NAND prices could fall in 2027. AI demand does not repeal the memory cycle.
A weaker NAND forecast can coexist with strong AI spending. Sandisk’s revenue mix shows why investors need to look beyond the industry label.
5 min read
AI can need more memory while memory investors have a worse year. There is no contradiction: demand is only half of a market, and “memory” covers products with different customers, economics and supply constraints.
That is why the latest NAND flash forecast from TrendForce deserves more attention than a broad verdict on whether the AI boom is over. The research firm expects overall contract prices in January to March 2027 to be flat to 5% lower than in the preceding quarter, with all product categories entering a downward price cycle from April to June.
Its explanation is specific. Reduced enterprise solid-state drive orders from Chinese cloud providers are expected to outweigh additional North American demand. Meanwhile, it forecasts NAND bit supply growing by 25-30% in 2027.
More AI infrastructure can therefore coexist with excess supply in part of the storage market. For shareholders, the useful question is not whether AI remains important. It is how much of the industry's growing activity turns into durable profit for the particular company they own.
First, separate the memory markets
NAND flash provides persistent storage, including the chips inside SSDs. DRAM is working memory. High-bandwidth memory, or HBM, is a specialised stacked form of DRAM that moves data quickly alongside powerful processors, as Micron's technical overview explains.
An SSD storing data and HBM feeding an accelerator both belong in an AI system. They are not interchangeable products. A shortage of one does not prevent oversupply of the other.
Even within NAND, “AI demand” is too broad a label to settle the investment argument. Enterprise storage and consumer devices have different purchasing decisions. A cloud provider expanding a data centre may be buying heavily while a household postpones a new phone or laptop.
For an investor, this creates a fairly basic research task that the AI label can obscure: identify what the company actually sells, which customers buy it, and what determines the price those customers will pay next time.
The answer can be quite different for a specialist supplier, a diversified chipmaker and a hardware reseller. A useful demand trend does not travel through the supply chain with the same effect on everybody's margins.
Sandisk's results show the split already matters
Sandisk's latest reported quarter, which ended on 2 October, provides a concrete example. Total revenue rose 23% from the previous quarter to $11.031bn, but its end markets moved at very different speeds.
The category names need care. Sandisk counts flash sold to PC and phone makers under Edge, alongside other equipment markets. Consumer covers retail and other end-user products. A weaker Consumer line therefore tells us little on its own about laptop or handset sales.
Sandisk’s growth came mainly from data centres
Revenue by end market, US$bn
Sandisk revenue increased from $8.965bn in Q4 FY2026 to $11.031bn in Q1 FY2027. Data-centre revenue increased while consumer revenue fell.
Quarterly revenue mix
US$bn
- Datacenter
- Edge
- Consumer
Reported end-market revenue. Edge includes supply to PC and phone makers; Consumer covers retail and other end-user products. Q1 FY2027 ended 2 October 2026; Q4 FY2026 is the preceding fiscal quarter.
Source: Sandisk Q1 FY2027 results
View data
| Category | Datacenter | Edge | Consumer | Total |
|---|---|---|---|---|
| Q4 FY2026 | $2,977,000,000 | $5,432,000,000 | $556,000,000 | $8,965,000,000 |
| Q1 FY2027 | $4,802,000,000 | $5,832,000,000 | $397,000,000 | $11,031,000,000 |
The data-centre increase accounted for about 88% of the net sequential revenue gain. That is a calculation from the reported figures, not a measure of profit contribution: revenue does not tell us the margin earned in each market.
The company also forecast $12.2bn-$12.6bn of revenue for the following quarter. Its near-term outlook remained strong when those results were issued on 29 October. That forecast and TrendForce's later warning can both make sense because they cover different periods and measure different things.
Company revenue combines prices, volumes and product mix. An industry price forecast concerns one part of that equation. Neither should be used as a shortcut for the other.
Supply can grow without a new factory opening
Investors often picture supply as a count of factories. For memory, the amount of storage produced also matters. More bits from a given manufacturing footprint can increase what the market must absorb even before an entirely new facility arrives.
TrendForce points to higher-layer NAND and greater use of QLC, which stores four bits per cell, as sources of additional bit output. It estimates industry capital expenditure will reach $29.9bn in 2027, up 26.7%.
Better manufacturing technology can be good for a producer and difficult for the industry's pricing at the same time. A supplier may reduce its cost per bit, but if competitors do the same and supply grows faster than demand, customers can capture part of that improvement through lower prices.
This is the uncomfortable feature of an investment story built around both rapid capacity expansion and permanently scarce supply. The first can undermine the second.
The counterargument is that suppliers can respond. They can slow investment, alter the mix of products or accept less volume in pursuit of better pricing. Growing applications can also absorb more storage than forecasters expect. The downturn is a forecast, not a fact already recorded in next year's accounts.
Spot prices are a signal, not everybody's selling price
TrendForce reports a recent fall of more than 20% in spot prices, while still forecasting an 8-13% quarterly rise in fourth-quarter contract prices. Those figures concern different markets and should not be blended into one claim that all NAND prices have fallen by a fifth.
A chipmaker's realised selling price depends on what it sells and the agreements under which it sells it. A sharp move in immediate-delivery trading may provide a warning about conditions without translating one-for-one into revenue on longer arrangements.
For buyers, falling wholesale prices also do not guarantee an immediate matching discount on a finished SSD. Existing inventory, other components and the retailer's pricing matter. The practical benefit is potentially more choice and negotiating room when stock and contracts turn over.
Timing affects suppliers too. Orders placed early during a shortage can make one period look stronger and leave a quieter period behind. The same issue appeared in Softcat's account of orders brought forward during memory shortages. Strong current sales and a less comfortable future can belong to the same story.
A low earnings multiple can hide a high assumption
A cyclical company can look cheapest when its earnings are unusually strong. Divide the share price by peak profit and the resulting multiple may look modest. If profit then falls, that apparent bargain can disappear without the share price moving at all.
Sandisk's results show why that assumption matters. Its company-wide GAAP gross margin reached 85.0% in Q1 FY2027, against 29.8% a year earlier. Neither figure is automatically the right one for a long-term valuation. The investor's job is to explain which pricing, mix and cost assumptions could sustain the chosen margin.
A sensible comparison would put current earnings beside a less favourable price environment, allowing for changes in volume, mix and cost. A business with lower manufacturing costs or better customer relationships may weather that environment well. The existence of those advantages needs evidence beyond an association with AI.
The next useful signals are contract pricing, inventory and the relationship between bit shipments and revenue. If shipments keep rising but revenue growth fades, the mix and price explanations become increasingly important.
AI demand remains part of the case for storage investment. It does not repeal the memory cycle. The companies worth understanding are those that can earn a satisfactory return when supplying the next bit becomes easier.
Comments
No comments yet - start the conversation.