DDR5 shortage in 2026: what the sources report and what it does to a server quote

Server DDR5 supply in 2026 in the words of TrendForce, HPE and Dell

Server DDR5 memory is reported as undersupplied in late 2026. TrendForce wrote in its press release of 30 September 2026: "Server DRAM is therefore expected to remain undersupplied in 4Q26." HPE named DDR5 as "a great example" of its supply constraints on 2 September 2026, and Dell put DRAM first in its list on 1 September 2026. For a buyer the effect shows in the server quote: it carries a validity date, memory becomes a line of its own, and module sizes may be substituted.

"DDR5 shortage" is the phrase people search for, and no source read for this article uses it literally. TrendForce's word for the fourth quarter is "undersupplied". It uses "shortage" for the following year, in its release of 9 July 2026: "Meanwhile, a server DRAM shortage is already anticipated for 2027." SK hynix uses the word once, in a heading of its results release of 28 October 2026. This article keeps to the sources' wording and prints no prices, no percentages and no durations.

TrendForce press release of 30 September 2026; earnings-call transcripts of HPE and Dell; Micron's prepared remarks for its fiscal Q4 2026 call; SK hynix's 3Q26 results release; a memory maker's third-quarter 2026 results release. All read on 12 November 2026.

TrendForce's forecasts describe the contract market between memory makers and large buyers such as cloud providers and server makers. They are not statements about what an individual supplier can deliver.

The causes the sources name for the undersupply of server DRAM

The first cause is demand. TrendForce wrote on 3 July 2026 that general-purpose servers with x86 processors and RDIMM memory "remain the primary memory platform for Agentic AI workloads", and on 29 May 2026 that "rising CPU deployment is expected to expand server DRAM capacity requirements". An AI project therefore adds demand twice: for the memory on the accelerators and for the RDIMMs of the servers around them.

The second cause is high-bandwidth memory (HBM), the memory of data-centre accelerators. In the release of 29 May 2026 TrendForce wrote that "the growing wafer consumption associated with HBM production is compressing available capacity for conventional DRAM", and on 30 July 2026 that HBM manufacturing "requires significantly more wafer input than conventional DRAM". HPE described the consequence for an ordinary server on its call of 2 September 2026: "Now, in the traditional server, we use DRAM. We do not use HBMs." The server takes no HBM, yet its modules come from the same wafer capacity.

Third, the makers place the capacity they have on server products. TrendForce wrote on 25 August 2026 that demand from cloud providers "is prompting suppliers to prioritize limited capacity for server applications". Fourth, new capacity takes time. Micron said on 30 September 2026 that "cleanroom additions are required to augment node-transition supply growth and help narrow the gap", and added: "we do not have line of sight to when supply and demand will return to balance". SK hynix named the "long lead time required to build new production facilities". Fifth, limits also sit outside the wafer fab. TrendForce on 30 September 2026: "constraints in component supply, packaging and testing capacity, and front-end DRAM production flexibility continue to prevent supply from fully aligning with demand."

The end of DDR4 as a cause finds little support in these sources. Only HPE mentions the move from DDR4 to DDR5, as a technology shift, and it listed "DDR4 for the older generation" among the constrained parts as well. No source read here names a date on which DDR4 ends.

Long-term agreements and the direction of DDR5 server prices

Large buyers have secured their supply by contract. TrendForce reported on 9 July 2026 that several US cloud providers "have entered into multi-year long-term agreements (LTAs), which restrict suppliers from raising prices for these clients". The makers describe the same from their side. Micron, in its remarks of 30 September 2026, calls its contracts "multi-year take-or-pay agreements". SK hynix wrote on 28 October 2026 that it has signed long-term agreements with large customers, cloud providers among them, and describes "a mechanism that reflects market conditions within a price range agreed upon with customers". HPE said on 2 September 2026 that it has signed "multi-year LTA agreements, to lock a capacity" with some of its suppliers.

For everyone else TrendForce drew the conclusion in the same release: it expects that "the primary source of server DRAM price increases will shift toward customers without LTAs", and that "as the market remains undersupplied, memory suppliers may continue revising quotations upward throughout the remainder of this quarter". On the direction, its release of 30 September 2026 says that contract prices keep rising in the fourth quarter while "the pace of contract price increases is expected to moderate". Dell said on its call of 1 September 2026 that part of the higher cost of its server configurations is "driven by the price increases as our input costs continue to go up". We print no figures from these documents; they are forecasts for contract markets, and a quote states the price that applies to one order.

What this does to a server quote

No primary source read for this article says anything about how long a server quote is valid: not TrendForce, not the memory makers, and neither Dell nor HPE on their calls. The effects below are our reading of the mechanism the sources describe.

A quote carries a validity date. If memory quotations can be revised within a quarter, a supplier can hold a price for memory only as long as its own source holds it. The date on the quote says until when the prices in it apply, and after that date the quote is renewed.

Memory becomes a line of its own. A quote that lists the modules separately, with type, size and number, can be renewed line by line, and a change in the memory does not reopen the cards or the chassis. Our guide to comparing GPU server quotes line by line shows how to read such a document.

The module population is stated. Where a supplier may substitute module sizes, the quote should say which population it contains, because the population decides the speed and the room for later upgrades, as the next section shows.

We build AI servers to order with DDR5 ECC memory sized to the GPU pool; send us the cards and the amount of memory you plan and we reply within one business day with the configuration, the memory population and a written quote.

RDIMM module sizes that can be substituted and what the swap changes

The market itself is changing module sizes. TrendForce reported on 9 July 2026: "Since the first half of 2026, CSPs and OEMs have gradually adjusted their RDIMM configurations, with some systems shifting from 96 GB and 128 GB modules to 32 GB and 64 GB modules." On 30 September 2026 it added that memory suppliers "have increased the share of lower-capacity RDIMMs in their product mix". No source ranks module sizes by supply, and neither do we.

What a swap changes follows from the board. NVIDIA's configuration guide for NVIDIA-Certified Systems (30 September 2026) asks for memory spread evenly across all sockets and memory channels, so the total comes in steps of the channel count. The table uses the platforms from our guide on how much RAM a GPU server needs.

Dell PowerEdge XE7745 Technical Guide Rev. A09 (24 slots, one DIMM per channel, RDIMMs of 32 to 256 GB at 6400 MT/s); Dell XE7740 Technical Guide Rev. A04 and HPE DL380a Gen12 QuickSpecs for two Xeon 6700P (6400 MT/s at one DIMM per channel, 5200 MT/s at two), as cited in our guide on system memory, read on 5 November 2026. Totals by our arithmetic; the maker's list for the exact server decides which module sizes it accepts.

On a board with one slot per channel, a smaller module means a smaller total: 64 GB modules in place of 96 GB modules turn 2,304 GB into 1,536 GB. On a board with two slots per channel the same total can be kept with twice as many modules of half the size, at the lower speed the maker gives for two modules per channel, and with every slot occupied. An equivalent module is therefore one that the server maker lists for that board, of the same type, in a population that keeps every channel filled. A request can say so in one line: equivalent modules accepted, total at least a stated amount, all channels populated.

How much memory to order when DDR5 is constrained

The amount should follow the workload, in both directions. NVIDIA's guide for NVIDIA-Certified Systems asks for system memory of at least twice the total GPU memory, and NVIDIA's reference architectures set a floor of 128 GB per GPU. Our reference configurations follow the first rule: 384 GB for two RTX PRO 6000 Server Edition cards, 768 GB for four and 2.3 TB for eight H200 NVL.

Three workloads justify more: KV cache offloading for many returning users, fine-tuning with the optimiser state held in system memory, and a host that also runs a vector database or virtual machines. Our memory guide and our guide to CPU, RAM and storage for a 4-GPU server go through them. A server that runs one model with one engine on bare metal can sit at the 128 GB floor, with less headroom for a second model or a larger page cache.

Buying less now and adding later has a cost that depends on the board. With one slot per channel a later upgrade replaces every module; with two slots it adds a second module per channel at the lower speed. That belongs in the decision, and it is one reason to fix the amount before the quote. We size the memory of each server against these rules and the workload; tell us the models, the number of users and any fine-tuning plans and we reply with the amount and the population.

Enterprise SSDs in the same quote

The drives follow a similar pattern. Dell named NAND second after DRAM on 1 September 2026. TrendForce wrote on 30 September 2026 about enterprise SSDs that "most incremental supply has already been committed", and that their price growth is "expected to accelerate in 4Q26, making it the only memory product category to see a larger increase than in the previous quarter". On 3 July 2026 it linked the two markets: "shortages of internally sourced DRAM continue to constrain the supply of small-capacity, high-performance enterprise SSDs".

On the outlook the sources differ. TrendForce's release of 30 July 2026 projects that NAND flash will "enter a looser supply environment" when new production capacity comes online; Micron said on 30 September 2026 that it expects the NAND industry "to remain supply constrained". In a quote, the SSDs are best treated like the memory: a line of their own, with the validity date of the whole document.

What a buyer can do

• Fix the configuration early. Cards first, then the amount of memory from the rule and the workload, then the population. What else a request should contain is in our guide on what to specify when buying an AI server.
• Name acceptable equivalents. State the total, the type and the rule that every channel is populated, and allow equivalent module sizes within it.
• Decide within the validity of the quote. After its date the lines for memory and SSDs are quoted again. A plan for the year should rest on a current quote, as our article on the AI hardware budget for 2027 explains.

Memory is one of several steps that decide when a server arrives; the others are in our article on what decides GPU server lead times. How we answer a price request is described in our article on the NVIDIA GPU price list for Europe. Our written quote names the price, the lead time and the warranty terms, and quantities and lead time are confirmed in writing with the quote.

FAQ