For three years Nvidia has been the company that sets the price of artificial intelligence. This week it became the company that pays one. Bloomberg reported over the weekend that Nvidia’s largest customers have been told prices on servers built around its AI chips are going up by more than 15% in many configurations, and the culprit is not silicon Nvidia designs. It is memory.
The One Input Nvidia Does Not Control
The increases hit systems shipping in early 2027 and cover the flagship parts, including Vera Rubin and Grace Blackwell configurations, with the exact number varying by chip generation and by how much memory a given rack carries. CNBC confirmed the reporting on Friday, and Fortune’s account pinned the driver squarely on the cost of high-bandwidth memory. Some estimates run higher: Tom’s Hardware put the top of the range near 17% depending on configuration.
That distinction matters more than the percentage. Nvidia’s pricing power has always rested on owning the scarce thing. It designs the GPU, it owns CUDA, it decides who gets allocation. Memory is the piece of the stack it buys, from a supplier base of three, in a market where those three have spent two years deciding that AI customers are worth more than everyone else.
So the pass-through is not a strategy. It is an admission. Nvidia can engineer around a lot of constraints. It cannot engineer around SK Hynix telling it what a stack of HBM costs.
The market understood the implication immediately, and read it as bad news for the memory names rather than good. Micron Technology dropped 5.8% on Monday, Advanced Micro Devices fell more than 3%, and Broadcom lost more than 2% as the Nasdaq closed down 0.76%, per CNBC’s session wrap. That reaction looks backwards until you think about who eats the cost. A 15% sticker increase on an AI rack is a demand test, and the first place demand gets tested is the component vendors sitting one rung below Nvidia in the same bill of materials.
What Memory Actually Costs Now
The numbers behind the hike are genuinely startling, and they have been building since late last year while the AI conversation stayed fixed on GPUs.
- Conventional DRAM contract prices are running up roughly 55% to 60% quarter over quarter, with NAND flash up 33% to 38%, according to TrendForce, which has been tracking the HBM4 transition and its spillover into general memory pricing.
- A 64GB RDIMM, the workhorse server memory module, went from about $450 in the fourth quarter of 2025 to more than $900.
- Samsung took a 32GB DDR5 module from $149 to $239, a 60% move on a single product line, and warned publicly that shortages would drive industry-wide increases through 2026.
- SK Hynix has described its HBM, DRAM and NAND capacity as essentially sold out for the year, and Micron walked away from consumer memory entirely to serve enterprise and AI buyers.
None of that unwinds soon. New fabs from Samsung and SK Hynix are not ramping before the second half of 2027, and Micron’s $200 billion American build-out does not start producing DRAM until the middle of that year. The supply answer arrives roughly 18 months after the problem. We flagged the setup in June when Micron jumped 10% on HBM4 certification, and again when a memory-led selloff hit Samsung and SK Hynix and dragged Asian markets with it.
The Bill Lands on Microsoft, Google and Oracle
Follow the paperwork and the story gets more interesting. The notifications did not go out from Nvidia’s sales team to end customers. They went out from the contract manufacturers that assemble servers for the hyperscalers, which means Microsoft, Google and Oracle are learning the new price the way a homeowner learns the contractor’s lumber went up.
Those three have spent the year committing to capital expenditure numbers that already assumed aggressive scaling. We put the combined hyperscaler figure at $886 billion in AI capex earlier this month. A 15% increase on the server line of that budget is not a rounding error, and it arrives at a moment when every one of those companies is being asked by its own shareholders to show returns on the previous round.
They have three options and none of them are comfortable. Absorb the cost and watch cloud margins compress. Pass it to enterprise customers and test whether AI demand is as price-insensitive as everyone has assumed. Or buy fewer racks, which is the outcome Nvidia can least afford heading into a quarter where growth expectations are close to a double.
The third option is why the memory story is really an Nvidia story. Every dollar of memory inflation that Nvidia passes through is a dollar that does not have to be spent on GPUs, and Nvidia is the only participant in this chain with an incentive to keep the total bill from breaking demand.
Wednesday Night Is the Real Test
Nvidia reports fiscal second-quarter results after the close on Wednesday, and the timing could not be less convenient. Consensus sits near $92 billion in revenue against company guidance of $91 billion plus or minus two percent, with earnings around $2.09 a share, roughly double the $46.74 billion and $1.05 posted a year ago.
Nobody is worried about the top line. The question analysts will actually press on is the one the price hike raises: what happens to data-center gross margin, currently modeled in the mid-70s, when the fastest-inflating input in the bill of materials is one you buy rather than make. Jensen Huang can tell that story two ways. He can frame the increase as evidence of demand so strong that customers will absorb anything, or he can frame it as a temporary supply distortion that normalizes when the 2027 fabs light up.
Watch which one he picks. If the price of an AI rack is now set in Icheon and Hwaseong rather than Santa Clara, the most valuable company in the world has a supplier problem it has never had to explain before.