DRAM Shortage to Push PC Memory Costs to 35% Through 2030

DRAM contract prices surged 90-95% in Q1 2026 and are expected to increase by 58-63% in Q2, the largest quarterly rise on record, according to TrendForce. This has raised memory costs to about 35% of PC build costs, leading Nvidia to raise the DGX Spark desktop price by $700 to $4,699. AMD is preparing for a Q3 launch of a 192GB Gorgon Halo APU.

The spike is primarily due to wafer allocation, with Samsung, SK Hynix, and Micron shifting capacity towards HBM3E and HBM4 for datacenter accelerators, where margins are higher. This has tightened conventional DRAM supply. SK Group chairman Chey Tae-won stated at Computex 2026 that the shortage is expected to last through 2030, with new fabs from all three manufacturers not reaching volume production until late 2027. Some vendors have begun adding flat memory surcharges to every purchase, and smaller buyers are reporting quotes that change by the hour.

Nvidia cited memory supply as the reason for the DGX Spark price increase to $4,699 in February. At Computex, Nvidia introduced the RTX Spark with 128GB unified memory, Tom's Hardware reports, while AMD's Ryzen AI Max 400 "Gorgon Halo" offers 192GB of unified memory, with up to 160GB addressable as VRAM—the first x86 client processor AMD claims can run a 300B-parameter model locally. Partner systems from Asus, HP, and Lenovo are scheduled for Q3, and an AMD developer box based on the previous 128GB Strix Halo is opening pre-orders in June at $3,999 through Micro Center.

However, bandwidth, not capacity, is the critical factor. Gorgon Halo maintains the same 256-bit LPDDR5X-8000 interface as Strix Halo, offering a theoretical peak of approximately 256 GB/s and an independently measured 212 GB/s on the GPU. Dense LLM inference is memory-bandwidth-bound: generation speed equals memory bandwidth divided by per-token weight footprint. At 212 GB/s, a dense 70B model operates in the low single digits of tokens per second, regardless of the 192GB capacity. In comparison, an Apple M3 Ultra delivers 819 GB/s, and an RTX 5090 achieves 1,792 GB/s. The 192GB pool aids KV-cache-heavy agentic workloads and MoE models that require additional space, but it does not enhance raw token generation for dense weights.

Architects designing edge inference must now consider memory as a volatile commodity cost with a multi-year horizon. HP informed investors in February that memory's share of PC build costs more than doubled from 15-18% to approximately 35% in a single quarter. The "agentic AI" narrative allows vendors to justify four-figure workstation prices against commodity mini PCs, but the hardware is constrained by the same DRAM shortage. Corsair's AI Workstation 300 review noted that the RTX Spark's GB10 chip outperformed Strix Halo as context length increased, highlighting that memory bandwidth still governs the workloads these systems are marketed to run.

Two caveats require attention. First, the 192GB configuration is based on a leaked PassMark entry showing eight 24GB SK Hynix LPDDR5X packages on an HP test board; AMD has not publicly confirmed this layout. Second, leaked roadmaps indicate a next-gen "Medusa Halo" moving to LPDDR6 with up to 80% more bandwidth, suggesting that AMD already views the current memory interface as insufficient for the model sizes the capacity invites. Until that silicon is available, architects are purchasing DRAM headroom without a proportional bandwidth upgrade.

Consider 192GB of edge DRAM as a cache-tier enabler for long-context agents and MoE offloading, not a throughput solution for dense models, and secure supplier contracts before the next quarterly quote revision.