Described as exponential dynamics, driven by massive investments, rapid algorithmic advances and an almost unlimited demand for computing power, artificial intelligence seems limitless. In this storytelling, everything seems extensible, whether it be models, uses, or even markets. But ASML’s last quarter results remind us that this growth is based on a physical infrastructure, constrained and above all highly concentrated.
Because behind each AI model, each data center, each GPU, there is an industrial chain of which ASML constitutes one of the obligatory points of passage. A chain whose capacity, today, no longer keeps up.
Frictionless growth…apparently
The announcement of ASML’s very good results during the last quarter must nevertheless be understood as a warning. Its two representatives, Christophe Fouquet, CEO of ASML, and its financial director Roger Dassen, expressed a market reality which leaves no room for ambiguity: “supply will not be able to meet demand in the foreseeable future”. Even more, “our customers tell us that they have already sold all of their capacities for 2026, and that the constraints will continue beyond that”.
Thus the semiconductor industry, long characterized by cycles of expansion and contraction, is entering a phase where demand permanently exceeds supply. This tension does not only affect cutting-edge components, but runs through the entire chain, from advanced logic to memory.
The real bottleneck
While the media attention is focused on chip designers and GPU makers, the real sticking point is upstream. ASML is today the only player capable of producing the EUV lithography machines necessary for the manufacture of the most advanced semiconductors, and without this equipment, no increase in capacity is possible.
The global production of advanced chips therefore depends on the pace of delivery of a limited number of machines, the manufacture of which itself requires a complex and difficult to compress supply chain.
Thus, ASML plans to deliver at least 60 EUV systems in 2026, then 80 in 2027. Although these volumes are growing significantly, they remain insufficient in view of the demand expressed.
A physical constraint
Unlike software, the semiconductor industry cannot replicate itself instantly. Each EUV machine concentrates thousands of components, cutting-edge technologies and years of development. Its production cannot be accelerated without compromising its reliability. The increase in capacity is, therefore, gradual and dependent on a global supply chain.
De facto, ASML does not intentionally slow down the AI, but, as a forced passage point, it becomes a real bottleneck.
Intensification rather than expansion
Faced with this constraint, of course the industry adapts. Rather than only increasing volumes, it seeks to optimize each unit produced. Christophe Fouquet talks about “higher lithographic intensity”, that is to say an increased use of lithography technologies for each chip manufactured.
This development has a paradoxical effect, because it reinforces the dependence on ASML. The more complex the chips, the more manufacturing steps they require, and the greater the demand for equipment.
Growth already arbitrated
Another determining element is that future capacities are already largely allocated. ASML’s customers (memory founders and manufacturers) secure their equipment over several years, backed by downstream contractual commitments. The growth of artificial intelligence is therefore not only constrained but it is already, in part, organized.
This point is central because it means that access to computing no longer depends solely on innovation or demand, but on the ability to position oneself upstream, in long and capital-intensive investment cycles. Therefore, only players capable of absorbing this constraint (hyperscalers, large manufacturers, States) benefit from the majority of resources.
