Why ASML Controls the AI Economy More Than OpenAI Ever Will
The Semiconductor Chokepoint
Everyone is wrong about what constrains the AI.
Not capital. Not talent. Not compute demand.
Semiconductor supply.
Specifically: one Dutch company shipping 40 machines per year that can print sub-5nm chips.
ASML. You’ve probably never heard of it. That’s the point.
The entire AI stack; every training run that matters, every inference cluster, every GPU that powers the models everyone is betting billions on; depends on a supply pipeline so narrow it could be shut down by a single policy decision in The Hague.
And everyone’s focused on OpenAI’s moat.
This is an analysis of structural economic power. Of where actual leverage lives in the AI economy. And of why the next decade will be decided not by which model is smarter, but by who controls the machines that make the chips that run the models.
The Chokepoint Is Real
Let’s start with the numbers, because they’re not subtle.
ASML manufactures Extreme Ultraviolet (EUV) lithography machines. These are the only tools capable of printing transistors smaller than 5 nanometers at scale. They are to the semiconductor industry what oil refineries are to energy: not the resource itself, but the infrastructure that converts raw material into usable output.
The company ships approximately 40 machines per year globally. This is not a constraint they chose. This is the maximum they can produce while maintaining quality and avoiding yield disasters.
Each machine costs $150-200 million. They take 3-4 years to design and build. They require 100+ component suppliers across multiple continents. When a component fails—which it does occasionally—the entire production line stalls.
ASML is not scaling. They can’t. The physics and logistics of EUV manufacturing don’t permit it.
So when global AI demand is doubling every 12-18 months, and semiconductor supply can grow at maybe 10-15% per year, you have an inelastic supply problem.
And it’s about to get worse.
Why This Can’t Be Solved
The usual response to bottlenecks is: someone will build a competitor.
This is where most analyses fail catastrophically.
You cannot compete with ASML on EUV lithography. Not because it’s expensive. Not because it’s hard. But because ASML has achieved a 15-20 year head start through accumulated knowledge and supply chain integration that literally cannot be replicated.
Here’s why:
ASML owns the intellectual property moat. ASML has 40+ years of lithography R&D. They own 15,000+ patents in EUV and related technologies. They acquired CYMER (the EUV pioneer) and integrated it vertically. They’ve spent $20+ billion in R&D on a single technology. Competitors cannot buy their way into this overnight; the knowledge is embedded in thousands of engineers and systems.
The supply chain is non-redundant. An EUV machine has over 100,000 components from 700+ suppliers. ASML spent decades integrating these suppliers, qualifying their processes, and managing yield. If you try to build a competitor, you’d need to either:
Rebuild every supply chain integration from scratch (8-12 years minimum)
Convince all 700+ suppliers to switch to you simultaneously (economically irrational)
Absorb catastrophic yield losses for years while you de-bug the supply chain
None of this is feasible. The supply chain advantage is not technology; it’s logistics and relationships. It’s the definition of a durable moat.
The physics is at the edge of what’s possible. EUV lithography uses light at 13.5 nanometers wavelength to print features 100x smaller. The tolerances are insane. A mirror can be off by a single atomic layer and the entire system fails. This isn’t just engineering; it’s physics at the limit of human manufacturing capability. The learning curve is vertical.
The regulatory moat is now hardening. ASML is Dutch. But the US controls ASML’s access to components (semiconductor materials, certain tools). In 2023, the US tightened export controls. ASML cannot sell advanced EUV machines to China without US approval. As geopolitical tension increases, this regulatory advantage becomes a policy chokepoint that makes ASML’s technical advantage even more durable.
So when someone says “oh, Samsung or Intel will just build their own,” you’re watching economic illiteracy in real time. Samsung and Intel have been trying to compete with ASML for 20 years. They’ve failed. Not because they’re incompetent, but because building an EUV machine is harder than putting a human on the moon and doing it 40 times a year.
The Supply Crunch Is Already Here
This isn’t a future problem. It’s happening now.
GPU supply is constrained by chip supply. Nvidia can design chips faster than TSMC can manufacture them. Not because TSMC doesn’t have capital. But because TSMC’s capacity is constrained by ASML’s 40 machines per year. Nvidia’s 3-year waiting lists for H100s and H200s? That’s ASML scarcity, translated through TSMC, back to you.
Chip fabs are operating at max utilization. TSMC is running 24/7 across its fabs. Samsung is doing the same. Intel is struggling to ramp new capacity because;you guessed it; they don’t have enough ASML machines. Every fab in the world that can print sub-5nm is maxed out. None of them can expand significantly without more EUV machines. And ASML can’t sell 60 machines because they can’t make 60 machines.
Lead times are extending. In 2024, lead times for advanced chips started stretching to 6-9 months. This isn’t demand softness; it’s supply. The entire pipeline is pinched.
China is locked out. By US export controls, ASML cannot sell the newest EUV systems to China. This means China cannot make sub-5nm chips at scale. (China has been working on its own EUV tool, SMEE, for years. They’re still decades behind. Realistically, they cannot catch up.) This is the ultimate strategic chokepoint: America’s semiconductor supply is constrained by physics. But China’s is constrained by policy. And policy can be enforced with sanctions.
Why Semiconductors Matter More Than AI Models
Let me be direct: the entire venture capital ecosystem is betting on AI model innovation.
Anthropic, OpenAI, Google DeepMind, Mistral; all running $500M-$1B+ bets on “we’ll build better models.”
This is the wrong bet. Model improvement follows a law of diminishing returns. Each 10% improvement in capability requires 100% more compute. The frontier of model capability is not limited by our ability to design architectures; it’s limited by how many chips we can build.
The real economic value in the AI era is not captured by whoever builds the smartest model. It’s captured by whoever controls the chips the models run on.
Think about the value chain:
ASML: Sells 40 machines at $150M each = $6B revenue. But ASML’s margin structure is extreme. ~30% operating margin = $1.8B annual profit from a single chokepoint.
TSMC: Manufactures chips, sells them to Nvidia/AMD. TSMC has ~25% gross margin on advanced nodes. Their entire fab capacity is ASML-constrained.
Nvidia: Designs chips, sells them to hyperscalers. Gross margins ~60%. But they’re waiting years for TSMC to manufacture GPUs.
OpenAI: Runs inference and training. Margins are negative in most years because compute costs $2-5B annually.
Follow the economic rents: ASML captures pricing power because they have an inelastic supply.
When supply is constrained and demand doubles every 18 months, the company with the constraint extracts maximum value.
This is elementary microeconomics. And nobody..not VC, not hyperscalers, not governments…is optimizing for the right chokepoint.
The Geopolitical Endgame
Here’s where this becomes strategic thinking.
The US has already realized this. In 2023, the US tightened export controls on advanced semiconductor tech to deny China access to leading-edge chips. This wasn’t a technical move…it was a supply chain chokepoint play.
The logic: If China can’t make sub-5nm chips, they can’t build competitive AI systems. They’re locked out for 5-10 years minimum.
But this strategy has a fragility: it depends on ASML remaining Dutch and US-allied.
What happens if:
The Netherlands faces political pressure (China has leverage through trade) to relax export controls?
A geopolitical crisis forces ASML to choose sides?
Someone else figures out how to build EUV machines (unlikely, but non-zero)?
ASML’s supply chain gets disrupted (Taiwan tension, component supplier failure)?
Right now, US semiconductor dominance is built on three legs:
ASML’s manufacturing
TSMC’s fabs (Taiwan; existential risk)
Nvidia’s design (US-based, but can be sanctioned or replicated)
The weakest link is TSMC (Taiwan geopolitics). But the chokepoint link is ASML.
China understands this. They’re investing heavily in SMEE (their domestic EUV company) and in alternative lithography approaches (like Imec’s more exotic techniques). These efforts are 5-10 years away from commercial viability. But China is patient.
Russia would love to sanction ASML or steal IP, but they lack the economic leverage.
Europe (specifically the Netherlands) is now the swing player. They control access to the tool that controls the world’s chip supply.
This is an unusual geopolitical position. The Netherlands has more structural power over the AI economy than the US does.
The Operator Implication: Chip Supply as Strategic Asset
If you’re running a company that depends on compute (which, now, is most companies), you need to understand semiconductor strategy.
For hyperscalers (Meta, Google, Amazon): The constraint is not capital; it’s chip allocation. You cannot order Nvidia H200s and expect delivery in 6 months. TSMC has a waiting list. Your competitive advantage is your ability to secure long-term chip supply contracts before your competitors do. This is now a procurement/supply chain game, not a technology game.
For chip designers (Nvidia, AMD, Intel): You are constrained by ASML machine availability. You cannot improve margins by designing better chips if TSMC can’t make them. Your economic value is now dependent on your ability to negotiate multi-year commitments with TSMC for ASML capacity. Intel’s recent struggles aren’t design failures—they’re capacity failures, which are ASML failures.
For governments: Semiconductor supply is now a national security asset. Which countries have access to advanced chips determines which countries can build competitive AI systems. The US has correctly identified this—hence the export controls on ASML. But this strategy only works if ASML remains US-aligned. Expect intense diplomatic pressure on the Netherlands in the next 3-5 years.
For startups: If you’re building AI products, your cost structure is determined not by model design, but by compute cost. And compute cost is determined by chip supply and pricing. This is why startups are increasingly turning to edge AI (smaller models, local inference) and synthetic data (reducing training requirements). They’re not doing this because it’s intellectually pure; they’re doing it because they can’t afford the compute bill if they’re waiting 2 years for chip allocation.
The Economic Constraint That Matters
The AI era will not be constrained by model innovation. It will be constrained by semiconductor supply.
This means:
The companies that win the AI race are not the ones with the smartest researchers
They’re the ones with the most durable chip supply contracts
This favors large hyperscalers (who can demand allocation) over startups (who can’t)
This favors countries with ASML access (US, South Korea, Japan, Taiwan) over those without (China, Russia, Europe outside the Netherlands)
The venture capital bet on “better models” is the wrong bet. The real bet is on whoever controls the supply chain that makes the models trainable.
ASML’s CEO Peter Wennink understands this. He’s operating a company with 15-20 year visibility into demand because demand is inelastic and growing faster than supply can grow. He can raise prices, extend lead times, and pick customers. He has maximum economic leverage.
The Scarcity Trap
There’s one more layer to this: the impossibility of escaping the constraint.
Chip fabs are considering how to expand. TSMC wants to build more fabs. Samsung is doing the same. Intel is making a massive capital bet.
But all of them are hitting the same wall: you cannot build a new fab without ASML machines. And ASML cannot scale production.
So the capital is flowing, but the actual productive capacity growth is limited by the speed at which ASML can manufacture 40 machines per year.
This creates a bizarre economic situation:
Hyperscalers are willing to spend $10B on new fabs
Chip fabs are willing to invest $20B in new buildings and equipment
But the actual constraint (ASML machines) is not scaling
So you end up with: tons of idle fab capacity waiting for ASML to deliver machines. This is a capital inefficiency. But it’s an inescapable one.
ASML could, theoretically, raise prices 50%. Demand would still exceed supply. They could 2x their EUV prices and the economics would still work; customers would still buy because they have no alternatives.
This is maximum pricing power. And it’s being exercised quietly, without drama or public awareness.
The 5-10 Year Implication
If semiconductor supply remains the constraint, here’s what we should expect:
Year 1-2:
Continued supply scarcity for advanced chips
Nvidia maintains pricing power (can charge what they want)
Hyperscalers hoard chips (anyone can sell if they have stock)
Startups get priced out of cutting-edge compute
China makes incremental progress on SMEE (still 5-10 years away from parity)
Year 3-5:
ASML may increase production to 50-60 machines/year (slow, continuous improvement)
TSMC and Samsung fabs come online (but gradually)
Chip pricing softens from current peaks, but remains elevated
ASML’s absolute profit grows as volume increases slightly
China’s EUV approach shows promise but hasn’t reached commercial scale
Year 5-10:
If ASML doesn’t increase output significantly, the constraint tightens further
If ASML does increase output (to say, 100 machines/year), we start seeing chip scarcity ease
Alternative lithography methods (more speculative) might start providing marginal capacity
Geopolitical pressure on ASML and Netherlands increases
The most likely scenario: ASML slowly increases production, but never fast enough to fully meet demand. Semiconductors remain expensive relative to the 2010s, and chip supply remains a strategic leverage point.
The Uncomfortable Conclusion
The AI economy is not built on intelligence. It’s built on sand.
No; worse. It’s built on geology. On rare earths, on silicon, on manufacturing precision that requires 15-20 years of accumulated expertise.
Every AI founder, every VC, every government betting on AI dominance is making an implicit bet: that semiconductor supply remains elastic.
It doesn’t.
ASML ships 40 machines a year. This is a hard constraint. Not a bottleneck to be optimized. A fundamental limit.
The companies that win the AI race will not be the ones with the best models. They’ll be the ones with the most reliable access to advanced chips. This favors scale (you can demand allocation), geopolitics (you have government backing), and supply chain integration (you own long-term contracts).
Every VC presentation about “the future of AI” is missing the real story: the future of AI depends on a Dutch company shipping 40 machines a year, and nobody’s building strategic reserves for when that’s not enough.
China’s AI ambitions are constrained by ASML policy, not their engineering.
Europe’s AI hopes are constrained by Netherlands export politics.
Russia’s AI dreams are constrained by sanctions that prevent buying components that go into ASML machines.
And the US’s AI dominance is built on an extremely fragile supply chain where a single point of failure (ASML production, Taiwan stability, US export control policy) could shift the entire balance.
This is not a technological problem. It’s an economic and geopolitical one.
And the person who figures this out first; who understands that semiconductors are the new oil, and ASML is the new Saudi Aramco—will have strategic clarity that 99% of the tech industry lacks.
The AI era is not being won or lost by who builds the smartest model.
It’s being won or lost by who controls the machines that make the chips.
And right now, that’s a Dutch company most people have never heard of.