SemiAnalysis: AI Silicon Shortage — HBM Bottleneck and N3 Wafer Dominance
Synthesis history
7 versions, newest first.
-
Version 7 2026-06-05 03:03 UTC · 100 items
New items this pass are thin. The one substantive addition is confirmation that Intel's Crescent Island inference chip does not use HBM [^25082], which deepens the existing observation that Intel is competing entirely o…
-
Version 6 2026-06-04 02:29 UTC · 94 items
Two new angles this pass. SemiAnalysis corrected a potential misread of AMD's competitive position: rack photos of the MI455 and VR200 at CoreWeave and Microsoft are engineering samples with incomplete software stacks a…
-
Version 5 2026-06-02 18:48 UTC · 80 items
The most significant new development is SK Hynix's disclosure that even doubling wafer capacity within five years will not relieve memory supply tightness until at least 2030 [^23422] — a substantial extension of the sh…
-
Version 4 2026-06-01 18:39 UTC · 58 items
The most significant new material this pass is quantitative confirmation of the shortage thesis from TSMC's own financials: +41% YoY growth with all-time-high margins [^23115], foundry industry at a record $48.8 billion…
-
Version 3 2026-06-01 08:18 UTC · 52 items
The most significant new development is the extension of the TSMC shortage horizon beyond 2027 [^23032], pushing the supply constraint from a near-term bottleneck into a potentially multi-year structural regime — a mate…
-
Version 2 2026-05-31 18:43 UTC · 43 items
The most significant new development is empirical confirmation of the HBM shortage thesis: sold-out order books at both SK Hynix and Micron through 2026 [^22728][^20307] move SemiAnalysis's central claim from forecast t…
-
Version 1 2026-05-31 02:14 UTC · 28 items
The primary constraint on frontier AI hardware has migrated from GPU availability to High Bandwidth Memory (HBM) wafer supply, with CoWoS packaging no longer the binding bottleneck [^22548][^22656]. SemiAnalysis resear…