Nvidia has spent the past three years as the company that sets prices in artificial intelligence, not the one that pays them. That changed over the weekend.
Some of Nvidia's biggest customers have been told that the prices of servers containing its AI chips are going up more than 15 percent in many cases, Bloomberg reported on Saturday, Aug. 22. The increases take effect on systems shipped early next year and cover both the current Grace Blackwell racks and the Vera Rubin generation replacing them. How steep the jump runs depends on which chip generation a buyer takes and, tellingly, how much memory is specified. DIGITIMES reported on Aug. 24 that certain flagship configurations slated for 2027 delivery are rising by about 17 percent.
The notices took an indirect route. Nvidia does not sell most of these racks itself, so the contract manufacturers that assemble them for Microsoft, Alphabet's Google and Oracle got the word first and passed it up the chain, people familiar with the process told Bloomberg.
The culprit is not the GPU. It is the memory sitting next to it.
The bottleneck moved
For two years the AI buildout has been narrated as a fight over GPU allocation. That constraint has quietly relocated to DRAM. Nvidia's accelerators are only as useful as the high-bandwidth memory they are paired with, and HBM is essentially a stack of DRAM dies. Samsung Electronics, SK Hynix and Micron Technology account for most of the world's production of both. All three are expanding output. None has caught up.
The pricing data is brutal. Contract prices for server DRAM jumped 53 to 58 percent in the second quarter of 2026 from the previous quarter, according to market researcher TrendForce, which expects another 13 to 18 percent increase in the third. Industry estimates now put memory at roughly a quarter of what a high-end AI rack costs to build, up from a mid-single-digit share two generations ago.
That is the arithmetic behind Nvidia's letter. The company runs a gross margin near 75 percent, meaning it keeps roughly three-quarters of revenue after the cost of its products. A firm with that much cushion choosing to pass costs through rather than absorb them is a statement about who actually holds leverage in the AI supply chain. As Bloomberg framed it, the inability of the industry's most dominant company to hold the line on prices shows how much power the memory makers have accumulated.
Analysts have been forecasting exactly this. "The bigger risk on the horizon is with advanced memory as NVIDIA's recent pivot to LPDDR means it is a customer on the scale of a major smartphone maker," said MS Hwang, research director at Counterpoint Research, whose team has projected a doubling of DDR5 server module prices in a highly constrained scenario. Hwang described the shift as one the supply chain cannot easily absorb.
Supply is not loosening on any near horizon. SK Hynix chief executive Kwak Noh-jung said in July that 2027 will bring the most severe memory supply crunch the industry has ever seen, and that customer demand will outrun his company's capacity well past 2030. Samsung's memory division has warned of significant shortages persisting through at least 2027. Reports in early August indicated that 2027 DRAM and HBM capacity across the three major suppliers is effectively spoken for.
Why It Matters
A 15 to 17 percent surcharge on a rack that already lists in the millions is not a rounding error. Grace Blackwell GB200 NVL72 systems have been shipping in the range of $2.8 million to $3.4 million each, and industry estimates for the Vera Rubin racks that succeed them run materially higher, with memory the single largest contributor to the gap. Scaled across a campus, the numbers get uncomfortable fast: applying the 17 percent figure, Korea JoongAng Daily estimated that building a 1-gigawatt data center could run at least $5 billion.
The strategic wrinkle is that this lands on Nvidia's would-be challengers just as hard. Amazon, Microsoft, Google and Meta are all building in-house accelerators specifically to reduce their dependence on Nvidia. Every one of those chips needs high-performance memory. Escaping Nvidia does not mean escaping Samsung, SK Hynix and Micron. It means becoming more dependent on them. The memory oligopoly collects regardless of which silicon architecture wins.
And it cascades downward. Memory fabs allocate wafers where margins are highest, and right now that is emphatically not consumer DIMMs. Apple and Qualcomm have both said chip shortages forced them to charge more. PC makers have pushed prices up by double digits. Counterpoint senior analyst Ivan Lam warned that the squeeze would spread across the broader consumer electronics ecosystem, with bill-of-materials increases upward of 15 percent on some mid-to-high-end phone models, "eating into margins or affecting growth. It will probably be both."
"If the cost of building AI infrastructure continues to rise, even major tech companies may find it difficult to maintain their current pace of investment," said Shin Joong-ho, head of the research center at LS Securities. "If higher memory costs continue to be passed on to customers through Nvidia's system prices, they could also squeeze returns on AI data center investments."
What to Watch
Nvidia reports fiscal second-quarter results this week, and the questions that matter are no longer about demand. Watch gross margin guidance for a read on how much of the memory bill Nvidia intends to eat versus pass along, and listen for any language about hyperscaler order pacing into 2027. So far the evidence suggests customers are simply paying: nobody in this segment delays a Rubin deployment over a 15 percent premium when the alternative is not deploying at all. The usual relief valve is demand destruction, and it has not opened. If capex committees do start quietly stretching timelines, the first place it surfaces will be the order books of the Taiwanese contract manufacturers that assemble these racks. The second will be TrendForce's fourth-quarter contract pricing, the earliest point at which any softening could plausibly appear.
“If the cost of building AI infrastructure continues to rise, even major tech companies may find it difficult to maintain their current pace of investment.”— Shin Joong-ho, Head of the research center, LS Securities