Even Nvidia can't absorb the AI memory squeeze. Server prices are going up.

Nvidia told its biggest customers this week that AI server systems shipping in early 2027 will cost 15% or more above current prices, blaming memory chip costs for the move. Bloomberg reported the customer notifications on 22 August, and Livemint's re-report reached the same conclusion. The increase applies to flagship systems built on Nvidia's next-generation Vera Rubin and current-generation Grace Blackwell chips; the precise figure will vary by chip generation and memory configuration. Systems already shipping are not affected.

Nvidia runs at roughly 75% gross margin, about 75 cents of every revenue dollar remaining after production costs, and is the dominant supplier of AI accelerators. When a chip leader with that kind of margin cannot absorb a supplier cost, the AI cost story is no longer being set in the GPU market. DRAM makers, the small group of suppliers that produce the dynamic random-access memory chips used alongside AI processors, are now setting terms.

The customer-facing mechanics are familiar. Contract manufacturers, the assemblers who build finished AI servers for major data centre operators, relayed the upcoming increase to their customers. CNBC's coverage names Microsoft, Google, and Oracle as illustrative buyers, not an exhaustive list, and confirms that the new prices will show up on quotes for systems shipping early next year.

The reason matters more than the headline number. Memory is the binding cost. Three companies, Samsung Electronics, SK Hynix, and Micron, dominate global DRAM supply, and the supply expansion for high-bandwidth memory (HBM), the stacked DRAM chips that sit next to AI accelerators on the same package, has lagged the surge in AI infrastructure demand. A single AI server pulls far more DRAM than a conventional server because every model in training or inference is constantly moving data between processor and memory. When HBM is short, the price of every system goes up regardless of who assembles it.

This is the first public confirmation that even Nvidia is passing the cost through rather than absorbing it. The company has been the price-setter in AI hardware for two years, partly because customers needed its chips more than the chips needed any one customer. Memory is the inverse: Samsung, SK Hynix, and Micron are not undercutting each other on HBM. They are rationing capacity, and their order books are full. Passing the cost through is what a price-setter does when its own input costs are being set by someone else.

The Apple and Qualcomm datapoints in the original report, both companies saying in recent weeks that chip shortages forced product price increases, are the same mechanism hitting different end markets. Consumer electronics, mobile devices, and AI accelerators all draw on overlapping memory and foundry capacity. TSMC, the Taiwan Semiconductor Manufacturing Company foundry that fabricates Nvidia's chips, has continued to fall short of customer demand, which keeps pressure on the whole stack. Fortune and Yahoo Finance both emphasised the 15% figure in their coverage of the same Bloomberg story.

The most useful update for a non-beat reader is this: from here, the AI cost story is best tracked through memory makers, not through Nvidia. If HBM capacity expands faster than expected, the 15% surcharge gets negotiated down. If it does not, the surcharge is the floor, and downstream cloud customers will pass it on to the workloads, the SaaS contracts, and eventually the consumer-facing products that run on top of them.

Nvidia's next earnings call will be the cleanest way to see whether the company frames the increase as a one-quarter wobble or as a longer-running reset. The answer will determine whether the AI hardware bill keeps climbing through 2027, or starts to ease.