NVIDIA has notified its largest customers that servers built around its AI chips will cost more than 15 percent extra in many cases, according to reports published on August 22. Configurations that include Vera Rubin and Grace Blackwell are covered, and the new pricing applies to systems shipped next year. The reason is the continued climb in the price of the memory that its GPUs and systems depend on.

The Increase Varies by Generation and Memory Configuration

The reported increase is not a flat rate. How much a customer pays extra depends on which chip generation is used and how much memory is installed. The new pricing takes effect on systems shipped next year, so equipment already delivered or arriving before the end of this year is not part of the picture.

Vera Rubin, one of the platforms named, is the rack-scale generation NVIDIA brought to market in 2026. It packs 72 Rubin GPUs and 36 Vera CPUs into a single liquid-cooled rack, with up to 288GB of HBM4 per GPU. Across the full rack that adds up to 20.7TB of HBM4, plus 54TB of LPDDR5X, for a total of 74.7TB of fast memory. Grace Blackwell, the preceding generation, is likewise designed around large HBM capacity.

When a system is built to carry that much memory, any movement in memory pricing feeds straight through to the bill of materials. The cost of manufacturing the GPU die can stay flat and the sticker price will still move. NVIDIA has not publicly commented on the reports so far.

Memory Prices Have Changed by an Order of Magnitude in Two Years

The underlying memory market is no longer creeping upward. Contract prices for conventional DRAM rose 90 to 95 percent quarter over quarter in the first quarter of 2026, and climbed another 58 to 63 percent in the second quarter, according to market researchers. Server DRAM finally cooled to somewhere around 13 to 18 percent in the third quarter, but that is a slower rate of increase, not a plateau.

Zoom out and the numbers get larger. One bank's research arm estimates DRAM prices will have risen more than 400 percent between the start of 2024 and the end of 2026. AI data center construction and hyperscaler procurement are absorbing a disproportionate share of global memory capacity, and memory makers are steering output toward the more profitable server segment. That leaves AI servers, which need high-capacity RDIMMs in volume, standing exactly where the price pressure lands hardest.

For the companies assembling those servers, absorbing that component inflation indefinitely is not realistic. Seen in that light, a price revision from the GPU supplier follows naturally from the same trend.

The Bill Lands With the Cloud Providers

The notification did not go from NVIDIA directly to end customers. It reportedly travelled through the contract server manufacturers that supply the large data center operators, with Microsoft, Google and Oracle named among those affected. In other words, the cost lands on the cloud providers pouring capital into AI infrastructure.

These companies have compute expansion plans mapped out for 2027. If the same budget buys 15 percent fewer racks, some of those plans need reworking. Decisions about whether to prioritize training clusters or secure inference capacity first become considerably tighter.

How far the increase travels downstream is the open question. If it shows up in cloud GPU instance rates or in API pricing for AI models, it stops being an infrastructure story. That said, AI companies are still in a phase of growing their user bases, so passing higher costs straight through to list prices is far from certain. For now the cloud providers and AI vendors look likely to swallow it, though a prolonged period of elevated memory prices would make that harder to sustain.

Summary

What makes this increase notable is that the driver is not the GPU itself but memory, a separate component. With DRAM prices moving by multiples over two years, systems like Vera Rubin and Grace Blackwell that carry enormous amounts of HBM and LPDDR take the impact directly. Applying the new pricing to systems shipped next year also means that 2027 AI infrastructure budgets are being rewritten right now. Alongside model performance and power availability, the unglamorous matter of memory pricing has become worth watching as a factor that sets the pace of AI development.