Analysis
DRAM contract prices are forecast to grow 13% to 18% in the third quarter, The Register reported Monday, a smaller jump than the price spikes memory buyers absorbed earlier in the AI buildout, but still a meaningful, persistent cost layer on nearly every device shipped this year.
Why memory keeps getting squeezed
High-bandwidth memory, the specialized DRAM that sits directly alongside AI accelerators to feed them data fast enough to keep GPUs busy, has driven several of 2026's AI hardware bottlenecks on its own. Every unit of HBM capacity a memory maker allocates to Nvidia, AMD or a hyperscaler's custom silicon is a unit not available for standard server, PC or phone DRAM -- the same manufacturing lines and, in some cases, the same underlying wafer capacity serve both markets. That is the direct mechanical link between Nvidia's GPU shipment volume and the price of memory in an ordinary laptop.
“That is the direct mechanical link between Nvidia's GPU shipment volume and the price of memory in an ordinary laptop.”
The forecast easing to 13-18% growth, down from sharper spikes earlier in the cycle, suggests memory makers -- Samsung, SK Hynix, Micron -- have added enough capacity to slow the rate of increase, even as absolute demand keeps climbing. That is consistent with Dell's own record $60.9 billion quarter of AI server orders, reported this same week, which requires DRAM and HBM at a scale that keeps upward pressure on price regardless of the growth rate slowing.
Who actually pays for it
The cost does not stay contained to AI infrastructure buyers. Every phone maker, PC OEM and mid-range server vendor competing for standard DRAM allocation against AI-driven HBM demand passes some of that cost through to consumer and enterprise buyers who have nothing to do with AI. The Verge has separately reported that rising RAM costs are a meaningful driver of higher phone prices this year -- a downstream consequence of the same capacity competition driving this DRAM forecast.
The counterweight
A 13-18% quarterly growth forecast is meaningfully below the panic-driven spikes DRAM has seen in past supply crunches, and TrendForce's own framing treats this as stabilization rather than crisis. Memory manufacturers have strong incentives to keep investing in new capacity given how profitable HBM allocation has become, which should ease the constraint further into 2027 assuming no new demand shock -- a large jump in edge-AI device memory requirements, for instance, would reset the trajectory upward again.
For hardware and infrastructure investors, DRAM and HBM capacity is one of the cleanest leading indicators available for how tight the broader AI supply chain actually is -- tighter than headline GPU allocation numbers suggest, because memory bottlenecks show up as a tax on unrelated consumer products months before they show up as a data-center capacity story.
Why the number keeps rotating
Every AI-buildout cycle so far has hit a different physical bottleneck in sequence: first GPU allocation itself, then grid interconnection and power -- the constraint behind Thailand's recent pause on new data center approvals -- and now memory. Each time a bottleneck eases, the next one becomes visible, because the buildout's total scale keeps growing faster than any single supplier can add capacity. Memory makers investing heavily in new fabs today are underwriting a demand curve that assumes AI server growth keeps compounding through the rest of the decade; if that growth rate slows even modestly, this quarter's capacity additions could just as easily produce a memory glut in 2027 as a continued shortage.
