Some datapoints from a memory supply-chain expert call a few weeks ago: → Conventional DRAM may now be competing with HBM for wafer allocation on economics. The call puts HBM LTA gross margins at ~58% and argues that tight conventional DRAM supply can make marginal capacity more attractive outside HBM. → The contracting model is getting much harder. Older LTAs were generally 2–3yr price/volume agreements with limited upfront commitment, while newer structures increasingly involve 5yr+ take-or-pay terms, deposits and price floors/ceilings. The bigger shift appears to be guaranteed volume and downside protection for suppliers rather than locking today's elevated prices for five years. → The call identifies Google, Meta, Microsoft and Amazon as Micron’s four largest contracted customers, with Apple, Alibaba and Tencent in the next tier. Automotive customers cited include GM, Ford and Tesla-related suppliers. → Suppliers are reportedly exercising much tighter control over the distribution channel. Distributors may be required to provide end-customer sales records, while manufacturers can conduct warehouse checks and potentially revoke allocation or authorisation if inventory is being hoarded rather than sold through. → SanDisk is said to be requiring NCNR terms alongside ~30% upfront deposits from some distributors, effectively pushing parts of the channel toward much harder long-term supply commitments. → Customer importance can also differ materially by supplier. The call gives an example of the same distributor being Tier 3 at Micron but Tier 1 at SanDisk, reflecting SanDisk’s greater reliance on large retail/channel partners and resulting differences in allocation and commercial terms. → HBM4/HBM4E is also being used as a pricing reset. The call cites HBM3/3E contracts around ~$530–550 per device versus ~$800–1,400 for HBM4/4E. The exact magnitude is difficult to verify given pricing varies materially by stack, density, customer and contract period. → Interestingly, the expert was less convinced that tight memory supply necessarily proves equally tight underlying AI demand. Data-center capacity can be built ahead of available power and utilization, making it difficult to distinguish actual consumption from capacity being secured ahead of future demand.