Nvidia has told the contract manufacturers that build its AI servers to prepare customers for a price increase of more than 15% on many configurations, according to reports citing people familiar with the discussions. The higher prices apply to systems shipping from early 2027 onward and cover both the current Grace Blackwell platform and the upcoming Vera Rubin generation, Nvidia's next flagship AI accelerator line.
The customers being notified include some of the world's largest cloud operators — Microsoft, Google and Oracle among them — which together account for a large share of the AI infrastructure being built this year. The increase, first detailed by Tom's Hardware and traced to Bloomberg reporting, was described only as "more than 15%," since exact pricing varies by chip generation, memory configuration and individual contract terms.
Nvidia has not issued a public statement confirming the figure. But the arithmetic is not complicated: the company currently runs a gross margin near 75%, and rather than absorb rising costs, it appears to be passing them down the chain to the hyperscalers that buy its servers by the tens of thousands.
A memory shortage even Nvidia can't dodge
The driver, according to the reporting, is not the cost of Nvidia's own chips but the price of the memory that surrounds them. High-bandwidth memory and server DRAM prices have been climbing at record rates through 2026, as Samsung, SK hynix and Micron — which together produce nearly all of the world's DRAM — shift capacity toward AI-grade memory and away from consumer products such as PC and phone chips, as Notebookcheck reported. Demand from AI data centers has simply outpaced the industry's ability to add capacity, giving memory makers unusual leverage over customers that otherwise hold most of the power in tech supply chains.
That dynamic was on display this week at Hot Chips 2026, the annual chip conference held at Stanford, where Samsung unveiled a new memory design that embeds computing logic directly inside DRAM chips, cutting how much data needs to move between memory and processor in the first place. It is one sign the industry sees the memory crunch as a structural shift rather than a temporary shortage.
Hyperscalers left with few good alternatives
For Microsoft, Google and Oracle, the timing is awkward. All three have spent the past two years racing to expand AI compute capacity for enterprise customers and their own AI products, and all three are simultaneously trying to reduce their dependence on Nvidia by developing custom silicon — Google's TPUs, Amazon's Trainium chips and, most recently, OpenAI's Broadcom-built Jalapeño inference chip, reportedly taped out on TSMC's advanced N3P manufacturing process earlier this year. But custom accelerators still need memory, and the same three suppliers dominating DRAM and HBM production hold leverage over homegrown chip programs just as much as they do over Nvidia's.
That leaves cloud providers with limited room to route around the increase. Passing the cost on to their own customers — the businesses renting GPU capacity to train and run AI models — is the most direct option, and would mark one of the clearest signs yet that AI infrastructure spending is beginning to show up in the price of the products built on top of it.
Nvidia has not disclosed exactly when affected customers will see the new pricing take effect beyond "early 2027," and it remains unclear whether the increase will apply uniformly or vary by customer and volume commitment. Investors will get an early read on the numbers at Nvidia's next quarterly earnings call; until then, how much of the increase hyperscalers ultimately absorb versus pass along is likely to remain a central question for an AI industry whose economics increasingly hinge on the cost of components it cannot make for itself.
