RAM shortage: why servers cost more and how long it will last
In the first quarter of 2026, a 64 GB DDR5 server module was already earning memory makers more per wafer than the memory that sits next to AI GPUs. And prices are still climbing.
If you need to refresh servers this year, memory will cost you far more than it did twelve months ago. DRAM contract prices rose 90–95% quarter on quarter in the first half of 2026, and Dell and Lenovo have already passed the increase on. Forecasts point to further rises, gradually smaller, until the second half of 2027. The room left to save is in how the memory is configured.
Until recently memory was the boring part of a server spec: you picked a round number, doubled it to be safe, and the final price barely moved. In 2026 that habit gets expensive.
Why there is a RAM shortage: AI memory uses three times the wafer
AI accelerators are paired with HBM (high bandwidth memory, stacked memory with very high throughput), which is made from the same wafers as the DRAM in an ordinary server but needs roughly three times the wafer area for the same capacity. Micron expects that ratio to get worse with HBM4. On top of that, Samsung, SK hynix and Micron, which make almost all of the world's DRAM, are handing HBM a growing share of their wafers. TrendForce estimates put it at 18% at the end of 2025, 22% at the end of 2026 and 30% by the end of 2027.
Every GB of HBM takes the place of about three GB of DDR5, which is why AI eats so many wafers. Conventional memory has become so expensive that a 64 GB server module now earns more per wafer than HBM itself. Hyperscale cloud providers add to the pressure by buying ahead for the shortage they expect in 2027, and a CPU shortage held up server assembly in the second quarter.
How much RAM prices have risen, and for how long
DRAM contract prices, the prices equipment makers pay, rose 90–95% quarter on quarter during the first half of 2026. For the full year, Gartner expects DRAM to end up 271% more expensive than in 2025, close to four times the price. Server vendors have passed it on: Dell raised server prices on 10 December 2025, and Lenovo told its channel partners in February that PC and server prices would go up from March.
For server memory, TrendForce forecasts another 13–18% in the third quarter of 2026 and quarterly increases through the second half of 2027, each one smaller than the last. Supply of RDIMM modules, the kind servers use, will grow only 15–20% a year, behind server CPU shipments, so the forecast has no price declines before the end of 2027. Anyone refreshing servers this year or next has no reason to expect cheaper memory by waiting.
What you can still control is the configuration, and if lead times are the problem, we have servers in stock with delivery in under 30 days.
How to configure server memory without overpaying
Hypervisors report two numbers for each virtual machine, the memory it has been allocated and the memory it is actually using, and the second is usually well below the first because VMs get generous headroom that nobody reviews afterwards. Measure usage at peak hours over a few weeks and keep headroom for what really matters, which in a cluster means being able to absorb the VMs of a failed node, and you will need fewer modules before vendor choice even comes up.
How the memory is laid out matters as much as how much you buy. Each processor accesses memory through channels, and vendors recommend balanced configurations with every channel populated by identical modules. Empty channels cut memory bandwidth for exactly the workloads that feel it most, such as databases, analytics and dense virtualisation. The capacity of a well-built server is therefore set by the platform: channels multiplied by module size.
This leads to the less obvious saving, which is the socket count. A two-socket server needs the same memory on each processor, so it doubles the channels to fill. With CPUs now reaching 128 performance cores on Intel Xeon 6 and 192 on AMD EPYC 9005, many workloads that used to need two sockets fit in one.
A virtualisation node needs around 750 GB. With two Xeon 6 6700P processors there are 16 channels to fill, which means 16 modules of 48 GB (768 GB). With a single AMD EPYC 9005 there are 12 channels and 12 modules of 64 GB: the same capacity with four fewer modules and one fewer CPU, as long as one socket has enough cores for the workload.
It also pays to decide up front how the server will grow. On platforms that support one module per channel, such as the Xeon 6 6900P series, there are no free slots, so expanding means replacing modules rather than adding them, at whatever memory costs when the time comes.
Waiting for AMD EPYC Venice only makes sense for 2027 projects
Venice, the sixth generation of EPYC, is built on a 2 nm process, reaches 256 cores per socket, and AMD puts its availability in the fourth quarter of 2026. For anyone buying in the middle of a shortage, the relevant detail is that it moves from 12 to 16 memory channels per socket and supports DDR5 at 8,000 MT/s. A properly configured Venice server therefore carries more modules, and faster ones, than a current-generation machine.
If you need capacity this year because hardware is reaching end of support, you are running out of headroom or it is time to move off VMware, EPYC 9005 and Xeon 6 are mature, and Dell, Lenovo and Supermicro already sell servers on both. Venice is a better fit for projects starting in 2027 that will push memory bandwidth hard, such as AI, HPC or in-memory databases.
SSDs are going up too, just less
NAND flash contract prices rose 55–60% in the first half, according to the same Gartner figures. With every TB of NVMe now costing more, it makes sense to keep it for databases and virtual machines and move backups and archives to high-capacity hard drives or object storage.
Servers in stock for this year
At SIXE we have servers and storage in stock with delivery in under 30 days. We usually recommend Supermicro, because it lets us pick the processor, memory and drives one by one and match the channel count to the capacity you need, and we also work with Dell PowerEdge and Lenovo ThinkSystem if they are already your standard. You can configure a server or browse servers by vendor.
Refreshing servers this year?
Send us the memory usage of your current hosts (a vCenter or Proxmox export is enough) and we will tell you how many modules you need and what size, on Supermicro or on the vendor you already use.
Blog!
Our mission
SIXE ensures that your infrastructures based on IBM®, Red Hat®, Canonical®, and SUSE® technologies remain stable, secure, and optimized for as long as you need them.
At SIXE, we are committed to sustainability. That’s why we help you design, implement, and maintain durable and efficient technological solutions.
