AI's Hardware Squeeze Spreads to Memory and CPUs, With Relief Years Away

For most of the AI boom, the scarce part was the GPU. This week's earnings from the companies that make memory and server processors showed the squeeze has spread. SK hynix said it expects memory demand to exceed supply through 2030, Samsung said its 2027 supply of high-bandwidth memory is already sold out, and Intel said its top priority is finding enough capacity to meet demand for its Xeon server chips. For enterprise technology leaders, the message is that AI infrastructure costs will be shaped by component scarcity for years, not quarters, and that planning has to change accordingly.

How tight is the memory market?

SK hynix, the largest supplier of high-bandwidth memory (HBM) to AI accelerator makers, reported third-quarter revenue of 104.6 trillion won, up 328% from a year earlier, and operating profit of 83.3 trillion won, an 80% operating margin, in results released this week. It was the first quarter in which the company's revenue topped 100 trillion won.

More telling for buyers was the outlook. SK hynix said it expects supply to lag demand through 2030 because leading-edge processes are hard to scale and new fabs take years to build. It has signed long-term agreements with 14 major customers, including global cloud providers, that cover more than half of its total production capacity for 2027 and 2028. Those contracts typically run five years, with prices tied to market conditions within an agreed range. The Korea Times reported that HBM and other DRAM products are sold out through 2029.

The company also said AI demand is broadening, from a handful of very large buyers to enterprises and national AI programs.

Samsung told a similar story at its full third-quarter results on October 29. Its HBM4 revenue tripled from the previous quarter and made up 60% of its HBM sales, Money Today reported, and negotiations for 2027 HBM volumes are finished at target prices, with that supply already sold out. Samsung said major customers are still asking for more. Earlier in October, Samsung's preliminary figures had put quarterly operating profit at a record 107.4 trillion won, and TrendForce noted that both Samsung and Micron expected memory supply to stay tight through 2028.

Why are CPUs now in short supply too?

Intel's results on October 29 showed the pressure moving to general-purpose server processors. Intel reported third-quarter revenue of $16.5 billion, up 21%, with its Data Center and AI segment up 78% to $7.3 billion, which the company called its strongest data center growth in 15 years. CEO Lip-Bu Tan said Intel's immediate priority is to "unlock additional supply to meet extraordinary demand" for Xeon.

Intel tied part of its pitch to agentic AI. Agents do far more than call a model: they query databases, run tools, move files and orchestrate steps, much of which runs on CPUs. Intel highlighted architectures for agentic and enterprise workloads and a multiyear deal to expand Xeon capacity on Oracle Cloud Infrastructure.

There is a cost on the other side. Intel guided fourth-quarter revenue to $15 billion to $16 billion, and CFO David Zinsner said the company expects the PC slowdown to continue at least through the first half of next year. Intel did not tie that slowdown to a single cause, but higher component costs make every device and server refresh more expensive.

What does the squeeze mean for enterprise budgets?

The direct effect is on anything that contains memory, from AI servers to ordinary database hosts and laptops. When the largest suppliers have committed more than half of their 2027 and 2028 output to long-term contracts with cloud providers, enterprises buying smaller volumes through resellers are last in line for allocation and have the least pricing leverage.

The indirect effect lands in cloud bills. Cloud providers are paying record memory prices as they build AI capacity, and those costs are likely to surface over time in instance pricing and reserved-capacity terms. Leaders from Bank of America, Carelon Health, Lockton and PRGX warned about a version of this dynamic on a COLLIDE 2024 panel. "In data center everything was free, and in cloud it is not," PRGX Chief AI Officer Manish Dingare said, arguing that unmonitored cloud and GPU spending erodes AI returns.

Efficiency becomes a strategy, not a nice-to-have. At COLLIDE 2025, Aerospike founder and CTO Srini Srinivasan argued that infrastructure cost decides whether real-time AI pays off. "Virtually everybody can do things at high scale," he said. "But what does it cost?" His point lands harder when the memory that holds working data is the most constrained part of the system.

How should CIOs plan for a multiyear hardware squeeze?

Leaders who have navigated these cycles suggest several moves.

Plan refreshes earlier and in larger blocks. If suppliers are signing five-year contracts, a buyer that orders quarter by quarter will pay spot prices. Work with hardware vendors and resellers on allocation commitments for 2027 now, and treat memory pricing as a line item to forecast rather than a fixed cost.

Split workloads by what they actually need. Not every step of an AI pipeline needs a GPU or a memory-heavy server. In a AI Connect webinar on AI infrastructure and inference costs, panelists described running preprocessing and postprocessing on cheaper CPU containers and reserving accelerators for inference only.

Compare cloud and on-premises again. A shortage changes the economics in both directions. Owned hardware locks in today's prices but may be hard to source; cloud capacity avoids procurement delays but passes through supplier costs over time. The answer will differ by workload, and last year's analysis is probably out of date.

Right-size models and data. Smaller models, shorter context windows and better data curation reduce both memory and compute per task. Every token not processed is capacity that does not need to be bought.

The long view

The memory makers' outlook is not a guarantee. Forecasts of shortages through 2030 come from companies that benefit from high prices, and new fabs, including SK hynix's Yongin site due for full operation next year, will add supply. Demand could also cool if AI spending slows. But the contracts already signed suggest that the largest buyers do not expect relief soon.

For enterprise leaders, the practical conclusion is that AI capacity is a supply chain problem as much as a technology one. The organizations that plan purchases early, measure cost per outcome, and engineer workloads to use less memory and compute will be the ones that keep their AI programs on budget while components stay scarce.

Frequently asked questions

How long will the AI memory shortage last?

SK hynix said in its October 2026 results that it expects demand to exceed supply through 2030, while Samsung and Micron earlier said supply would stay tight through 2028.

Why are server CPUs in short supply?

Intel reported 78% growth in data center and AI revenue in Q3 2026 and said its priority is adding supply to meet demand for Xeon processors, which run much of the non-accelerator work in AI and agent systems.

How should enterprises respond to rising memory costs?

Forecast memory as a variable cost, negotiate allocation for 2027 early, run non-inference steps on cheaper CPU infrastructure, revisit cloud versus on-premises economics and right-size models.

Filed under: Data & AI Infrastructure, AI Strategy & ROI