JPMorgan and Citi Both Raise Global Semiconductor Equipment Forecasts on Same Day: Memory Cost Cycle as Core Driver of Equipment Upcycle
Claire Weston
JPMorgan and Citi both raised their global wafer-fab equipment (WFE) forecasts on the same day, reaching the same core conclusion: this equipment cycle is almost entirely AI-driven, and memory — especially NAND — has moved from supporting role to lead driver.
How much did the two banks raise their forecasts?
JPMorgan lifted its 2026 and 2027 WFE growth estimates from 21% and 18% to 28% and 29%, and added a new 16% growth forecast for 2028.
Citi's bull case is more aggressive: WFE reaching $145 billion, $170 billion, and $250 billion from 2026 to 2028.
This means → two major banks are placing their bets in the same direction almost simultaneously — AI-driven equipment expansion is far stronger than previously priced in.
Why has NAND suddenly become the headline?
AI inference generates ballooning KV caches — intermediate data the model holds during operation. HBM and DRAM are expensive and supply-constrained, so workloads are migrating toward higher-capacity, lower-cost NAND.
Nvidia has already cut DRAM capacity in half in its next-gen Vera Rubin NVL72 system, confirming the shift.
Citi estimates that bridging the DRAM gap requires 2 to 4 new NAND fabs, translating to $15–30 billion in NAND equipment demand. In plain terms = almost no one had this number in their models — it is an "unpriced surplus" for the equipment sector.
How far are cloud giants stretching their wallets?
JPMorgan raised its 2026 capex growth forecast for the top four US cloud companies from 63% to 80%, and 2027 from 40% to 50%. In dollar terms: over $575 billion in 2026, over $860 billion in 2027.
The more striking figure: historically these four companies spent 25–30% of operating cash flow on capex. This year that ratio rises to 82%; next year it is projected at 94%. In plain terms = they are emptying their pockets.
Memory's share of cloud capex is jumping from a prior single-digit-to-15% range to roughly 50% in 2026. This reflects a structural re-ranking — memory is no longer a GPU accessory but a co-equal investment pillar.
Can one variable — AI — really carry the whole cycle?
Citi's three-scenario model shows AI-related demand reaching 65% of total WFE by 2028, while China and non-AI legacy demand (consumer electronics, industrial, auto) contribute near-zero annual growth.
This means → the traditional equipment-cycle logic — how many phones ship, how many auto chips are needed — is effectively broken this round. AI is the sole growth engine.
The risk is equally concentrated: if AI inference demand growth slows, the entire forecast framework faces repricing.
Where is the bottleneck — not money, but space and people?
JPMorgan identifies the binding constraint for 2026 capacity ramp-ups as cleanroom space and engineering talent, not equipment procurement budgets. Existing fabs will be reshuffled first; large-scale cleanroom buildouts begin only in 2027.
Supply is deliberately restrained: Kioxia says its current fabs can meet demand through 2028 or even early 2029; new fab investment starts only in late 2028. This means → the market stays tight — not because demand is too hot, but because supply is released slowly.
On advanced packaging, TSMC's next-gen CoPoS — a chip-level packaging interconnect — is running behind schedule. The near-term fix: extending CoWoS life and expanding reticle size (up to 14× by 2028). CoWoS monthly capacity is planned to grow from 115,000 wafers at end-2026 to 220,000 at end-2028.
What other demand sources are worth tracking?
Samsung and Intel foundry are showing signs of life. Intel foundry is gaining traction with Google and Apple. Samsung is accelerating its Taylor fab, collaborating with Google on a 2nm memory I/O die for TPU v10, reportedly negotiating CPU foundry work with AMD (target: 2028 mass production), and in talks with BYD on 2nm/4nm autonomous-driving chips. In plain terms = if TSMC's near-monopoly loosens and second sources are reactivated, equipment demand gets another layer.
China market: domestic equipment makers report that memory is the primary order driver this year, with YMTC and CXMT as the main contributing customers. A useful conversion: each additional 10,000 wafers/month of capacity requires roughly two electroplating tools.
The risk: if this wave of domestic capacity lands all at once and creates oversupply, pricing pressure becomes the next inflection point to watch.
Content is for reference only, not financial advice.