How much of the world’s advanced chip packaging and high-bandwidth memory does AI consume? Almost all of it.
We estimate the four largest AI chip designers consumed ~90% of global advanced packaging and HBM supply in 2025, suggesting these inputs were bottlenecks in 2025.
Frontier AI chips like NVIDIA’s B200 need three key inputs: advanced logic dies, high-bandwidth memory (HBM), and CoWoS advanced packaging, which integrates these components together on a single chip.
CoWoS packaging and HBM are almost entirely consumed by AI demand. But logic fabrication serves a much broader market — e.g. smartphones, laptops, and automotive — with AI chips consuming only 12% of supply in 2025.
However, as next-generation chips move towards more advanced process nodes, we may see AI chips claiming a much larger share of advanced logic capacity.
To reach these estimates, we drew on Epoch’s AI Chip Sales hub, chip technical specs, unit costs, and market size data. We also modeled manufacturing lags and inventory timing to make the figures as accurate as possible for CY2025.
For full methodological details and an interactive plot, check out the data insight on our site!
Very interesting piece, but I think it's not doing enough to justify *why* consumption share is (conclusive) evidence of a bottleneck. It *could* be the case that we observe high utilisation in resource X, but because production is more elastically scaled to actual/other bottlenecks.
My quick research:
- HBM: Almost certainly independently binding. Sources say 3-4x wafer intensity per GB, there are sold-out-through-2026 statements, price spikes in downstream DRAM
- CoWoS: Likely also independently binding, but evidence is TSMC's capex behaviour and customer allocation reports
- Logic fabs: Possibly binding at the frontier nodes too, but harder to see because TSMC has more pricing power and allocation flexibility there.