Nvidia has informed buyers of its Grace-Blackwell and Vera-Rubin NVL72 rack-scale systems of a 15 percent price increase next year. The warning surfaced via Bloomberg at the Hot Chips 2026 conference, where the memory market set the tone for the week.

Hyperscalers and cloud builders have absorbed nearly every available GPU component for 2026, with allocations largely locked through the end of 2027. Smaller system vendors now have little leverage in a market where the largest buyers set the terms.

Jim Handy, general director at Objective Analysis, opened the memory tutorial on Sunday with a blunt diagnosis. "It's all made on the same manufacturing line," he told attendees, referring to flash, DRAM, and HBM. "Because of that, it's causing DDR prices to go crazy, and all of NAND flash prices are going crazy."

What is driving the demand shock?

AWS, Google, Microsoft Azure, and Meta Platforms dominate AI infrastructure investment from 2021 through the second quarter of 2026, according to Handy's aggregation. Alibaba, the largest Chinese hyperscaler, barely registers against that group, and Apple's 2026 AI capex sits below the prior five-year average.

The agentic AI layer compounds the strain. Agent sandboxes require large CPU fleets alongside substantial DRAM and flash, and even hard disk drives are moving as hyperscalers absorb near-cold archival capacity. Enterprise, government, and academic buyers are also rebuilding back-office and analytics infrastructure for GenAI, which tightens the market further.

Who controls HBM allocation?

Nvidia has long been the largest HBM buyer thanks to its dominance in AI training silicon. Broadcom has climbed to third place, behind Nvidia and AMD, propelled by its custom XPU business for hyperscalers and AI hardware startups that lack direct foundry access.

Broadcom and Marvell can compete for HBM capacity on price on behalf of those smaller customers who would otherwise miss the queue. Nvidia has reportedly inserted flexible pricing provisions into contracts so it can pass rising memory costs to ODMs, OEMs, and ultimately to system buyers. A 15 percent list-price increase looks almost modest given the sevenfold rise Handy cited on the spot market.

The antitrust question is never far away. Eighty-five percent revenue or shipment share has historically drawn attention from the US Department of Justice, though political climate now matters more than any fixed threshold.

How does the industry escape the memory wall?

Most forecasts Handy has reviewed see AI demand growing through 2030 at minimum, with shortages persisting into that horizon. The memory and flash vendors, after a brutal late-2022 and early-2023 trough tied to the early GenAI investment dip, are once again seeing revenues climb.

The architectural response is processing in memory, or PIM. "We are probably headed towards having smarter memory chips that will end up reducing the amount of traffic it needs to go to the processor," Handy said in his keynote. The processor instructs the memory to sort, the memory sorts, and the processor retrieves a single byte. That exchange would have required a flood of traffic without on-chip intelligence.

The roadmap extends well beyond HBM base dies. External flash tiers for KV caches and DPU-driven storage fabrics are in the mix, and Handy expects AI inference to reach microcontrollers, communications channels, and layers users never see. If the boom holds to 2030 and PIM-era memory ships at volume, the next contest will be over who collects the rents in a market that has finally remembered how scarce silicon memory can be.