Nvidia has told some of its largest customers that the price of servers built around its AI accelerators will rise, with increases reported above 15 percent in many cases. The notice, first reported by Bloomberg on Saturday and confirmed by CNBC, applies to systems built on the company’s flagship Vera Rubin and Grace Blackwell chips and is set to take effect on machines shipped early next year.
The details come from CNBC’s account of the Bloomberg report, which cites people familiar with the process. The Information separately reported that the increase runs to about 17 percent and that it covers Grace Blackwell 300 and Vera Rubin 200 systems due for delivery in 2027. Nvidia has not published its own list prices, so the figures describe what selected customers have been told rather than a public price sheet.
What is changing
The hikes target complete server systems, not bare chips alone. That distinction matters because those systems bundle Nvidia’s accelerators with high-bandwidth memory, networking, and the surrounding board and thermal hardware that AI factories depend on. According to the reporting, the new prices will be set according to chip generation and memory configuration, so a dense Vera Rubin rack and a smaller Grace Blackwell node will not move by the same absolute amount.
Rising memory costs are a central driver. The same reports tie the increase to soaring prices for the HBM stacked on Nvidia packages, where supply has stayed tight as accelerator demand outruns capacity. For buyers, the effect is a higher cost per unit of installed compute at precisely the moment they are scaling data-center footprints.
The move follows a year in which Nvidia has taken a more direct hand in financing and siting the AI data centers that buy its systems. As we covered in our report on the $500 billion AI compute financing push, the company has been tying its hardware roadmap to the buildout of the facilities that consume it. Raising system prices extends that control from where the factories go to what they cost to fill.
Why it matters for buyers
Cloud providers, neoclouds, and enterprises planning 2027 capacity now face a moving cost baseline. A mid-teens percentage increase on accelerator systems reshapes the math behind the AI provider comparison that developers use to pick inference suppliers, because the amortized hardware cost sits underneath every token priced by a cloud. It also raises the stakes for the open-weight and local-AI alternatives that let teams avoid per-token frontier pricing altogether.
Nvidia’s own open-weights effort adds a twist. The company has been expanding its Nemotron family of open models, and a separate wave of reporting this weekend describes a multibillion-dollar arrangement with startup Poolside aimed at strengthening that open-model work. Pricing the proprietary silicon higher while pushing open models downstream is a two-track strategy: capture more margin on the hardware that feeds AI factories, and widen the ecosystem that runs on Nvidia software.
The clearest next signal is Nvidia’s upcoming earnings call, where management is expected to address demand, supply, and pricing. Until then, the practical takeaway for buyers is that 2027 system quotes should be re-checked against these revised numbers before capacity is committed.
