AI Demand Is Squeezing Conventional Memory as DRAM and NAND Prices Surge

Bank of America’s forecast is best understood as a supply-allocation crisis arriving at the worst possible time for buyers. Memory manufacturers are prioritizing high-bandwidth memory (HBM) and other AI-related products, while PC, smartphone, console, and server makers are preparing for the traditional year-end production season. That combination is tightening conventional DRAM and NAND and giving prices further room to rise. 17

The immediate forecast: another 10%–20% in September

Bank of America expects DRAM and NAND spot prices to increase by another 10%–20% in September 2026. The call follows a period of strong pricing momentum and renewed purchasing by device manufacturers ahead of peak-season production. 17

The underlying market data shows why the forecast is significant. TrendForce’s August 28 spot-market update listed a session average of $91.293 for 16Gb DDR4-3200 and $53.933 for 16Gb DDR5-4800/5600. These are chip-level spot references, not the prices consumers necessarily pay for complete memory kits.

That distinction matters. Spot prices, quarterly contract prices, module prices, and retail prices reflect different points in the supply chain. A sharp move in one does not translate immediately or uniformly into the others.

How severe is the shortage?

The clearest signal is not one dramatic headline percentage but the combination of elevated spot prices, scarce inventory, and manufacturers’ willingness to reserve capacity in advance. SK Hynix said its DRAM, HBM, and NAND output for 2026 had already been sold out, while customers were securing conventional-memory production slots into the following year. 2 The company described DRAM and NAND markets as extremely tight. 13

Consumer pricing provides another view of the stress. Tom’s Hardware’s U.S. price tracking listed a 64GB DDR5-6000 kit at $849, compared with a historical low of $159. The same snapshot showed 96GB kits at $1,799 and 128GB DDR5-6400 kits at $3,399.

Other retail reports have shown selected 64GB DDR5 kits above $1,200 and year-over-year increases approaching 500%. Those figures describe particular tracked configurations or listings, not every DDR5 product. The variation reinforces the need to treat retail examples as market snapshots rather than universal prices.

Why HBM is tightening ordinary DRAM

HBM is central to the shortage because it is not simply an additional product line that manufacturers can produce without trade-offs. HBM for AI accelerators requires substantial manufacturing and packaging resources. One estimate cited by The Motley Fool, drawing on TrendForce data, suggests HBM could use about 30% of the DRAM production capacity of the three largest memory makers in 2027 while accounting for only 13% of their output bits. 4

That imbalance leaves less capacity available for conventional products. DDR5, legacy DDR4, mobile DRAM, and server DRAM must compete for manufacturing resources while AI customers secure supply through long-term commitments. As production mix shifts toward higher-value HBM, conventional memory becomes less flexible and more exposed to even modest demand increases.

The pressure is especially acute for DDR4. Although it is an older standard, many existing PCs, servers, industrial systems, and other devices still depend on it. Manufacturers have less strategic incentive to expand legacy capacity when HBM and newer products offer stronger returns. That can make an older memory standard unusually expensive during a shortage.

Why NAND is rising too

The AI build-out requires more than accelerator memory. Data centers also need high-capacity, fast storage for datasets, checkpoints, and inference infrastructure. That demand supports enterprise SSDs and improves the outlook for NAND, even though NAND’s market dynamics differ from DRAM’s. 5

NAND is also receiving support from conventional electronics demand. OEMs are bringing forward purchases as they prepare phones, PCs, consoles, and other products for the year-end season. Bank of America’s September view therefore combines AI-related infrastructure demand with a shorter-term rush for inventory across the electronics supply chain. 17

The evidence is not equally strong across every NAND category. Bank of America’s later outlook cited by Kaohoon International anticipated further fourth-quarter increases, followed by stable or mildly correcting prices through 2027 and possible declines in 2028. That suggests NAND may recover alongside DRAM, but its price path could be more uneven by product type.

What it means for PCs, servers, and consumers

Higher memory costs create several possible outcomes for device makers:

• higher prices for PCs, gaming systems, servers, and storage products;
• lower-memory base configurations;
• fewer promotions and discounts;
• margin pressure for manufacturers that cannot immediately pass through costs; and
• delayed or redesigned product refreshes.

The largest near-term exposure is in memory-heavy products and systems that still require DDR4. Consumers building or upgrading PCs may encounter especially wide price differences between capacities and speeds, because high-capacity kits have experienced some of the sharpest increases in available retail tracking.

However, a retail kit price should not be read as a direct measure of the cost of a DRAM chip. A kit includes module manufacturing, testing, distribution, retailer margins, inventory conditions, and sometimes a premium for a particular speed or capacity. The chip-level spot data and retail examples show the same direction of stress, but they measure different things.

Nvidia benefits from demand but faces a supply constraint

The memory shortage is broadly consistent with continued demand for Nvidia’s AI accelerators: strong HBM demand indicates that data-center operators are still competing for AI infrastructure. SK Hynix’s position as an Nvidia supplier and its reported advance bookings illustrate how tightly accelerator demand is linked to memory availability. 2

But the same shortage can constrain Nvidia’s system shipments. If HBM supply, advanced packaging, or related memory components cannot keep pace, accelerator demand cannot automatically become completed and shipped systems. The available evidence supports viewing memory as both a demand signal and a potential bottleneck; it does not establish a current Bank of America Nvidia rating or price target.

Why SK Hynix is associated with a “profit super-cycle”

The super-cycle thesis is broader than an HBM boom. SK Hynix is positioned to benefit from rising prices across HBM, conventional DRAM, and NAND at the same time. The company reported that its 2026 output across those categories was sold out and planned to begin supplying HBM4 in the fourth quarter. 2

Analysts have described the current setup as unusually broad because HBM expansion restricts commodity-memory supply while AI storage demand helps support NAND profitability. 5 6 In that environment, high utilization, advance customer commitments, and stronger average selling prices can support unusually strong earnings for leading suppliers.

The thesis still carries risk. Memory is a cyclical industry, and new capacity, weaker AI infrastructure spending, or a slowdown in electronics demand could eventually change the balance. The question is timing: the available forecasts do not point to an immediate normalization.

When could memory prices fall?

Bank of America’s cited outlook anticipates additional increases in the fourth quarter, followed by stable or mildly correcting pricing through 2027. It projects potential DRAM and NAND declines in 2028, characterized as a healthier correction rather than a collapse.

That timeline is consistent with SK Hynix CEO Kwak Noh-jung’s warning that 2027 could be the industry’s worst-ever year for supply. 1 Other industry reporting has also described shortages lasting through at least 2027 as AI demand continues to outpace available capacity. 10

The practical conclusion is that broad price relief before 2027 appears unlikely under the current supply-and-demand assumptions. Even a 2028 decline depends on new capacity arriving on schedule and AI infrastructure spending growing more slowly. For buyers, the shortage is therefore not just a short-lived spike in consumer RAM prices; it is a structural contest between AI infrastructure and the rest of the electronics market.