The world's AI labs are racing toward 28 gigawatts, 50 gigawatts, more. ASML makes fewer than 100 EUV lithography machines per year. That's the math.
But it's not the whole story.
ASML recognized revenue on 48 EUV systems in full-year 2025, up from 30 in 2023 ASML Q4 2025 investor deck. Analyst Ming-Chi Kuo projects 67 tools in 2026 and 80 to 85 in 2027 Luminix. Bank of America projects near-full ASML capacity at roughly 90 tools per year by late 2027 Luminix. The ceiling is real. But here's what the doomsters are missing: ASML unveiled a 1 kilowatt EUV light source in February 2026, capable of delivering a 50 percent throughput gain by 2030 without a single new machine being built Reuters, Feb 23 2026. SemiAnalysis reported the development ahead of broader public reporting. The ceiling exists. It keeps getting re-negotiated.
The compute race is running into three walls simultaneously: logic, memory, and power Dwarkesh Patel / SemiAnalysis. On the power side, SemiAnalysis correctly called the US AI power scaling from 3 GW to 28 GW Luminix. The big four hyperscalers — Amazon, Meta, Google, and Microsoft — are on track to spend $600 billion in combined capital expenditure in 2025, driving toward roughly 50 gigawatts of AI compute demand Dwarkesh Patel / SemiAnalysis. A single gigawatt-scale AI data center needs about 55,000 wafers of 3nm logic, 6,000 wafers of 5nm, and 170,000 wafers of DRAM memory Dwarkesh Patel / SemiAnalysis. That memory number is where the next wall appears.
By 2026, 30 percent of tech giant capital expenditure will be consumed by storage and memory, driven by HBM demand from inference models Longbridge. Nvidia already has $90 billion in long-term contracts and is negotiating three-year deals with memory vendors to lock in supply Dwarkesh Patel / SemiAnalysis. Even if you solve the EUV tool shortage, memory is the next bottleneck. Dylan Patel of SemiAnalysis calls ASML the number one constraint for AI compute scaling by 2030 — but that framing may be too generous to the other two problems Dwarkesh Patel / SemiAnalysis.
ASML itself is planning a five-times EUV capacity increase compared to 2020 through a combination of unit growth and productivity improvements SemiAnalysis. Zeiss, which makes the optics inside every EUV system, is expanding capacity at 20 to 25 percent year-over-year Luminix. The optics bottleneck is separate from the machine bottleneck — even if ASML hits 90 tools per year, Zeiss has to be able to fill them with mirrors and lenses that don't yet exist at that scale.
The geopolitical dimension adds its own friction. No EUV system has ever been sold to a customer in China, ASML confirmed to Reuters Reuters. China accounted for roughly 42 percent of ASML's total revenue through the third quarter of 2025, but that figure is almost entirely DUV immersion systems — the older, less restricted technology. Sales to China fell about 23.6 percent year-over-year as export controls tightened Tom's Hardware. China imported 70 percent of ASML's DUV immersion systems in 2024 CNAS. The export control regime keeps EUV out of Chinese fabs but creates a growing依赖 on DUV revenue that ASML's long-term customers are systematically reducing.
What does this mean for someone building or investing? The scarcity story is real but overstated in the short term. The 1kW source is not a press release — it's a hardware development that, if it yields as projected, changes the effective supply math by 2030 without a single new fab. The ceiling is lower than the ambition. It's just not as low as the pessimists say, and the timeline for when it bites keeps slipping.
The deeper risk is that the industry is solving one bottleneck while running straight into the next. EUV tools get more productive. Memory demand compounds. Power infrastructure takes years to permitting. The compute race isn't constrained by one wall — it's a relay race through three.
