The notice arrived over the past week, and it carried a message that procurement teams at the world’s largest cloud providers had been dreading: Nvidia Corp. is raising prices on its newest AI servers by more than 15%. The notifications went out to a number of the chip maker’s largest customers, according to people familiar with the matter, who described the communications as a heads-up rather than a negotiation.
The increases apply to systems built on the Vera Rubin and Grace Blackwell architectures, the chip lines that anchor Nvidia’s data-center lineup through next year. New pricing takes effect with shipments beginning early next year, the people said. Nvidia told customers the move reflects a surge in the cost of the memory that surrounds its processors, which has become the binding constraint on how many servers the industry can build.
The root cause is high-bandwidth memory, or HBM, the ultra-fast memory stacked beside AI accelerators to feed them data. Supply of HBM is tight, prices have climbed sharply, and the cost of conventional DRAM and the packaging materials used to assemble AI servers is rising alongside it. The pressure is moving down the supply chain in sequence: memory suppliers raise prices, chip makers pass them along, server vendors follow.
CNBC, Reuters and Fortune subsequently confirmed the notifications, citing people familiar with the matter. Upstream memory production lines are already booked out to 2028, according to those reports, a signal that the constraint is structural rather than seasonal. For Nvidia’s customers, that means today’s increase is unlikely to be the last.
Nvidia reports fiscal-second-quarter earnings on Aug. 26. The price increase lands after the quarter closed, so it will do little for the numbers due next week, but executives have telegraphed that margin support from pricing is on the way. Analysts said the question for the earnings call is how much of the increase Nvidia can keep, given that some customers signed long-term contracts months ago at lower prices.
For hyperscalers — Microsoft Corp., Amazon.com Inc., Alphabet Inc.’s Google and Meta Platforms Inc. among them — the increase arrives at an awkward moment. These companies are pouring tens of billions of dollars a year into AI infrastructure, and their own investors are watching capital spending closely. A 15% rise in the cost of the industry’s most sought-after servers compounds the pressure on cloud profit margins at exactly the moment those businesses are trying to prove the AI buildout pays.
The memory crunch traces back to the peculiar economics of AI hardware. A single AI server carries roughly six times the DRAM and eight times the NAND of a traditional server, and its most expensive component after the accelerator is the HBM stack beside it. HBM supply rests with three companies — SK Hynix Inc., Samsung Electronics Co. and Micron Technology Inc. — and all three have said their capacity is effectively sold out. When the three producers of an essential input are full, the price of that input moves in only one direction.
Nvidia could have absorbed the rising memory bill itself. The company’s gross margin, though down from the peaks of 2024, remains among the highest in the semiconductor industry, and its accelerators still sell out months in advance. Instead it chose to pass the cost through, a decision that reflects the balance of power between the chip maker and its buyers: with demand still outstripping supply, Nvidia sees little reason to eat the increase.
The move is the latest evidence that the bottleneck in the AI buildout has shifted from compute to memory. A year ago the industry worried about whether Nvidia could ship enough accelerators. Today the accelerators are ready, and the industry cannot make enough of the memory that surrounds them. Suppliers have responded by raising prices and promising capacity additions that will not arrive until 2027 or 2028.
For customers, the calculus is uncomfortable. They can wait for prices to normalize, gambling that the memory crunch eases; or they can lock in capacity now at higher prices, protecting their AI roadmaps at the cost of worse near-term margins. Most are choosing to pay, according to people familiar with the matter, because for a company racing to deploy models, the alternative — no servers at all — is worse.
The increase also gives cover to Nvidia’s rivals. Advanced Micro Devices Inc. and the custom chip efforts inside Amazon, Google and Microsoft all compete for the same memory supply, and none of them is in a position to undercut Nvidia on price while their own input costs are climbing. The pricing power Nvidia is exercising today is a function of its supply position, analysts said, not just its brand.
Whether the increase holds depends on the memory market, not on Nvidia. If HBM supply catches up with demand faster than expected, customers will have negotiating room again by the time next year’s shipments begin. If the crunch persists — and the 2028 booking horizon suggests it will — today’s 15% will look like an opening bid.
The answer arrives with the first shipment batches early next year, when the new prices appear on invoices and the market sees how many customers pushed back. For now, the direction is clear: the cost of building AI infrastructure is going up, and the industry is being told to get used to it.
