Mobile NAND demand is expected to be flat year over year in 2026, while memory-industry revenue is forecast at approximately $550–570 billion for the year.
The central capex question is not whether AI infrastructure is being funded, but whether NAND suppliers can add capacity and improve yields without converting scarcity into structural oversupply.
2026 Demand and Spending Set the Capex Test
Memory-industry revenue is forecast at approximately $550–570 billion in 2026, compared with a trajectory of $800–850 billion by 2027. These figures establish the scale of the AI-linked spending cycle, but they do not by themselves justify unrestricted NAND capacity additions.
Memory revenue is projected at approximately $550–570 billion in 2026, or roughly 55–58% of the global semiconductor market, after the industry moved from about $85–90 billion at its 2023 trough toward a projected $800–850 billion by 2027 [S2]. The unusually large revenue share makes memory capex discipline consequential across the semiconductor supply chain: each producer must balance prospective market growth against the risk that simultaneous expansion creates excess bits, weaker average selling prices, and lower utilization.
The wider hyperscaler estimate also supports a clear distinction between funding capacity and consuming capacity. Hyperscaler budgets can finance servers, data-center buildings, networking, power systems, and memory procurement, so a larger capex total does not translate one-for-one into NAND wafer starts. A detailed breakdown of hyperscaler AI capex in 2026 is therefore more useful for memory planning than the aggregate headline alone.
Capex Discipline Matters More Than a Demand Headline
Mobile NAND demand is expected to be flat year over year in 2026 despite DRAM pricing being on track to increase roughly 275–300% from 2025 through 2027. This is the practical difference between capacity expansion and overspending: the first follows qualified demand, while the second follows a temporary revenue surge.
Three controls should anchor supplier planning. First, incremental wafer capacity should be tied to visible consumption rather than the theoretical maximum implied by AI use cases. Second, capital spending on process migration and yield improvement should be separated from spending that adds total wafer starts. Third, procurement and inventory signals need to be reviewed by product segment because mobile NAND demand is expected to be flat year over year in 2026 [S2].
That mobile result is an important constraint on the headline growth rate. NAND demand is not moving as one synchronized market: enterprise SSD demand linked to AI infrastructure can accelerate while smartphone and PC economics deteriorate. Mobile NAND demand is expected to be flat in 2026, the first annual flatness identified in the NAND market’s history, as OEMs reduce specifications or reduce shipments because memory costs are harder to absorb [S2].
Demand Type Determines the Capacity Decision
AI storage demand has a much larger theoretical ceiling than realized consumption. Capacity planners should not convert that upper bound into committed NAND investment.
Workload type changes the relevant storage mix. Persistent context, long-term memory, and key-value-cache storage can expand data volumes, while image, video, and multimodal workloads require additional capacity beyond the text-only calculation [S2].
For NAND suppliers, that spending supports enterprise SSD demand but does not remove product-level constraints. Capacity decisions must distinguish client flash, consumer flash, and enterprise SSD layers because their qualification cycles, endurance requirements, interface generations, and customer economics differ. Flat mobile NAND demand therefore supports selective capacity and technology spending, not a uniform expansion across every NAND product family.
Top Makers Face Different Discipline Tests
Top memory suppliers cannot be evaluated through a single capex ranking because their planned spending, clean-room schedules, technology portfolios, and exposure to AI-related products differ. The available evidence supports comparing the direction and timing of investment, but it does not provide a verified 2026 company-by-company NAND capex table for Samsung, SK hynix, and Micron. [S2]
Because long construction lead times allow 2026 capex decisions to influence supply beyond 2026, some spending is committed before meaningful output is available.
A disciplined comparison should therefore use four criteria: percentage of spending tied to additions versus process improvement, expected wafer-output timing, exposure to high-value enterprise products, and balance-sheet capacity under weaker pricing scenarios. Without those disclosures, the wrong conclusion would be to treat the largest announced investment as the strongest competitive position. Larger capex can create scale, but it can also depress utilization if qualification and customer demand lag.
Price Signals Can Expose an Oversupply Build
Average selling prices, rather than unit growth, are driving the memory revenue expansion in the available forecast. Semiconductor unit volumes are projected to grow roughly 4% annually through the end of the decade, while revenue is compounding at 14%; estimated semiconductor average selling price reaches about $0.91 in 2026 and is tracked toward $1.22 by 2030 [S2].
That spread is the clearest warning against indiscriminate capacity additions. A supply response calibrated to forecast NAND demand may look conservative against a theoretical AI ceiling, yet it may be excessive if pricing begins to weaken because too many bits reach the market at once. The first practical warning signs would be falling NAND contract prices, rising customer inventories, slower enterprise SSD qualification, and supplier utilization falling below the level required to justify new clean-room spending.
Technology migration can add complexity to this calculation. Conventional DRAM average selling prices are projected at approximately $1.20–1.30 per gigabit in the forecast period, approaching HBM3E 12-Hi pricing, while DDR5 scarcity pushes conventional product pricing upward [S2]. The same price pressure can improve cash generation for NAND producers, but it can also accelerate customer design changes and make demand harder to forecast, increasing the value of staged capex commitments.
Procurement and Funding Data Need Separate Treatment
Hyperscaler capex is the strongest observable demand-side financing signal, with one estimate placing 2026 spending across Amazon, Microsoft, Alphabet, Meta, and Oracle at $602 billion and describing a wider 75% allocation to AI-related uses [S1]. Another forecast estimates top-five hyperscaler capex at $720–750 billion in 2026 and $800–840 billion in 2027 [S2]. The differing totals reflect forecast scope and methodology, so they should not be combined into one synthetic number.
For NAND capex analysis, the defensible use of these estimates is comparative. A $602 billion baseline and a $720–750 billion alternative both describe a very large funding pool, while the $720–750 billion range provides a useful 2026 planning band [S1][S2]. Suppliers should track where that money flows across compute, memory, networking, buildings, and power rather than inferring a proportional increase in NAND demand.
The broader semiconductor context also requires caution. The NAND market grew by 67.10% in 2024, and 2026 capacity spending was linked in analysis to semiconductor-market growth during 2025, illustrating how cyclical revenue can pull capital plans forward [S3]. That historical acceleration is evidence of cycle sensitivity, not a fixed rule for 2026.
Decision Framework for NAND Capacity
A disciplined 2026 plan should allocate capital first to bottleneck relief and product qualification, then to output additions justified by contracted or forecast consumption. A supplier can stress the plan under lower, base, and upper cases, including flat mobile NAND demand and the risk that server storage qualification takes longer than planned [S2].
The decision threshold is explicit: add NAND wafer capacity when expected bit consumption, acceptable pricing, and usable output from funded projects justify the investment after a pricing downturn. Do not use the theoretical AI storage ceiling, a hyperscaler capex headline, or a fab announcement as a substitute for that test.
Track four indicators through 2026: NAND average selling prices, customer inventory days, enterprise SSD qualification schedules, and supplier capital spending separated between maintenance, process migration, and new wafer starts. Together, these signals reveal whether the cycle is absorbing capacity or merely financing it.
The combination of flat mobile demand, projected memory prices, and long construction lead times favors staged commitments over a capacity race.