
Photo: Dell Inc., 2016, via Wikimedia Commons (CC BY-SA 2.0)
If you sign a contract for AI servers today, the price on it may not be the price you pay. The Financial Times reported this weekend that Dell, HPE, Lenovo and Supermicro are asking customers for as much as 45% more for Nvidia-powered AI servers, months after deals were signed and just as the machines are ready to ship. Trading firms building their own data centers, including Jane Street, Optiver and Hudson River Trading, are among the buyers hit, according to the report.
The short answer: memory. JPMorgan research cited by the FT puts the rise in memory costs at more than 400% since the start of 2024, and server makers say they can no longer absorb it on orders priced last year. One buyer has already gone to court: XTX Markets is suing Dell in Dublin over a demand for $62 million to $70 million more on a $549 million server order. The same squeeze is now showing up in laptops, phones and even cars.
Here is what was reported, what the court papers say, why memory became the most expensive part of a server, what the vendors themselves have said, and the two earnings calls this week that should show how far it goes.
What the FT reported
The FT’s sources described price increases of 25%, 30% and 45% on AI server orders that had already been agreed. One person put the rise for GPU servers at 20% to 30% or more, and said CPU nodes, the general-purpose servers that sit alongside the GPU racks and carry a lot of memory, had tripled in price or more. Those are different machines, so the figures are a range across orders, not one across-the-board increase.
According to the report, some contracts now start with an "indicative" price, which works as an estimate rather than a promise, and buyers have largely paid the higher bills because they need the compute. HPE and Supermicro declined to comment to the FT, and Dell and Lenovo did not respond. None of the four has said publicly how many orders have been repriced.
A quote is now closer to an estimate than a promise. For years, enterprise hardware buyers negotiated a price, signed, and waited for delivery. Memory has broken that habit.
The test case: XTX Markets v. Dell
The clearest public example is a lawsuit filed in Ireland in June. XTX Finland OY, part of the London-based algorithmic trading group XTX Markets, sued Dell Products Unlimited Company in the High Court in Dublin, and the case was admitted to the fast-track Commercial Court in mid-June, according to The Irish Times.
- The order: 1,680 servers for a 15,000-square-meter data center, at an agreed price of about $549 million. That works out to roughly $327,000 per server.
- The price lock: XTX says it paid an extra $5 million above the quoted price specifically to avoid exposure to future price increases.
- The demand: after delivery slipped by two months, Dell said higher component costs meant XTX would have to pay an additional $62 million to $70 million, or about 11% to 13% more. That is roughly $37,000 to $42,000 extra per server, and 12 to 14 times the $5 million XTX paid for price certainty.
- Dell’s side: Dell disputes XTX’s claims. It said it was not willing to lose tens of millions of dollars on the deal, could not deliver on the terms it thought valid when the agreement was signed in November 2025, and that other customers facing increases had gone to other suppliers and then come back and accepted its price adjustments.
- Where it stands: the servers are being delivered, about $70 million sits in escrow, and the judge adjourned the case to December. No ruling has been made.
XTX director Cameron Scott Hill said in an affidavit that the company was shocked by the increase, given it had paid to avoid exactly that, and that Dell told it it was "embarrassed" by the situation. The FT says the court filings also record Dell telling XTX that it "was not able to absorb those increased costs" from higher memory prices.
Why memory is behind it

Photo: 4300streetcar, 2025, via Wikimedia Commons (CC BY 4.0)
An AI server is mostly GPUs, but it also needs large amounts of ordinary DRAM, flash storage and, inside each GPU package, stacks of high-bandwidth memory (HBM). All three come from the same few chipmakers, mainly Samsung, SK hynix and Micron, and they are being pulled in different directions at once.
HBM is eating the wafers
HBM sells for far more than ordinary DRAM, so memory makers keep moving production toward it. Research firm TrendForce said in September that AI server demand keeps HBM and conventional DRAM competing for limited wafer capacity, keeping memory supply tight through 2027. It now expects the blended average selling price of HBM to rise 121% in 2027 compared with 2026.
Memory is now most of a regular server’s bill
On HPE’s fiscal first-quarter call on March 9, 2026, CEO Antonio Neri said DRAM and NAND flash already made up over half of the bill of materials of a traditional server, and that the share would keep rising. He also described triple-digit percentage price increases across the memory industry between the fourth quarter of 2025 and the first quarter of 2026. That is why the CPU nodes in the FT report rose so much faster than the GPU servers: in a GPU server, the GPUs dominate the cost, while in a CPU node, memory does.
Prices are still rising, just more slowly
TrendForce’s latest quarterly forecasts show memory contract prices still going up in the last three months of 2026, though by less than earlier in the year. Storage is the first place where the trend may turn.
| Memory type | TrendForce forecast | Published |
|---|---|---|
| Conventional DRAM contract prices | Up 10% to 15% in Q4 2026 | Sept. 30, 2026 |
| PC DRAM contract prices | Up 13% to 18% in Q4 2026 | Nov. 11, 2026 |
| NAND flash contract prices | Up 8% to 13% in Q4 2026, then down 0% to 5% in Q1 2027 | Nov. 19, 2026 |
| HBM blended average price | Up 121% in 2027 | Sept. 29, 2026 |
The constraints remain the same. DRAM, DRAM, DRAM followed by NAND, NAND, NAND.
Jeff Clarke, Dell Technologies, fiscal second-quarter earnings call, Sept. 1, 2026
As an Amazon Associate, Aglena earns from qualifying purchases.
What the server makers have said
None of the four companies has commented on the FT’s report, but all of them have talked about memory costs on recent earnings calls.
HPE: repricing is in the contract
HPE is the only one that has spelled out the mechanism. In February it told partners it was cutting quote validity to 14 days for most customers and changing its terms so server and GreenLake orders could be repriced until shipment, CRN reported. On the March call, Neri said HPE had "amended our quoting terms with the right to reprice existing orders for commodity cost increases between quoting and shipment," that it raised prices in November, December and January, and that customers can cancel an order before it ships. He said HPE expects elevated memory prices to "persist well into 2027."
Dell: a record backlog, and inflation inside it
On its Sept. 1 call, Dell reported a record $60.9 billion in AI orders for the quarter, $16.4 billion in AI server revenue and a record $95 billion AI backlog. Jeff Clarke said demand had outrun supply for two quarters in a row, that servers "cost more than they did last quarter and the quarter before and the quarter before," and that part of the growth in Dell’s server business is "driven by the price increases as our input costs continue to go up."
Lenovo: higher prices for PCs too
Lenovo reported on Nov. 19 that its Infrastructure Solutions Group had record revenue of $11.3 billion, up 178% from a year earlier, with GPU server revenue up more than 300%. On the call, Luca Rossi, president of Lenovo’s Intelligent Devices Group, said he expects PC average selling prices to keep growing, "also as a consequence of commodity costs going up." Supermicro has not commented on the report.
Why buyers mostly pay
For a trading firm or an AI lab, a server that arrives late can cost more than the price increase. Demand for AI hardware still exceeds supply, so a buyer who walks away goes to the back of someone else’s queue, and the next supplier faces the same memory bill. That is the leverage Dell described in court when it said customers who left came back and accepted the new prices.
A quote is now closer to an estimate than a promise.
Big buyers are not immune either. Elon Musk wrote on X early Sunday that Tesla’s in-car computer "AI4 costs twice as much now due to memory and storage price hikes." The same forces that push up an AI server’s price push up anything with a lot of memory in it. And the sums at stake keep growing, as the $44 billion SpaceX chip financing showed this week.
What it means for laptops, phones and PC parts
The server fight is between big companies, but the memory shortage behind it reaches ordinary buyers too. TrendForce’s recent reports describe how device makers are coping:
- Laptops are selling less and shipping with less: notebook shipments fell 15.8% from the previous quarter in the third quarter of 2026, according to TrendForce, and some PC makers plan lower-spec memory and CPUs with fewer cores in 2027 models. PC brands are also cutting SSD capacity in mainstream models to hold down costs.
- Budget phones are using recycled memory: TrendForce estimates that reballed memory, lower-cost recovered chips fitted with new solder balls, will be in about 3% to 4% of smartphones in 2026 and 7% to 9% in 2027, mostly in entry-level models.
- Storage may ease first: TrendForce expects NAND flash contract prices to start falling in the first quarter of 2027, as bit supply grows 25% to 30% and Chinese cloud providers cut orders, and says spot prices have already dropped more than 20% recently. It still expects DRAM prices to keep climbing into early 2027.
If you are buying a laptop or phone this season, check the RAM and storage figures, not just the model name; a cheaper configuration may be how a maker held the price. Prices for SSDs may come down sooner than prices for RAM.
What to watch this week
Two of the four companies named in the report have earnings calls in the next two days, and analysts are likely to ask about repricing directly.
Monday, Nov. 23: HPE reports fiscal fourth-quarter results, with a call at 2 p.m. Pacific. Watch for any update to Neri’s view that memory prices stay high well into 2027.
Tuesday, Nov. 24: Dell reports fiscal third-quarter results, with a call at 1:30 p.m. Pacific. Its AI backlog was $95 billion at the end of the last quarter; how much of it carries old prices is the open question.
December: XTX Markets v. Dell returns to the Commercial Court in Dublin. A ruling on whether a paid-for fixed price can be reopened would matter well beyond one trading firm.
What we still don’t know
The FT did not say which supplier sought the 45% increase, how many contracts have been repriced, or whether the affected contracts allowed changes before delivery; buyers it approached did not confirm their terms. Dell’s demand to XTX, at 11% to 13%, is well below the top of the reported range. It is also unclear how much of the increase reflects memory costs and how much reflects suppliers’ pricing power in a market where buyers have few alternatives. For background on why the chips inside these servers matter so much, see our explainer on why GPUs became the engine of AI.
Further reading

Samsung 990 PRO 2TB SSD
A fast PCIe 4.0 drive built on the same Samsung NAND flash whose prices have swung with AI demand this year.

Chip War
Chris Miller’s history of the semiconductor industry, including how a handful of companies came to make the world’s memory chips.

NVIDIA Jetson Orin Nano Super Developer Kit
A small Nvidia AI computer for running models at home, with 8GB of memory shared between its CPU and GPU.

The Nvidia Way
Tae Kim on how Jensen Huang built the company whose GPUs sit at the center of every AI server in this story.

Seagate Portable 2TB Hard Drive
A spinning-disk backup drive, a simple way to add storage without paying for flash memory.
