Nvidia's AI Servers Are Getting 15% More Expensive. Watch the Margin

Nvidia has warned its largest customers that servers built around its newest AI chips will cost more than 15 percent more in many configurations, Bloomberg reported on 22 August. The increases cover Vera Rubin and Grace Blackwell systems shipping early next year.

Four days later, Nvidia reports earnings. The increase cannot appear in the quarter being reported, but management will likely be asked whether it is simply passing on higher memory costs or using the shortage to collect additional margin.

That question matters because Nvidia described its pricing differently in its latest quarterly filing:

Our product and solution pricing generally does not fluctuate with short-term changes in our costs.

Memory prices have risen enough to test that policy.

Memory suppliers now have the leverage

An AI accelerator is only as useful as the memory it is paired with, and memory is the hardest component in the chain to obtain. Three companies make almost all of it: Samsung, SK Hynix and Micron. All three have sold their output well into next year, so buyers compete for allocation rather than negotiate on price. TrendForce forecast conventional DRAM contract prices rising 55 to 60 percent quarter on quarter in the first quarter of 2026, with server DRAM up more than 60 percent, followed by a further 13 to 18 percent in the third quarter.

The economics have already reversed. Micron reported a GAAP gross margin of 84.6 percent for its quarter ended 28 May, and guided to approximately 86 percent for the next. Nvidia's GAAP gross margin for its quarter ended 26 April was 74.9 percent. SK Hynix reported operating profit of 60.5 trillion won for the June quarter, up 557 percent on the year.

For most of the past three years, Nvidia supplied the scarce component in an AI system. The constraint has now moved to the memory beside it, which Nvidia buys from three suppliers that have already committed most of next year's production.

Switching from Nvidia does not fix it

A price increase of that size would normally push buyers towards a competitor. There is no obvious alternative, because the rival products are built on the same memory. AMD's Instinct MI455X carries 432GB of HBM4 per accelerator, and a single Helios rack holds 31 terabytes of it. The chips the cloud companies design for themselves are in the same position: Google's TPUs, Amazon's Trainium and Microsoft's Maia all rely on high bandwidth memory.

Estimates of Nvidia's share of AI accelerators run from about 70 percent to about 86 percent, depending on whether hyperscaler custom silicon is counted, in a data centre accelerator market worth more than $200 billion this year. A memory shortage is unlikely to move much of that share, because it raises the cost of building every one of these systems at the same time.

Server makers cannot absorb the cost

The companies that assemble the systems have the least room. Dell recognised $16.1 billion of AI server revenue in its April quarter, and its gross margin fell to 17.8 percent from 21.1 percent a year earlier. Super Micro reported a 17.5 percent gross margin on $11.1 billion of sales for the June quarter. Both have told investors they are passing component cost increases on to customers, which is the route by which memory contracts agreed in Korea end up in the budgets of data centre operators in the United States.

The increase will not appear on Wednesday

Nvidia guided to July-quarter revenue of $91.0 billion, plus or minus 2 percent, and a GAAP gross margin of 74.9 percent, plus or minus 50 basis points. The price increase applies to systems shipping early next year, so it cannot appear in a quarter that ended in July. Any effect will instead emerge in Nvidia's guidance for the quarters that follow.

Watch what happens to the margin

Nvidia's quarterly filing says nothing about the level of that margin, which has moved: 75.0 percent for the year ended January 2025, 71.1 percent for the year ended January 2026, and 74.9 percent in the April quarter. All three are GAAP. The fiscal 2026 figure was pulled down by a $4.5 billion charge on H20 inventory and purchase obligations.

If Nvidia is recovering a cost increase and no more, it adds roughly the same number of dollars to revenue and cost of revenue. The pass-through generates no additional gross profit, while the gross-margin percentage falls because revenue has increased without a corresponding increase in gross profit. Guidance for a lower margin would therefore be consistent with the company passing memory costs along rather than a sign that demand had weakened.

That leaves the question for Wednesday's call. If Nvidia guides to a lower margin, the increase may be doing little more than recovering its memory bill. If it holds the mid-70s, that would suggest Nvidia is recovering more than the memory bill, although product mix could obscure the result.

The servers are getting at least 15 percent more expensive either way. What matters is how much of that increase Nvidia keeps.