SMIC: AI Memory Supply-Demand to Balance by Mid-to-Late 2028, Computing Chips Still in Shortage
nashnova research
SMIC co-CEO Zhao Haijun told the Q3 earnings call that AI-related memory chips should reach rough supply-demand balance by mid-to-late 2028, while compute and supporting chips will still be in shortage — signaling that this AI chip upcycle may last longer than the market expects.
How long before the AI chip supply gap closes?
Based on announced data-center and GPU deployment scales converted to wafer area, 2027 will be extremely tight, with the gap extending into mid-2028.
By mid-to-late 2028, announced memory capacity additions will outpace compute and supporting chips — AI memory is likely to balance first.
This means → the compute-chip shortage cycle runs longer than memory, and equipment makers and foundries retain pricing power for an extended period.
Zhao said AI is still in its early stages. Enterprises buying compute power remain undersupplied; many are building their own small data centers because cloud-compute pricing and availability are unstable.
Utilization at 97.3%, quarterly shipments past 3 million — how full is SMIC?
Q3 shipments crossed 3 million 8-inch-equivalent wafers for the first time, up 5.7% quarter-on-quarter; blended ASP rose 1.8% QoQ.
Utilization climbed to 97.3%, up 3.6 percentage points QoQ — a multi-year high even as monthly capacity added 20,000 8-inch-equivalent wafers in the quarter.
In plain terms = SMIC normally reserves 5% of capacity for engineering wafers and R&D. This quarter, customer orders hit a multi-quarter peak with rush-delivery requests, so nearly all capacity went to volume production.
Behind the pull-in — which segments are rising, which are falling?
Chips supporting AI data centers and enterprise compute workstations remain in persistent shortage. Zhao's exact words: "The gap between demand and capacity is still widening, not narrowing."
Auto and industrial demand recovered with no seasonal pullback; mid-to-high-end connectivity and IoT orders held steady; LED display driver ICs and fingerprint ICs are also recovering.
Smartphones and low-end consumer products stayed weak — Q3 revenue fell 8% and 1% QoQ respectively. This reflects AI demand crowding out capacity and resources from non-AI segments.
Equipment lead times stretched to two years — what does that mean for expansion?
Shortage fears pushed the industry to pre-order equipment, including tools not urgently needed in the current year. Vendors now quote 16-to-24-month lead times; the tightest tools exceed 24 months.
Equipment planned for this year slipped to next year; next year's deliveries pushed into the year after — causing SMIC's quarterly capex to run persistently below plan.
This means → even if SMIC wants to accelerate expansion, the equipment bottleneck slows the pace. Industry-wide equipment hoarding is itself amplifying the shortage.
Full-year capex will come in slightly below the original forecast, but management stressed the delay does not affect this year's expansion plan — only the spending cadence.
Depreciation rising 30% a year — can gross margins hold?
Zhao said depreciation grows roughly 30% annually; when revenue grows at a similar pace, the depreciation-to-revenue ratio stays stable and the gross-margin impact is limited.
Q4 depreciation will rise per full-year guidance, but the ratio changes little.
Next year depreciation will rise again; the exact scale will be disclosed at the next earnings call. Neither executive indicated when depreciation will peak.
Can specialty memory prices keep climbing?
For specialty memory — including NOR flash, a type of non-volatile storage chip widely used in embedded devices — SMIC follows market pricing. The current round of increases is complete, and prices will hold steady for a period.
64Mb and below (low-density) products are roughly in balance with limited upside; high-density products remain undersupplied, with room for further increases if market prices keep rising.
Zhao noted DRAM prices have risen several-fold since Q1 2025, while specialty-memory increases have been comparatively mild. In plain terms = surging DRAM costs have already pushed smartphones priced below RMB 1,000 (about $149.53) out of the market — AI demand is squeezing out non-AI demand across the board.
市场有风险，内容仅供研究参考，不构成投资建议。