Calibrating the Right Policy Response to the AI-Driven Memory Chip Shortage

The AI-driven surge in data center demand is revealing some of the fragilities in the global semiconductor industry's production capacity. The memory chip market—integrated circuits that store data—offers a vivid example. Production capacity simply isn't enough to meet rising demand driven by data centers. Consequently, prices for some memory chips will likely more than triple in 2026. The semiconductor industry expects this shortage to last at least until 2027, potentially until 2030. This shortage directly affects consumers—Apple, for example, has already announced that “price increases are unavoidable” for its products. Depending on how Washington responds to the memory chip shortage, this moment could represent a unique opportunity to keep boosting U.S. semiconductor production capacity—but policymakers should avoid short-termism that could derail the industry’s long-term competitiveness.

Memory chips matter because they let computers store large volumes of information and rapidly supply data to processors, with different memory technologies serving different functions. Dynamic random-access memory (DRAM) functions as a computer’s short-term working memory, holding the instructions and data a processor needs immediately but losing them when power is removed. Flash storage (NAND) allows retaining data without power and serves as the storage medium in smartphones and solid-state drives. High-bandwidth memory (HBM) is a specialized form of DRAM that stacks memory dies vertically and places the resulting stack next to an AI accelerator, greatly increasing how quickly data can reach the processor.

These capabilities make memory chips ubiquitous in most modern electronics, such as mobile devices, automobiles, and gaming consoles. Counterpoint, a market intelligence company, estimates that DRAM and NAND represent 14 percent and 11 percent of the raw materials cost of a representative mid-end smartphone, respectively. The same analysis also found that memory chips could reach 43 percent of a low-end smartphone's bill of materials (BOM) costs. TrendForce, another market intelligence company, forecasts that DRAM and NAND will absorb 47 percent of major cloud providers’ capital expenditures in 2026.

The explosive growth of AI-related data centers is squeezing memory chip supply, driving up prices for many electronics that use them as inputs. The major memory makers—Micron, Samsung, and SK Hynix, also called the “Big Three”—have redirected or prioritized production capacity and investment toward HBM to meet AI demand, at the expense of some DRAM and NAND supply. (The Big Three account for 87 percent of the global DRAM market share.) This prioritization of HBM production over DRAM/NAND can be explained by three factors: HBM can be differentiated, helping de-commoditize production; HBM demand is more predictable; and HBM offers better margins.

Another important factor is that the memory chip industry is cyclical. By 2018, total industry revenue was forecast to increase 33 percent, but a year later, the sector contracted by the same share. During the pandemic, the demand for electronics (and therefore for memory chips) surged. When the boom reversed, Micron announced in 2022 that it had to cut production of DRAM and Flash Storage (NAND wafers) by 20 percent, later expanding the cut to 30 percent in 2023. Micron’s operating margins fell to negative 9 percent in 2023, while SK Hynix reported a negative 24 percent operating margin for the same period.

The industry's cyclical nature limits production capacity expansion. Supply moves slowly, demand moves quickly, and much of the DRAM and NAND market is commodity-like. This creates repeated phases of price surges, a race for investment, overproduction, price declines, stockpiling, and back to square one. A producer must spend billions of dollars on a fab, tools, process development, and customer qualification before the first shipment. Once a fab is online, however, it is often cheaper to continue production because most fixed costs have already been incurred. Suppliers therefore have an incentive to keep producing during a downturn. Demand, by contrast, depends on volatile end-market sales. This cyclical nature has taught incumbent companies to invest in capital with discipline—in other words, to be cautious about investing.

This industry's picture makes it tempting for policymakers to intervene: a highly concentrated market, a strategically critical input, and prices that are peaking, which might affect consumers.

However, the core challenge for U.S. policymakers is to distinguish genuine market failure from industry self-discipline. Price increases in consumer goods driven by higher input costs are of public interest, but any intervention should be cautious, given the complexity and importance of this market. First and foremost, Washington should avoid micromanaging the supply/demand balance.

One example is cherry-picking who gets first in line. In April 2026, Senator Bernie Moreno (R-OH) issued a letter urging the Department of Commerce to work with the Big Three “to ensure that the U.S. auto industry is fully supplied with memory chips.” While Senator Moreno calls for dialogue and a public-private agreement, the approach still tips the scale in favor of one particular sector.

To be clear, U.S. advanced manufacturing is strategically important, and ITIF has long advocated for smart, targeted industrial policies, including for the auto industry. However, forcing domestic memory makers to allocate production does nothing to increase the underlying supply, it merely shifts scarcity from one industry to another.

There are limited short-term options, but the best response is to maintain a stable, predictable, supportive policy environment while maximizing market flexibility and access. To that end, Washington should avoid semiconductor tariffs, which raise costs without adding near-term capacity. It should also keep export controls tightly focused on genuinely security-sensitive technologies rather than unnecessarily constraining commodity-memory supply.

Policymakers need to assess whether U.S. firms should source conventional memory from Chinese producers such as CXMT for non-sensitive applications or for China-market production, subject to appropriate security safeguards. Such a policy would need to be carefully calibrated. In August 2026, Commerce Secretary Howard Lutnick urged Apple not to buy Chinese memory chips because doing so could strengthen state-backed Chinese competitors and deepen U.S. dependence on suppliers that Washington already views as national-security risks. To source these memory chips, U.S. companies would need a U.S. government license, since the largest producers are sanctioned. YMTC has been on the Commerce Department’s Entity List since 2022, while the Pentagon’s 2026 list of Chinese military companies includes both YMTC and CXMT. These companies could conceivably use the shortage to gain customers through low prices and lock them into their ecosystems.

The potential impact of allowing U.S. companies to source Chinese-made memory chips, even if only to serve China’s local market, needs to be evaluated in the context of how the resulting demand could strengthen Chinese memory-chip manufacturers and, in turn, bolster China’s techno-economic power. A 2021 ITIF report estimated that there would be 5,100 more U.S. patents in the semiconductor industry annually if not for China’s innovation mercantilist policies in the sector. That same report also noted that Chinese state-backed companies such as YMTC have a precedent for distorting markets.

From a market-based perspective, one could argue that the main Chinese semiconductor firms meet all the conditions to be considered illegitimate companies.

The policy issue, thus, is finding ways to allow U.S. companies to use Chinese memory chips for products sold in the Chinese market (and nowhere else in the world), without creating long-term dependencies. One way to square the circle would be to add conditions for U.S. companies to get a temporary license to deal with YMTC and CXMT. For example, the party seeking an exemption should feasibly prove that: 1) the memory chips acquired would anyway be produced, and 2) the transaction is an extraordinary, one-off measure that won’t lead to longer-term contracts that incentivize Chinese memory makers to invest in serving U.S. firms.

Even if the U.S. federal government allows the use of Chinese memory chips for products to be sold in the Chinese market in the short term, it should also seek capabilities and incentives to encourage U.S. companies to favor U.S. or U.S.-based chipmakers.

A feasible medium-term measure is to help U.S. manufacturers qualify compatible replacement memory more quickly. Medical-device manufacturers must assess whether substitutions require additional Food and Drug Administration (FDA) review, while vehicle manufacturers must self-certify compliance with safety standards administered by the Department of Transportation’s National Highway Traffic Safety Administration (NHTSA). FDA’s draft guidance already addresses certain memory-chip supplier changes. Both agencies should clarify testing and documentation expectations for compatible substitutions, and FDA should prioritize necessary reviews when memory shortages threaten medical-device availability.

In the long term, Washington should focus on adding capacity, especially domestic capacity. Congress should make the 35 percent Advanced Manufacturing Investment Credit permanent, giving firms a durable incentive to build and expand U.S. fabs. The $275 million CHIPS award to expand and modernize Micron’s Manassas, Virginia, facility—which focuses on legacy chips rather than the latest generations—is a good example of federal efforts to build this capacity. Additional manufacturing support should target commercially viable projects that address persistent supply gaps, backed by credible customer commitments. Long-term agreements, such as Micron’s supply agreement with General Motors, can help demonstrate sustained demand and reduce investment uncertainty. It should also work with allies—especially Japan and South Korea—to expand trusted memory production capacity and support supply chains.

Finally, policymakers should strengthen semiconductor R&D and workforce development so U.S. producers can improve memory architectures, manufacturing productivity, and the technologies needed for the next cycle. For example, federal support could focus on improvements in stacking, bonding, testing, and repair, which could raise HBM packaging yields (the share of production that becomes usable output). Losing fewer memory dies during assembly would let manufacturers deliver more usable memory from a given wafer volume.

Certainly, there’s much policy can do to support America’s semiconductor manufacturing economy, but the ultimate objective should be to produce more memory in the United States, not to create mechanisms for deciding who gets supply.