Dylan Patel on Dwarkesh Patel: How EUV Tools Cap AI by 2030

AI compute scaling hits bottlenecks in chips, tools, and memory

1. A.Hyperscaler AI capex largely pre-pays long-lead items like turbines, land, and construction years ahead, so spend today does not translate linearly into gigawatts online this year.
2. B.GPU economics are shifting because chip value is increasingly set by the utility of today’s best models under scarcity, making older GPUs (e.g., H100) potentially more valuable now than years ago despite newer generations.
3. C.By the late 2020s the binding constraint to AI compute is expected to move upstream to semiconductor manufacturing capacity—ultimately ASML EUV tools and their intricate supplier network—rather than power or data center construction.
4. D.Falling back to older nodes (e.g., 7nm) is not a clean escape hatch because modern performance depends heavily on system-level factors like interconnect, memory bandwidth, and packaging, not just raw FLOPS or node shrink.
5. E.A major memory crunch (HBM/DRAM) is projected to raise prices, reallocate wafers away from consumer devices, and potentially shrink low-end phone/PC volumes while AI captures a growing share of global memory output.