The Memory Supercycle Is Here, and AI Just Priced Everyone Else Out of RAM, GENZ TECH

Memory has become the bottleneck of the AI boom, and the bill is landing on everyone. DRAM contract prices are set to rise 50 to 55 percent this quarter versus the end of 2025, high-bandwidth memory is sold out for all of 2026, and the three companies that control more than 95 percent of the world's DRAM have redirected their capacity toward AI accelerators. The result is a textbook supercycle, one Bank of America compares to the 1990s boom. SK Hynix says it will double production over five years, but that capacity is years away, and the squeeze is already making the phone and laptop in your pocket more expensive.

DRAM contract prices are projected to climb 50 to 55 percent this quarter; some contract DRAM has risen more than 200 percent since early 2025.
HBM, the memory that feeds AI GPUs, is sold out for 2026, with hyperscalers locking multi-year supply from Samsung, SK Hynix, and Micron.
The capacity reallocation is squeezing consumers: 2026 smartphone shipments are projected to fall 12.9 percent and the PC market 11.3 percent.
SK Hynix plans to double overall production over five years and is deepening ties with TSMC on HBM4 base dies.

What actually happened

The memory market has split into two worlds. On one side sits HBM, the stacked, ultra-fast memory bolted directly onto AI GPUs, where demand is effectively infinite and supply is fully spoken for through 2026. On the other sits commodity DRAM and NAND, the ordinary memory in phones, laptops, and servers, which is now being starved of capacity. Because the same fabs and the same wafers can make either, the three giants that dominate DRAM, Samsung, SK Hynix, and Micron, have rationally shifted production toward the higher-margin AI memory. HBM now consumes a large and growing slice of total DRAM wafer output. The predictable consequence: prices for everything else are spiking, with contract DRAM up 50 to 55 percent this quarter and compounded increases topping 200 percent since early 2025.

Why is AI eating all the memory?

Modern AI accelerators are memory-bound, not compute-bound. A GPU can only run a large model as fast as it can feed that model's weights and activations in and out of memory, which is why HBM, with its enormous bandwidth, has become the single most contested component in the AI stack. As one Kioxia executive put it, the next constraint on AI performance is not raw compute but the ability to move data fast enough to keep the compute busy. That makes memory a strategic chokepoint. Hyperscalers, understanding this, have locked in future supply through multi-year deals, pulling capacity off the open market before anyone else can buy it. TSMC's advanced CoWoS packaging, which assembles HBM onto GPU dies, is booked solid through 2027. The entire pipeline, from wafer to packaged accelerator, is reserved years out. There is simply nothing left for the commodity market.

The mechanism most coverage skips

The cruel logic here is that consumer electronics and AI data centers draw from the exact same well, and the well is fixed in the short term. Building a new memory fab takes years and tens of billions of dollars, so when AI demand surged, the only lever the manufacturers had was to reallocate existing capacity. Every wafer turned into HBM for a data center is a wafer not turned into the DRAM for a smartphone. This is why your next phone and laptop will likely cost more even though nothing about them got better: the memory inside them is competing, directly, with Nvidia's customers, and Nvidia's customers have deeper pockets and signed contracts. The supercycle is not just a price story. It is a reallocation of a scarce physical resource away from consumers and toward AI infrastructure, and the manufacturers have every financial incentive to keep it that way.

Who this affects

Consumers feel it first and most directly, through pricier devices and an estimated 12.9 percent drop in smartphone shipments and 11.3 percent contraction in PCs for 2026 as makers absorb or pass on higher memory costs. Device manufacturers are caught in the middle, forced to either eat margin or raise prices into a weakening market. The memory makers themselves are the clear winners: Samsung and SK Hynix are posting strong earnings, and analysts forecast DRAM revenue up 51 percent and NAND up 45 percent year over year, helping push the global semiconductor market past $975 billion. Geopolitically, the squeeze has deepened the alliance between Korea and Taiwan. SK Group chairman Chey Tae-won met TSMC chairman C.C. Wei on June 3 to map cooperation on HBM4 base dies and custom AI memory, with SK Hynix pledging to double capacity over five years.

What is next

The key variable is when new capacity arrives and whether AI demand holds long enough to absorb it. SK Hynix's five-year doubling and Samsung's HBM5 roadmap will eventually loosen the supply, but not soon, and some analysts already warn that HBM prices could enter a correction after 2026 if too much capacity comes online at once. For now, watch contract DRAM pricing and HBM allocation as the leading indicators. If hyperscalers keep signing multi-year deals, the consumer squeeze continues regardless of how many fabs break ground. The risk for the memory makers is the oldest one in the chip business: everyone expands at once, demand softens, and the supercycle ends in a glut.

Our take

This is the clearest example yet of how the AI boom imposes costs far outside the data center. There is no dramatic shortage headline, no single failed product, just a quiet reallocation of memory capacity that makes ordinary electronics more expensive for everyone. The manufacturers are behaving rationally, chasing the highest-margin demand, but the externality is real: a contraction in the consumer market driven not by weak demand for phones and laptops but by the fact that their most critical component is being bought up by AI. The supercycle will mint enormous profits for three companies and quietly tax billions of consumers. That is the trade the industry has made, and it will hold until the fabs catch up or the AI demand finally blinks.

Reporting via SK hynix Newsroom, analysis by GenZTech.

The Memory Supercycle Is Here, and AI Just Priced Everyone Else Out of RAM