The memory bill for AI data centers
Bloomberg reported on Saturday, August 23, 2026, that Nvidia has informed some of its largest customers that prices for servers containing its artificial intelligence chips will rise by more than 15% in many cases — and memory is the primary culprit behind this increase. The rise will apply to systems with both the Grace Blackwell and the next-generation Vera Rubin platforms that ship from early 2027 onward, according to people familiar with the communications who spoke on condition of anonymity about information not yet made public. The magnitude of each increase will depend on the chip generation and memory configuration involved, meaning this is not a uniform price table but rather a range of adjustments that could be significantly larger in certain configurations.
RAMageddon and the real cost of AI
The phenomenon already has a name circulating among industry analysts: RAMageddon. Contract prices for conventional DRAM have risen at record rates throughout 2026. Analysts projected that conventional DRAM contract prices would climb between 58% and 63% quarter-over-quarter in Q2 2026, following a Q1 surge of 90% to 95%. SK hynix announced in October last year that it had already sold out its entire 2026 memory production capacity, while Samsung and SK hynix raised HBM3E supply prices by nearly 20% before the year even began.
The irony is that the same supply crunch now inflating Nvidia's systems was partially created by the very demand it helped fuel. The three major memory manufacturers spent 2025 and 2026 redirecting advanced nodes and new capacity toward HBM and server DRAM, where margins are substantially more attractive. This hollowed out the conventional product market. Consumer DDR5 pricing has more than doubled since late 2025: a mainstream 32 GB DDR5-6000 kit that cost between 10 and 40 a year ago was listed around 92 in August, according to Tom's Hardware's RAM Price Tracker.
The weight of memory in the bill of materials
A Rubin GPU system can carry up to 288 GB of HBM4 per package. The rack-scale NVL72 system combines 72 of these GPUs, placing more than 20 TB of HBM in a single rack before even accounting for the LPDDR memory attached to Vera CPUs. Since HBM production consumes roughly four times the wafer area of equivalent conventional DRAM, memory has become one of the largest line items in an AI server's bill of materials — and it continues rising at a stratospheric pace.
An analysis from GF Securities, cited by Wccftech in a July 28 TrendForce report, put memory at about 29% of a roughly .1 million bill of materials for a Vera Rubin VR200 system if Nvidia made no configuration cuts. Nvidia's own preferred memory share of the system bill of materials was 20%. The same report suggested the company may reduce the SOCAMM memory attached to Vera CPUs from about 55 TB to 28 TB per rack while keeping GPU HBM4 capacity at 20.7 TB — a maneuver that would lower the amount of costly LPDDR5X memory per system without touching the high-bandwidth memory that feeds the GPUs.
Implications for the market
Nvidia already passed rising costs through to consumers on the gaming GPU segment by raising prices on GeForce graphics cards earlier in August. A 15% increase on rack-scale systems selling for several million dollars each means hundreds of thousands of additional dollars per rack across deployments that run to thousands of racks. The company runs a gross margin of roughly 75% non-GAAP, among the highest in the semiconductor industry, and the reported hikes indicate Nvidia intends to pass memory cost inflation on to customers rather than absorb it — something it can more than afford given the context.
Meanwhile, supply of Nvidia's accelerators from TSMC still cannot meet demand, which limits buyers' immediate negotiation leverage. The question that remains is whether these increases will push the largest customers toward AMD accelerators or their own custom chips — and whether those alternatives will be able to absorb displaced demand before the marginal cost of AI becomes unsustainable for smaller companies that lack long-term agreements with the three dominant memory suppliers.
Sources: Fortune, Tom's Hardware, Rogers' Lab
✓ Independent sources cross-checked and verified before publishing
