Nvidia Warns Customers of 15% Price Hikes on AI Servers

Nvidia has told some of its largest customers to brace for higher bills. Servers built around the company's newest AI chips will cost more than 15% more on many orders shipping early next year, according to Bloomberg, which first reported the notices on 22 August. The reason is not Nvidia's own margins. It is memory.

The increases land on systems that carry the flagship Vera Rubin and Grace Blackwell processors, the parts data-centre operators are racing to install. Fortune and The Decoder report that the contract manufacturers assembling these machines for Microsoft, Google and Oracle have already passed word down the line to their own buyers.

Why memory is the culprit

The squeeze comes from a shift inside the memory industry. Chipmakers have been moving production capacity away from ordinary DDR5 modules and toward high-bandwidth memory, the specialised stacks that sit beside an AI accelerator and feed it data. That is where the fat margins are right now, so that is where the silicon goes. The result is a shortage of the mainstream parts and a price climb across the board.

The numbers are steep. DRAM contract prices have risen roughly 50% over the course of this year, according to industry figures cited by Business Standard. A single 32GB DDR5 module that sold for about $149 last September now runs closer to $239. A modern AI server can hold dozens of those modules, so the arithmetic adds up quickly.

The cost lands downstream

The timing is awkward for anyone betting that AI compute will keep getting cheaper. Only last week Samsung raised its own foundry prices on the same wave of demand. Meanwhile the labs have been cutting what they charge for the models themselves, with OpenAI trimming the price of its flagship only days ago. Those two trends cannot both run forever. Someone absorbs the difference, and for now that looks like the cloud providers, and eventually the companies renting their servers.

Nvidia has not spelled out the exact figures publicly, and the warnings reached customers through the manufacturers rather than any formal announcement. But the direction is clear. The bottleneck in AI has quietly moved. For a while it was the accelerators themselves. Now, at least for the next few quarters, it is the humble memory chips sitting next to them.