Micron tech workshop (CTPO DeBoer): management guides 'memory undersupply persists for next 2 years, no supply/demand balance in sight'…

training (per-GPU HBM up 3.6x, from the H100's 80GB to the B300's 288GB), inference (per-server memory doubles in three years), agents (context length grows 30x per year) and physical AI (L4+ robotaxi DRAM usage is 10x+ that of an L2+ car). The structural kicker on the supply side is vertically stacked HBM (vHBM, memory stacked directly on the GPU), which launches in 2028. DeBoer calls its 'trade ratio high', meaning the same capacity consumes meaningfully more wafers. Every HBM generation gets more wafer-hungry, and that is the supply-side anchor of the structural shortage call. The node roadmap puts the next generation of ramps in H2 2027. 1δ DRAM at Boise brings a bit-density gain of about 25% and is designed specifically for HBM5 and vertical HBM (Samsung's equivalent is reportedly due at end-2027). 1ε uses wafer bonding, and development is 'going well'. 1ζ and 1η are still in path exploration, with 3D DRAM slotted in as a parallel route between 1ζ and 1η. The HBM annual cadence runs HBM4 (2026, in mass production, 2.8TB/s per stack), HBM4E (2027), HBM5 (2028), HBM5E (2029) and HBM6 (2030), with three parallel product lines (standard, custom with logic on the base die, and vertical). G11 NAND has more than 400 layers, wafer-to-wafer bonding and 2x the bit density of G9, with mass production in H2 2027. There are also DDR6 (2x the bandwidth of DDR5) and LPDDR6 (35% better energy efficiency). The implication is that the HBM4E repricing thesis for 2027 discussed last week now has a matching node timetable, but 1δ and G11 both ramp in H2 2027, so H1 2027 supply gets little relief. DeBoer responded directly to the CXMT 4F² threat. He says 4F² gives almost no density advantage over 6F², is doable at legacy nodes but is costlier and more complex, and faces 'fundamental' scaling problems at advanced nodes. That aligns with last week's Mizuho view that CXMT's 4F² is at research-demo level. The caveat is that this is one competitor talking down another, and real CXMT progress needs mass-production data. On Taiwan labour: the Taichung union (Micron's largest Taiwan site) signed a labour agreement with the company on 21 Oct and withdrew the second mediation that was scheduled for today. Terms were not disclosed. The Taoyuan union's negotiation at the city government on 21 Oct ended with no result (the union says the company gave no specific numbers, and the profit-sharing plan has been pending for nearly a month). Taoyuan runs a 24-hour warning strike at the plant gate from 9:00 today to 9:00 tomorrow, with a full-member vote on the company's all-hands proposal at 10:00 today. On scale, Taoyuan has 2,258 strike-eligible members against Micron Taiwan headcount of about 15,000. Taichung, the biggest risk, is now resolved, and Taoyuan is a one-day warning action with limited production impact. Watch today's Taoyuan vote and whether Taichung's bonus terms push up company-wide labour costs.