Memory Shortage Won't End Until 2028: AI Impact

Micron CEO Sanjay Mehrotra told investors on September 30 that demand for his company's memory products will exceed available supply for at least the next two years. Samsung executives echoed that assessment days later. Together, the two statements amount to the clearest signal yet that the memory shortage will persist into 2028 — and that AI infrastructure planners should stop treating tight supply as a temporary disruption.

Why Chip Executives Say the Memory Shortage Runs Through 2028

Mehrotra's remarks came during Micron's quarterly investor communication, where he said demand would outstrip supply "for at least the next couple of years," according to Ars Technica. Samsung, the world's largest memory manufacturer by revenue, offered corroborating guidance the same week. When the two largest players in a consolidated market independently project the same multi-year imbalance, the memory shortage 2028 timeline stops being a forecast and starts being a planning assumption.

The consolidation matters. DRAM and HBM production is concentrated among three firms: Samsung, SK Hynix, and Micron. TrendForce data shows that the top three suppliers control the overwhelming majority of global DRAM output, and HBM capacity is even more concentrated. In a market this narrow, a unanimous signal from two of the three largest producers carries unusual weight.

The structural logic reinforces the executive consensus. Memory fabs are among the most capital-intensive facilities on earth. Building new capacity takes 18 to 24 months from groundbreaking to volume production, assuming no permitting delays, equipment bottlenecks, or yield problems. A shortage that executives say lasts two more years cannot be solved by capacity that does not yet exist.

Micron's own business decisions underline the shift. The company no longer sells consumer RAM at all. Its memory output now flows to business-to-business customers: HBM for AI accelerators and DRAM for servers. That is not a temporary allocation choice. It is a strategic reorientation around the highest-margin, fastest-growing segment of the market.

HBM and Server DRAM: The AI Demand Engine

High-bandwidth memory is the bottleneck inside every AI accelerator. Analyst firms tracking the segment — TrendForce and IDC among them — have repeatedly revised HBM market forecasts upward as training clusters and inference fleets scale. HBM is not interchangeable with standard DRAM. It is stacked, packaged, and co-engineered with specific GPU architectures, which means each new accelerator generation locks in demand years before it ships.

Server DRAM faces parallel pressure. Every AI rack needs host memory alongside accelerator memory, and hyperscale operators are provisioning both at levels that would have seemed implausible before the generative AI boom. The result is a demand curve that bends upward faster than any fab construction schedule can follow. TrendForce has noted that HBM production consumes significantly more wafer capacity per bit than conventional DRAM, meaning the AI mix shift effectively shrinks total industry output even when wafer starts hold steady.

SK Hynix, which supplies HBM to Nvidia, has been operating in a sold-out condition for multiple quarters. Micron's guidance now confirms the same dynamic on its side of the market. This is the core of the memory shortage 2028 projection: not a demand spike that fades, but a structural reallocation of the entire industry's output toward AI.

Semiconductor economics explain why supply cannot respond quickly. A leading-edge memory fab costs tens of billions of dollars in capital expenditure, and the equipment lead times for advanced lithography and deposition tools stretch well beyond a year. Even when capital is committed, yields on new HBM lines ramp slowly. The industry's capex cycle simply does not run on the timescale that AI buyers want.

How Prioritizing AI Memory Squeezes Consumer Device Supply

Consumer RAM is the release valve, and Micron has closed it. By exiting consumer memory sales entirely, the company removed a supply source that PC builders, laptop OEMs, and DIY system integrators had relied on for years. Samsung and SK Hynix have not made identical exits, but their capacity allocation follows the same gravity: HBM and server DRAM earn more per wafer than consumer-grade modules.

The arithmetic is straightforward. When fabs dedicate more wafers to HBM — which consumes roughly two to three times the die area per gigabyte compared with standard DRAM, according to TrendForce analyses — fewer wafers remain for the commodity memory that goes into phones, laptops, and gaming hardware. Consumer device makers then bid against server buyers for a shrinking pool, and they lose.

The historical parallel is instructive. During the 2017–2018 memory upcycle, DRAM prices roughly doubled over seven quarters, and PC makers responded by shipping machines with less base memory. That episode lasted about two years. The current shortage, by executive guidance, will run longer and is driven by a demand source — AI infrastructure — that has deeper pockets than any consumer OEM. IT procurement teams should expect device configurations to stretch and prices to firm across the consumer catalog well into 2028.

Implications for AI Infrastructure Buildout

For organizations planning AI clusters, memory availability is now a first-order constraint alongside GPU allocation. A training cluster without sufficient HBM does not train. An inference fleet without server DRAM cannot serve requests. The practical consequence is that memory procurement windows are lengthening, and the era of buying accelerators and memory on the same timeline is ending.

Three implications stand out. First, capacity planning must extend beyond the typical annual budget cycle. If executives at Micron and Samsung are guiding to multi-year tightness, a twelve-month procurement plan is structurally inadequate. Second, memory should be contracted earlier and further out than GPU orders, because HBM supply is negotiated directly with the three manufacturers and those negotiations are getting more competitive, not less.

Third, architecture decisions gain new weight. Systems that economize on memory — through quantization, mixture-of-experts routing, or tiered inference — reduce exposure to the shortage. Those are engineering choices, not procurement ones, and they are made months before hardware arrives.

The hyperscalers have already adapted. Their scale lets them sign multi-year supply agreements and co-develop memory with manufacturers. Enterprises buying at smaller volumes do not have that option. For them, the shortage translates into longer lead times, higher prices, and a real risk that accelerator purchases sit idle waiting for memory that has not arrived.

What Buyers, Builders, and IT Teams Should Do Now

Extend planning horizons to 2028 and beyond. The consensus between Micron and Samsung makes a two-year horizon the minimum defensible assumption. Procurement calendars built around quarterly cycles will consistently miss.

Lock in memory contracts before accelerators. HBM and server DRAM are the scarcer inputs. Buyers who secure memory first and match compute to it will build more functional capacity than those who do it in the reverse order.

Model higher memory costs into total cost of ownership. The pricing environment of the past decade, shaped by periodic oversupply, is not the environment of the next three years. Business cases that assume declining memory costs per gigabyte need revision.

Diversify supplier relationships across Samsung, SK Hynix, and Micron where design allows. Qualification takes time, and single-sourcing in a shortage is a single point of failure.

Design for memory efficiency. Memory-optimized model architectures and inference stacks deliver operational relief that no purchasing strategy can match when supply is fixed.

Outlook: Will Supply Catch Up Before 2028?

Not on current evidence. The memory shortage 2028 guidance from Micron and Samsung rests on capacity that cannot be built fast enough and demand that shows no sign of plateauing. AI infrastructure spending continues to absorb every wafer the industry can allocate to HBM and server DRAM, while consumer memory supply shrinks as a consequence.

Three variables could shorten the timeline. A slowdown in AI capital expenditure would relieve pressure, but no major hyperscaler has signaled one. Faster-than-expected yield improvements on HBM lines could add effective capacity, though stacking and packaging remain difficult. A demand shift toward more memory-efficient model architectures could reduce bits required per unit of compute, but efficiency gains have historically been reinvested into larger workloads rather than banked as savings.

The most likely path through 2028 is a market that stays tight, with periodic regional and segment-level volatility, and with AI buyers consistently outbidding consumer and enterprise buyers for the same limited output. Organizations that treat this as a structural condition — and plan accordingly — will build AI infrastructure that actually runs. Those that wait for relief will keep waiting.

Source: Ars Technica - All content