Micron CEO: 75% of 2027 Supply Already Sold, No End in Sight for Supply-Demand Gap
nashnova research
Micron CEO Sanjay Mehrotra says 75% of 2027 supply is locked up and the company is already negotiating 2028 deliveries, with no timeline for supply to catch demand — a signal that memory is shifting from a cyclical commodity to a seller's market resembling long-term energy contracts.
What does it mean to sell 75% of capacity two years early?
Mehrotra told investors that 75% of 2027 supply is already sold, with 2028 negotiations under way.
This means → customers are not placing orders — they are racing to secure capacity two full years before delivery.
Four forces are accelerating at once: larger AI models, rising agent usage, longer context windows, and broader application rollouts — each one directly increases memory consumption.
In plain terms = memory used to be "buy what you need"; now it is "lock it in early or risk getting none."
Why are these contracts more than paper promises?
Unlike the old norm — one-year deals based on customer forecasts — Micron's current supply commitment agreements (SCAs) are five-year contracts with take-or-pay clauses.
Customers must post cash deposits and make financial commitments. Micron has signed 26 SCAs involving $32 billion in upfront cash and financial obligations.
This means → customers are betting real money that AI compute demand will not reverse over five years; once signed, the economic cost of exit is steep.
Why are defaults unlikely this time?
Asked what happens if a customer tries to walk away, Mehrotra said he "absolutely does not believe" past-style defaults will recur.
His logic: memory's role has changed — it is now the key component that helps customers lower per-token inference cost and raise ROI on compute platforms, no longer a fungible commodity part.
In plain terms = memory used to behave like a bulk commodity — prices dropped, customers broke contracts and bought cheaper on the spot market. Now memory is the performance bottleneck of AI systems; losing supply means the entire compute chain stops.
Why does HBM complexity lock in the competitive landscape?
CTO Scott DeBoer called HBM — high-bandwidth memory that vertically stacks multiple DRAM layers using through-silicon vias, then bonds them to a foundry-made logic die — "one of the most complex semiconductor products ever manufactured."
The DRAM industry shrank from 10–15 competitors around 2000 to effectively three today (Samsung, SK Hynix, Micron).
This reflects a self-reinforcing loop: the technology is hard enough that only three survive, and because only three survive, the long-term contract model works.
When does 3D-stacked DRAM arrive, and what does it change?
DeBoer gave a timeline: market entry in 2028–2029, broad adoption by 2030–2031.
3D-stacked DRAM — placing multiple DRAM layers directly on top of a GPU or accelerator chip — maximizes bandwidth and minimizes transfer power, while allowing custom logic to be integrated into the memory product.
This means → once a customer adopts a custom base die deeply integrated with an accelerator, switching costs become extremely high — stickiness even greater than today's HBM.
The biggest engineering challenge is thermal management: temperature-sensitive DRAM sitting directly on a hot GPU creates irregular hotspots, requiring co-optimization of thermal profiles, bonding materials, and cooling systems at the design stage.
How much optimism is already priced in?
Current market optimism is largely reflected in the scale of long-term contracts and the 75% lock-up ratio for 2027.
The real proof point is whether 3D DRAM can ramp to volume on schedule in 2028–2029, and whether AI infrastructure demand continues to support the existing pricing framework at that point.
In plain terms = the market already knows the good news. The next critical question is whether both technology and demand can deliver simultaneously two years from now.
市场有风险,内容仅供研究参考,不构成投资建议。
