo radiation and environmental testing before launch, but the facilities available are limited enough that companies routinely wait months for a slot. A delay in component fabrication becomes a delay in testing, which becomes a delay in integration and launch. The report recommends that government agencies and major customers provide clearer, longer-term demand forecasts to give suppliers the confidence to invest — but that is a policy fix running up against a hardware constraint that is not responsive to procurement announcements.

The report also flags workforce aging and a limited pipeline of skilled technicians and engineers as structural drags on scaling production. This is real. But it is a slower problem than the EUV constraint. You can train more machinists. You cannot fab more EUV tools next quarter.

What this looks like in practice: a connector manufacturer with three qualified domestic suppliers. A fastener supplier whose facility fire — Anderson didn't specify a year — disrupted supply chains across both aerospace and commercial space. FPGAs sourced from a handful of radiation-tolerant fabricators, each of which is simultaneously fulfilling orders from defense programs, satellite manufacturers, and AI hardware companies. The market signals are all pointing in the same direction — more demand, same supply — and the suppliers qualified to serve space are not adding capacity fast enough to matter in the current cycle.

"Without deliberate steps to strengthen suppliers and modernize regulations, we risk turning today's momentum into tomorrow's bottlenecks," Eric Fanning, the AIA's president and CEO, said. "A resilient space supply chain is not optional — it's a national imperative."

He's right about the momentum. The momentum is also running into physics. ASML's factory in Veldhoven, Netherlands, is not doubling its output next year. The tools already allocated to TSMC and other major customers are not becoming available to space manufacturers because the AI labs decided to be more conservative on compute spend. The nine-component shortage in the AIA-PwC report is a symptom of a constraint that will outlast any single procurement cycle — and the EUV tool constraint is the part of that story that no procurement policy can address in this decade.

The report recommends better government-industry coordination on demand visibility, regulatory modernization to reduce qualification overhead for new suppliers, and investment in shared testing infrastructure. These are sensible. They are also insufficient on their own to change the trajectory of a supply chain that is now competing, in the same sub-component markets, against the most capital-rich semiconductor buildout in history.

What's worth watching: whether the Space Development Agency's optical intersatellite link programs — the SDA's mesh network architecture depends heavily on OISLs — can hold schedule as laser component suppliers stretch to fill orders from both government and commercial programs. If the bottleneck bites harder there, the DoD's LEO architecture has a problem that budget cannot solve.