Memory Prices Surge as AI Demand Collides With Limited Supply

Hot Chips 2026 Opens With a Memory-Market Overview

Hot Chips 2026 began with half-day tutorials on Sunday. Jim Handy, general director at semiconductor market research firm Objective Analysis, opened the memory tutorial with an overview of the DRAM, HBM, and flash markets. His presentation provided context for later sessions on high-bandwidth memory and high-bandwidth flash.

Separate presentations from Micron Technology, Samsung, SK Hynix, d-Matrix, Meta Platforms, and Oxmiq Labs addressed the technical details. Handy's market overview highlighted the sharp rise in DRAM and flash prices as strong demand meets constrained supply.

Hyperscalers Drive Memory Demand

Hyperscalers and cloud providers are investing heavily in infrastructure to train increasingly capable GenAI models and run inference in production. That spending is driving demand for HBM and GPU/XPU systems, along with the DRAM and flash required by those platforms.

The growth of agentic AI is adding another source of demand. Agents require substantial CPU capacity, as well as DRAM and flash for the sandboxes in which generated Python code and other workloads execute. These resources are also used as applications are modernized or developed from the ground up.

Handy's spending data, covering the period from the beginning of 2021 through the second quarter of 2026, shows the scale of investment by major technology companies.

The chart also illustrates the difference between Alibaba, the largest of China's hyperscalers with a cloud business, and the four leading US technology investors: Amazon Web Services, Google, Microsoft Azure, and Meta Platforms. Apple, which uses AI but does not build AI systems at the same scale, represents only a small amount of investment. Its AI-related capital spending in 2026 is also lower than in each of the previous five years.

Storage Costs Rise Across the Market

Demand from hyperscalers and cloud providers is being reinforced by new investments from enterprises, governments, and academic institutions. These organizations are seeking more efficient back-office, web infrastructure, and data-analytics systems while preparing capacity for GenAI workloads.

The resulting pressure extends across the storage market, including hard disk drives, which technology companies primarily use for nearly cold storage. The five largest memory and flash manufacturers have benefited from the resulting increase in revenue.

The market experienced a pronounced memory and flash downturn in late 2022 and early 2023, when GenAI was only beginning to gain momentum. If the GenAI expansion continues, revenue could keep increasing. Industry forecasts cited at the conference anticipate growth through at least 2030, with supply shortages potentially lasting through that period as well.

Commodity Products and Volatile Pricing

Flash, DRAM, and HBM are produced using related manufacturing capacity. According to Handy, this connection means that demand for one type of memory can affect the availability and pricing of the others, contributing to sharp increases in DDR and NAND flash prices.

These products are commodities, meaning buyers can often switch between suppliers with relatively little difficulty. Price and availability typically matter more than the identity of the manufacturer. During a shortage, however, suppliers can favor customers offering higher prices, which can push spot-market prices sharply upward.

Handy said spot-market prices had increased by approximately seven times. He also noted that spot pricing is not a perfect market indicator, although it is relatively easy to measure.

Spot prices represent a type of average. Companies producing systems or processors with integrated memory, including HBM and LPDDR5X DRAM, negotiate their own prices, volumes, and commitments. Depending on those agreements, their costs may be better or worse than the spot-market average.

HBM Capacity Shapes AI-System Growth

Nvidia has been the largest-volume HBM buyer for much of the deep-learning and early AI boom, giving it substantial influence over HBM pricing. Broadcom has now become the third-largest HBM buyer, behind Nvidia and AMD, through its custom XPU business. Its financial resources allow it to compete for HBM capacity on behalf of AI hardware startups that might otherwise struggle to obtain chip manufacturing and packaging support from companies such as Broadcom and Marvell.

GPU components for 2026 have reportedly been sold out for some time, while component allocations may already extend through the end of 2027. Reports in Bloomberg said Nvidia had warned customers purchasing Grace-Blackwell and Vera-Rubin NVL72 systems that prices would rise by 15 percent the following year.

The increase may allow Nvidia to pass higher HBM, DRAM, and flash costs to its original design manufacturer and original equipment manufacturer customers, which would then pass those costs to system buyers.

Nvidia's growth is limited by the amount of HBM it can obtain, as is the case for other GPU and XPU manufacturers. At the same time, memory suppliers and Nvidia must consider competition concerns and ensure that competing companies can obtain capacity. Historically, a revenue or shipment share of 85 percent or more has often been treated by the US Department of Justice as a potential monopoly threshold, although enforcement decisions depend on the circumstances and prevailing policy.

Moving Computation Closer to Memory

The industry's response includes efforts to reduce the amount of data that must move between processors and memory. Examples include external flash-based, DPU-driven storage tiers for key-value caches, compute capabilities in HBM base dies, and processing directly within memory subsystems. The last approach is known as Processing In Memory, or PIM.

Handy said communication requirements between computing devices can be reduced when the memory system performs more of the work. He described a model in which a processor instructs memory to sort data, receives a completion signal, and then requests only the top result. Without computation inside the memory chip, the same operation would require much more data movement.

This approach could help address the memory wall by reducing traffic to the processor and increasing the intelligence of memory devices.

Handy also expects AI to spread across the computing stack, including microcontrollers, endpoint devices, and communication channels. As developers apply techniques that work in one area to other parts of a system, AI could become increasingly integrated into infrastructure until its presence is largely invisible to users.