HBM4 arrives and memory prices go through the roof

On 12 February, Samsung said it had begun shipping HBM4, the latest generation of high-bandwidth memory, the specialized DRAM stacks that sit next to AI accelerators and feed them data faster than any other memory can. The announcement was one line in a bigger story: the memory industry has spent the past year in a pricing super-cycle, and the arrival of HBM4 is the moment the bottleneck moved from the chip to the memory that feeds it.

The market numbers are the clearest summary. Samsung expects its high-bandwidth memory sales to more than triple in 2026. Micron’s chief financial officer said in mid-February that the company’s HBM supply for the year is sold out, a statement it had been making since the previous year without the supply catching up. TrendForce, the market research firm, forecast that contract prices for conventional DRAM would rise by 90 to 95 per cent in the first quarter alone, which it described as the largest quarterly jump ever recorded. The last time memory prices moved this hard, the cause was a supply shock; this time the cause is demand that will not stop growing.

The demand comes from one place: AI infrastructure. Every frontier model trained in the past year consumed enormous amounts of memory bandwidth, and the inference systems that serve those models to users consume even more. High-bandwidth memory is the component that determines how fast a GPU can actually run, which is why NVIDIA’s newest accelerator designs, including the Vera Rubin systems announced for the coming years, are built around HBM4 and its successors. The memory makers have responded by converting as much conventional DRAM capacity as they can to HBM production, which is itself part of the reason ordinary DRAM prices are jumping: the same fabs are producing fewer of the chips that go into laptops and servers.

The structure of the market deserves attention. Memory manufacturing is one of the most concentrated industries in technology, with effectively three suppliers, Samsung, SK hynix and Micron, controlling nearly all of the world’s DRAM. All three have been disciplined about capacity for years, which is why the current boom is translating into prices rather than overcapacity. The strategic question is whether that discipline survives contact with the current order book: every hyperscaler is signalling capital spending at record levels for 2026, and the largest buyers are signing long-term agreements to lock in supply, which is exactly the kind of behaviour that, in past cycles, has ended with overbuilding and a crash.

The arrival of HBM4 itself matters technically. Each generation of HBM roughly doubles bandwidth, and the jump from HBM3E to HBM4 changes not just speed but the way the memory is packaged and how close it can sit to the compute die. For the AI labs and cloud providers building the next generation of systems, HBM4 availability is a scheduling constraint: systems designed around it cannot ship without it. That is why Samsung’s shipping announcement, routine for most components, is treated as an event, and why the shortage of advanced memory has become a topic at earnings calls, government meetings and trade negotiations in equal measure.

The uncomfortable question for anyone watching the industry is where the cycle turns. Memory is famously cyclical, and every up-cycle in the past has been followed by a down-cycle, often brutal, as capacity catches up with demand. The current bet, from the hyperscalers’ capital budgets and the memory makers’ long-term contracts, is that AI demand keeps growing faster than supply can be added. That bet has held for two years. It has not, yet, survived the moment when the largest buyers finish their current build-out, and the industry’s own history says that moment arrives with little warning.

For now, though, the story is a simple one: the component that was once the boring part of a computer has become the most constrained, most valuable and most discussed piece of the AI supply chain, and the companies that make it are having the best year of their existence. HBM4 is shipping, the prices are going up, and everyone who needs memory is paying attention to it the way they used to pay attention only to processors.