Nvidia's 15% price hike shifts AI's cost burden onto cloud giants

Nvidia's biggest customers just got a warning that will ripple through every AI budget for 2027: server prices are going up by more than 15% in many configurations, and memory shortages, not chip scarcity, are to blame. Bloomberg reported on 22 August, citing people familiar with the matter, that Nvidia notified some of its largest customers of the increase ahead of shipments of its flagship Vera Rubin and Grace Blackwell platforms early next year.

The size of the increase depends on chip generation and memory configuration, according to the report. Contract server builders that supply Microsoft, Google, Oracle and other large data-centre operators have already passed the warning on to customers. Nvidia has not confirmed the figures, and Reuters said it could not independently verify them.

The driver is memory, not logic. Server DRAM roughly doubled in price in the first quarter of 2026, and Counterpoint reported quarter-on-quarter increases of 80-90% across DRAM, NAND and HBM. Memory now accounts for roughly 25% of the cost of a high-end rack, and Deloitte expects AI-server DRAM prices to quadruple over the course of 2026.

The dollar figures at stake are large. A Blackwell NVL72 GB200 rack already costs $2.8 million to $3.4 million; inquiries for Vera Rubin systems run $5 million to $7 million per unit. A 15% increase on those totals adds hundreds of thousands of dollars per rack, multiplied across the thousands of units hyperscalers order. The news landed days before Nvidia's 26 August earnings, putting fresh scrutiny on how the company frames margins and supply costs to investors.

The clearest winners are the memory makers. Samsung, SK Hynix and Micron control most of the world's DRAM, NAND and HBM supply, and a quadrupling of AI-server DRAM prices this year hands them pricing leverage they have not had in this cycle. Nvidia, despite sitting atop the AI hardware stack, is exposed to a cost structure it does not fully control.

For Microsoft, Google, Oracle and other hyperscalers, the increase lands directly on capital budgets already stretched by the scale of AI build-outs. Expect those costs to show up eventually in cloud AI pricing and in the economics that labs use to justify model training runs. Watch memory-maker earnings and hyperscaler capex guidance over the next two quarters for signs of how much of this gets absorbed, and how much gets passed on.