Bloomberg’s Oct. 27, 2026, headline framed the AI memory-demand debate as a $2 trillion sustainability test. It described memory-chip giants as having that combined valuation and asked whether AI demand could sustain it. AI workloads can hit a real memory-bandwidth bottleneck, but that technical need alone does not answer whether the investments behind it will earn adequate returns.
Why AI workloads can be memory-bound
Memory bandwidth is the rate at which data moves between memory and a processor. If data cannot move fast enough, an AI workload can be memory-bound even when a powerful GPU has computing capacity available. Adding more compute alone may not make that workload run faster.
IBM Technology presenter Kaoutar El Maghraoui explains how the gap between processor compute and data movement creates this constraint. Arithmetic intensity, the number of floating-point operations relative to bytes accessed, helps identify whether a workload is limited by computing power or memory bandwidth. The roofline model visualizes that distinction.
That mechanism explains why memory performance matters to AI workloads. Whether the investments in that memory generate adequate returns is a separate business question.
Why this memory cycle could differ
Memory is a cyclical business: capacity expansions can overshoot demand when a cycle turns. Lazard Asset Management portfolio manager and analyst Celine Woo argues that structural changes may make the current cycle harder to compare with earlier ones. She points to greater manufacturing and capital intensity, as well as difficult technology transitions, as factors shaping both supply and demand.
Woo’s view is one investor’s analysis, not a settled forecast. Her argument puts structural pressures alongside the industry’s history of overbuilding; it does not remove that risk.
What executives and analysts forecast
The Motley Fool reported on Oct. 27, 2026, that Nvidia CEO Jensen Huang said a memory glut by 2029 was not his base case. The report placed his remarks on Oct. 22 at a Berlin conference. His reported outlook is not a guarantee that a glut will not occur.
The same article reported that Micron Technology and SK hynix had signed three-to-five-year deals with major customers setting price ranges and sales volumes. It also reported that SK hynix plans to triple wafer capacity by 2034. Contracts and a capacity plan describe different parts of the market: neither, on its own, determines whether future supply will match demand.
Counterpoint Research forecast that DRAM prices would fall 23% in 2028 compared with 2027, while remaining 230% above 2025 levels, according to The Motley Fool. That forecast describes a price path, not a direct measure of whether the market will have a glut.
On July 23, 2026, iXBT.games reported ADATA chairman Simon Chen’s forecast that DRAM shortages could last at least another decade. Chen also forecast third-quarter contract-price increases of 20–30% for DRAM and 35–40% for NAND. These are distinct supplier-side projections, with different time horizons and measures.
A separate reported cancellation involved server hardware. Yahoo Finance reported on April 24, 2026, that Oracle canceled an order for 300–400 Nvidia GB300 NVL72 racks from Super Micro Computer. Those were server racks, not memory-chip orders.
Technical demand and investment returns
A memory-bandwidth bottleneck gives AI workloads a concrete technical reason to need faster data movement. It does not establish how much memory companies will buy or whether AI spending will earn enough to justify the valuations Bloomberg’s headline framed. The market’s sustainability question remains tied to both sides of the cycle: demand for AI infrastructure and the capacity memory makers bring online.
