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 and covers two Nvidia platforms: Grace Blackwell and Vera Rubin. The size of the increase varies by chip generation and by how much memory the system carries.

The outlets that relayed the story did not agree on the subject of the sentence. In Bloomberg's own text, the party doing the notifying is described this way:

"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 changed the subject and 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 has said nothing of its own: "Nvidia representatives didn't respond to requests for comment."

The difference is not a matter of phrasing. The price of a rack and the price of a chip are two different numbers, and the distance between them is the story here.

A rack's price is the sum of the GPU module, the CPU, system memory, storage, the printed circuit board, cabling and cooling. The party that combines those line items into a single invoice is not Nvidia but the contract manufacturer that assembles the system. A rack's price can rise 15% without Nvidia changing its price list at all.

How Memory Got Here

The cause is not in dispute: DRAM. According to the market researcher TrendForce, server DRAM contract prices rose between 53% and 58% in the second quarter of 2026 against the first. The same firm expects a further increase of 13% to 18% in the third quarter. DRAM — the working memory a computer uses while running — spent years being treated as a commodity with a falling price. AI data centers ended that classification.

Almost all global DRAM output sits with three companies: Samsung, SK Hynix and Micron. All three are adding capacity and none can keep up. For the first time in a long while, the seller sets the price.

The squeeze does not come from demand alone; it comes from a choice inside the fabs. HBM is made in the same plants, from the same wafers as ordinary DRAM, and it consumes far more capacity per wafer. When a producer leans into HBM, the capacity available for server DDR5 shrinks. AI demand therefore pushes memory prices up from two directions at once: it raises the price of HBM itself, and it makes conventional server memory scarce because capacity has migrated to HBM. The 53% to 58% jump TrendForce measured in the second quarter is the sum of both pressures, not the movement of a single product.

The consequence on the buyer's side is that a long-standing reflex has stopped working. Memory has traditionally been a cyclical product: when the price rises, the buyer defers orders, runs down inventory, and comes back when the price falls. In the data center, that option does not exist. A server cannot ship without the memory that goes inside it, and there is no substitute rack to put in its place. When the buyer cannot do anything but accept the price, the negotiating table has only one side.

The chart is not evidence for a claim; it is the backdrop. Memory producers' shares moved in the same direction over this period as the bill the buying side was paying. The question is where in the chain that bill actually lands.

Who Buys the Memory Inside the Rack

The decisive detail is this: no single company buys all the memory in a rack.

HBM — high-bandwidth memory — goes inside the GPU package, and Nvidia buys it and hands it to TSMC. A rise in the HBM price therefore lands in Nvidia's own cost of goods sold. The LPDDR5X, server DDR5 and NAND storage in the rest of the rack are bought by the assembler, and a rise in those prices lands on the assembler's books. That the notification came from the assembler is not odd; it is what you would expect.

The Same Increase, Three Times Heavier

Morgan Stanley's bill-of-materials breakdown gives the scale. In the previous generation, the Grace Blackwell NVL72 (B300), the memory line came to $373,939, or 9.3% of the rack's roughly $4 million cost. In the Vera Rubin NVL72 (VR200) the same line is $2,001,600, or 26% of a $7.8 million cost. For comparison: each of the 72 Rubin GPUs in that rack costs $55,000, and each of the 36 Vera CPUs costs $5,000.

When the share changes, the same price move produces a different result.

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

How Much Each Line Item Moved

A caution is needed here. 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; the previous generation carried 17 TB of LPDDR5X. Part of the line item comes from memory getting more expensive and part from more memory being installed. The sources do not separate the two. Saying "memory got 435% more expensive" is therefore wrong; the accurate statement is that the memory bill per rack rose 435%.

Those two lines explain on their own why server pricing is back on the negotiating table. The $1.63 million between them is more than 40% of the rack's entire previous cost.

Where Everyone in the Chain Stands

The range on that last line is wide; estimates for 2026 hyperscaler capital spending run from $630 billion to $725 billion. Whichever proves right, a 15% hardware increase is a line item worth tens of billions inside that budget, and it was in none of the capacity plans drawn up so far.

The Market Has Not Spoken Yet

The report landed on Saturday, with U.S. markets closed. QQQ closed Friday up 0.47% and SPY up 0.41%; that move has nothing to do with this news, which had not yet been published. This story gets its first price at Monday's open.

The three months the chart covers do not include this report. The week that will is this one.

The Line to Watch on Wednesday

Nvidia reports on Wednesday, 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.

But on the question of where the bill lands, one line decides it: gross margin. The market expects 73.5%. Because Nvidia buys the HBM itself, more expensive memory shows up there first. A margin below expectations means Nvidia has not yet passed the increase through to customers. A margin above it means most of the bill is sitting further down the chain — with the assembler and the end buyer. That single ratio determines which side of the chain 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.