Intel CEO Lip-Bu Tan Warns AI Memory Shortage Could Worsen as Prices Surge

AI infrastructure may be running into a constraint that is less visible than GPUs but just as consequential: memory. Speaking at the AI Infra Summit in Santa Clara on September 15, Intel CEO Lip-Bu Tan said the shortage had already delayed business projects and lifted memory prices by roughly five to seven times. He warned that the situation would worsen as AI systems consume more supply. 45

Memory, not just compute, is limiting AI buildouts

Tan’s central point was that AI deployment depends on more than processors. AI servers require substantial memory capacity, including high-bandwidth memory (HBM), and the rapid expansion of AI infrastructure is tightening availability across the market. Reporting from the summit said that the resulting supply pressure was spreading beyond AI servers to PCs and smartphones. 4

This matters because a system builder can have compute hardware planned or ordered yet still be unable to complete a deployment if memory is unavailable. In that environment, memory becomes a practical scheduling constraint—not simply another line item in a server bill of materials.

Why the shortage is proving difficult to fix quickly

The available evidence points to a capacity mismatch: demand for AI infrastructure is rising faster than suppliers can make memory available. Tan said that memory had become a bottleneck and that conditions would get worse. 5

A separate report from February quoted Tan as saying that, based on discussions with key industry players, he had heard of “no relief until 2028.” That should be read as an executive assessment of the supply outlook rather than a confirmed industry-wide delivery schedule. 8

The pressure is especially significant for HBM, which AI servers use in large quantities. The September summit reporting linked the shortage to AI servers absorbing HBM supply, contributing to shortages and price increases that reach conventional devices. 4

How long could tight supply last?

There is no single, verified timetable for normalization in the materials available here. But the warnings consistently describe a multi-year risk rather than a short-lived disruption:

• Tan’s view: no relief until 2028, based on what he said key suppliers had told him earlier in 2026. 8
• Phison’s view: K.S. Pua said NAND flash could face its worst shortage in 2027. 17
• Broader market context: an analysis cited sustained tight supply, rising prices and multi-year procurement commitments, while noting that Samsung Electronics, SK hynix and Micron account for the overwhelming majority of global DRAM production capacity. 57

These are forecasts and market assessments, not guarantees. Supply additions, changes in AI investment, and improvements in memory efficiency could alter the outcome.

The cost and availability consequences

The immediate effect is likely to be tougher procurement and higher system costs. When scarce memory is allocated toward AI infrastructure, lower-priority or price-sensitive device categories can face reduced availability or greater component costs. The summit reporting specifically identified spillover into smartphones and PCs. 4

For AI operators, the practical implications are straightforward:

1. Plan memory procurement early. Compute availability alone does not ensure a deployable system.
2. Treat memory as a design constraint. Capacity, bandwidth and allocation risk should be considered at the architecture stage.
3. Expect interdependent bottlenecks. Adding more accelerators can increase memory, power and cooling requirements at the same time.

Power and cooling are the other limits Tan highlighted

Tan did not frame the challenge as a memory-only problem. He also pointed to securing electricity and advancing cooling technology as major issues for the chip and AI-infrastructure market. 5

That broader framing is important: scaling AI requires a coordinated system of compute, memory, networking, power delivery and thermal management. More available memory would remove one obstacle, but it would not by itself solve the physical constraints of operating dense AI systems.

Bottom line

Tan’s warning is that AI’s hardware bottleneck is moving beyond processors. Memory supply is delaying projects, prices have climbed sharply, and the near-term outlook remains tight. His earlier expectation of no relief before 2028 and Phison’s warning of a potentially severe NAND shortage in 2027 underscore how long the pressure could last—but both remain forecasts, not settled outcomes. 4817

For companies planning AI capacity, memory availability, power and cooling now belong in the same category as accelerator procurement: essential infrastructure risks that need to be designed around from the start.