CoWoS bonds the GPU die to its HBM stacks on a silicon interposer. No interposer, no high-bandwidth memory, no usable accelerator. Because nearly all leading-edge capacity sits with one supplier, each hyperscaler's GPU allocation is rationed by how many packaging slots it can secure — making this the true rate-limiter on AI compute, upstream of the chips themselves.
2026–27SoIC + CoWoS hybrid lines scale for next-gen accelerators
End 2027Capacity targeted to ~170k wafers/mo; glass-substrate pilots could relieve interposer scarcity
What would loosen it
Faster-than-guided capacity from ASE and Amkor, a demand air-pocket if GPU orders soften, or a shift to panel-level / glass substrates that sidesteps silicon-interposer scarcity. None look likely to bite before late 2026.
Latest developments
2026-01-28
Analysis: even with $56B 2026 capex, supply-demand gap won't fully close until 2027