AI Server Prices Jump 15% as Memory Shortage Bites

Contract server builders have begun notifying Nvidia’s largest customers that AI servers shipping in early 2027 will cost more than 15% more, in many cases, than the equivalent systems today. Bloomberg reported the notices on Saturday; Tom’s Hardware puts some configurations closer to 17%. The increases cover Grace Blackwell and Vera Rubin platforms and vary by generation and memory configuration, and they reached buyers through the assembly middlemen rather than from Nvidia directly.

Nothing about the accelerators changed. Memory did. On Implicator’s numbers, memory is about 29% of a Vera Rubin VR200 system’s roughly $2.1 million bill of materials — a single NVL72 rack carries 20.7TB of HBM4 alongside tens of terabytes of LPDDR5X. When a line item that large moves, the sticker moves with it, and DRAM, LPDDR and high-bandwidth memory have all been moving in the same direction for the better part of a year.

The scale of that move is the part worth sitting with. 24/7 Wall St. notes server DRAM roughly doubled in the first quarter of 2026, with Counterpoint Research tracking 80–90% quarter-over-quarter increases across DRAM, NAND and HBM, and Deloitte projecting that AI-server DRAM prices could quadruple over the course of 2026. Against that, a 15% bump on a finished server is a remarkably well-absorbed shock — evidence of how much of the rack is still GPU, and of how hard Nvidia has worked to keep the number in the teens.

Some of that absorption is engineering. Nvidia is reportedly weighing a cut in SOCAMM capacity from 55TB to 28TB per rack while leaving HBM4 at 20.7TB, trading bulk system memory to protect the bandwidth tier that actually feeds the GPUs. That is a design decision made by the memory market rather than by a roadmap, which is a new posture for a company that has spent three years setting the terms.

Nobody expects relief soon. Samsung and SK Hynix warned back in April that the shortage could run through 2027. TrendForce expects HBM bit shipments to grow 50–60% year over year in 2027 and still fall short of demand. Gartner sees the squeeze lasting at least through mid-2027; Deloitte does not expect meaningful new capacity before 2029 or 2030. The suppliers — Micron, which sold out its entire 2026 output, plus SK Hynix and Samsung — hold the pricing power for the duration.

Who actually pays depends on how big you are. Microsoft, Alphabet and Oracle negotiate long-term memory commitments and in some cases invest directly in suppliers; enterprises buying through Dell, Lenovo, HPE or Supermicro have neither lever and take the pass-through. The spillover is already visible outside the data centre, with Apple raising some product prices up to 20% and Amazon taking the Echo Dot from $49.99 to $79.99 and the base Kindle from $109.99 to $149.99, both citing components. And there is a second-order effect worth watching: when a standard rack costs 15% more to buy, the arithmetic behind custom accelerator programmes and in-house chip efforts gets easier to defend — though custom silicon needs the same memory, from the same three vendors, at the same prices.