Nvidia-Powered Servers Rise 15%: The Assemblers Said So, Not Nvidia

The Subject of Saturday's Sentence

Bloomberg reported on Saturday, August 22, that prices for servers carrying Nvidia's AI chips are going up more than 15%. The increase applies to systems shipping in early 2027 across two platforms, Grace Blackwell and Vera Rubin, and its size varies by chip generation and memory configuration.

The outlets that relayed it did not agree on the subject of the sentence. Bloomberg's own text:

"Companies who build the servers under contract for large data center operators such as Microsoft Corp., Alphabet Inc.'s Google and Oracle Corp. have recently notified their customers of the forthcoming increases."

CNBC attributed the notification directly to the chip designer: "The chipmaker has told some of its largest customers that the prices of servers containing its AI chips could move more than 15% higher." Nvidia did not respond to requests for comment.

The difference is not phrasing. A rack's price and a chip's price are two different numbers.

A rack's price sums the GPU module, CPU, system memory, storage, circuit board, cabling and cooling, and the contract manufacturer combines them into one invoice. A rack can rise 15% without Nvidia changing its price list.

How Memory Got Here

The cause is not in dispute: DRAM. Per TrendForce, server DRAM contract prices rose between 53% and 58% in the second quarter of 2026, with a further 13% to 18% expected in the third.

DRAM spent years being treated as a commodity with a falling price; AI data centers ended that. Almost all global output sits with three companies — Samsung, SK Hynix and Micron — all adding capacity, none keeping up.

The squeeze is not demand alone: HBM comes from the same plants and wafers as ordinary DRAM and eats far more capacity per wafer, so leaning into HBM shrinks what is left for server DDR5.

On the buyer's side a reflex has stopped working. Memory was cyclical: when prices rose, buyers deferred orders and ran down inventory. In the data center a server cannot ship without its memory and there is no substitute rack, so the buyer can only accept.

Who Buys the Memory Inside the Rack

HBM goes inside the GPU package, bought by Nvidia, so its price lands in Nvidia's own cost of goods sold; the LPDDR5X, server DDR5 and NAND in the rest of the rack land on the assembler's books.

The Same Increase, Three Times Heavier

Morgan Stanley's bill-of-materials breakdown gives the scale. In the Grace Blackwell NVL72 (B300) memory came to $373,939, 9.3% of the rack's roughly $4 million cost.

In the Vera Rubin NVL72 (VR200) the same line is $2,001,600, 26% of a $7.8 million cost. For comparison, each of the 72 Rubin GPUs costs $55,000 and each of the 36 Vera CPUs $5,000.

The margin assumption is ours; manufacturers do not disclose per-rack margins. But the gap holds regardless: a 75% gross margin and a single-digit one cannot absorb the same dollar the same way.

Not all of the fivefold jump in the memory line is price. The VR200 carries 20.7 TB of HBM4 across 72 GPUs and 54 TB of LPDDR5X across 36 CPUs, against 17 TB of LPDDR5X in the previous generation.

The sources do not separate the two, so "memory got 435% more expensive" is wrong; the memory bill per rack rose 435%. The $1.63 million between the generations is more than 40% of the rack's entire previous cost.

Where Everyone in the Chain Stands

Whichever end of that range proves right, a 15% hardware increase is worth tens of billions inside it — and it was in none of the capacity plans drawn up so far.

The Line to Watch on Wednesday

The report landed on Saturday with U.S. markets closed; QQQ closed Friday up 0.47% and SPY up 0.41%, unrelated. The story gets its first price at Monday's open.

Nvidia reports August 26 after the close. Consensus is $91.9 billion in revenue and $2.08 per share, with third-quarter guidance expected at $103.96 billion; the company's market value is $5.21 trillion.

On the question of where the bill lands, one line decides it: gross margin, which the market expects at 73.5%. Because Nvidia buys the HBM itself, more expensive memory shows up there first.

A margin below expectations means Nvidia has not passed the increase through; above it means most of the bill sits further down the chain, with the assembler and the end buyer. That single ratio determines which side gets revalued.

This piece draws on Bloomberg's August 22 report and on the relays of it published by CNBC, Business Standard, Investing.com and the Korea JoongAng Daily; Bloomberg's own text was behind a paywall and was not read directly, and the quotations were taken only where they matched word for word across more than one independent relay. The rack cost breakdown comes from Morgan Stanley Research's bill-of-materials analysis via Tom's Hardware and Wccftech and is an estimate. Memory contract prices are TrendForce data. Nvidia has not confirmed the report and did not respond to a request for comment.