The 2026 DRAM squeeze: what it means for dedicated server buyers

Memory has become one of the most volatile lines in a server’s bill of materials. TrendForce’s estimates of quarterly DRAM contract-price increases climbed from 8–13% in September 2025 to 90–95% for the first quarter of 2026, driven by AI servers and a shift of capacity toward high-bandwidth memory. For anyone renting dedicated servers, that changes which configurations are good value and how much RAM is worth paying for.

The numbers: a timeline

TrendForce, a market research firm that tracks memory contract prices, publishes quarterly estimates. The sequence below shows how quickly the outlook moved. Every figure is a quarter-over-quarter change in contract prices.

These increases compound: each quarter’s rise applies to a base that had already jumped. TrendForce expects server DRAM contract prices to keep rising every quarter through the second half of 2027, at a more moderate pace.

Takeaway: the Q3 2026 figure marks a slower rate of increase, not a decline. TrendForce’s early estimate is that RDIMM bit supply will grow only 15–20% in 2027, well behind projected server CPU shipments.

Why it happened: AI, HBM and a capacity shift

Three forces reinforced each other.

• Capacity moved to HBM and server DRAM. In September 2025, TrendForce reported that the three major suppliers were allocating advanced process capacity primarily to high-end server DRAM and HBM, crowding out PC, mobile and consumer memory. As of the second quarter of 2025, HBM3e still sold at more than four times the price of DDR5.
• Cloud buyers locked in supply. U.S. cloud providers pulled orders forward from late 2025 and negotiated long-term agreements, leaving less for everyone else. By March 2026, TrendForce named high-capacity RDIMMs, the modules inside high-memory servers, as their primary procurement target.
• DDR5 margins closed in on HBM. As DDR5 prices rose, TrendForce projected its profitability would overtake HBM3e’s from the first quarter of 2026, giving suppliers a reason to allocate more capacity to server DDR5 rather than to consumer products.

The consumer market felt it quickly. In December 2025, Micron announced it would exit its Crucial consumer business to improve supply for larger, strategic customers in faster-growing segments, citing AI-driven growth in data-center demand.

A second bottleneck is hiding part of the demand. In July 2026, TrendForce noted that server CPU shortages had slowed system assembly, so DRAM inventories were building up at U.S. cloud providers. As CPU supply improves from the second half of 2026 through 2027, server production should ramp up, and TrendForce already anticipates a server DRAM shortage in 2027.

What it does to dedicated-server prices

Memory cost scales with capacity, so the squeeze is not spread evenly. If the per-gigabyte price of server DRAM doubles, the memory bill of a 32 GB machine doubles, and so does that of a 768 GB machine, but the second adds 24 times as many dollars. High-memory configurations, and the high-capacity RDIMMs they use, absorb most of the increase. Storage is not immune either: TrendForce expected NAND flash contract prices to rise 55–60% in Q1 2026 and 70–75% in Q2.

Hosts can pass that on through higher list prices for new orders, fewer promotions or longer lead times on large-memory builds. Hetzner, for example, adjusted its server prices on 15 June 2026, citing “the massive increase in procurement costs”; the change applied to new orders and rescales, while existing contracts kept their terms. Questions worth asking any provider:

• Does a price change apply to servers you already rent, or only to new orders?
• What would a RAM upgrade cost later, compared with ordering the capacity now?
• How many DIMMs does the configuration use? On 12-channel EPYC platforms, the same capacity on fewer, larger modules means fewer populated channels and less memory bandwidth.

Right-sizing memory by workload

The cheapest gigabyte is the one you don’t rent. Size from measurements rather than habit. On Linux, MemAvailable in /proc/meminfo estimates how much memory is available for new applications without swapping, and /proc/pressure/memory shows whether tasks are stalling while they wait for memory. Watch both at peak load. Free memory on a busy server is usually low because the kernel uses spare RAM as page cache and releases it on demand, which is why MemAvailable, not MemFree, is the number to track. Documented starting points:

PostgreSQL also relies on the operating system’s cache, which is why its recommended share is smaller than MySQL’s. For virtualization hosts, add up the memory actually assigned to guests plus the host’s own needs instead of planning on overcommit: ballooning and page sharing help at the margin, but a host that swaps slows every guest at once.

Practical ways to spend less on memory

• Compare price per gigabyte, not only per core. Two servers at similar prices can differ several-fold in RAM, and in this market RAM is where much of the money goes.
• Let fast storage take some of the caching. On ZFS, an L2ARC cache device on NVMe sits between the disks and the in-memory ARC; the Proxmox documentation notes that it can help when the ARC must be reduced because of memory constraints.
• Size for the coming year, not the next five. With monthly billing, moving to a larger server later is possible. A migration has its own cost, so weigh it against paying now for memory you may never use.
• Don’t plan on waiting it out. TrendForce expects server DRAM contract prices to keep rising into the second half of 2027, and that from Q3 2026 most of the increase will come from buyers without long-term agreements and from supply sold outside them.

ECC or non-ECC

DDR5 blurs a distinction that used to be simple. DDR5 chips include on-die ECC: they store 8 extra bits for every 128 bits of data and correct single-bit errors inside the memory array. That does not protect data on the channel between the module and the CPU. Server ECC, known as side-band ECC, adds bits that the memory controller checks end to end, so single-bit errors are corrected anywhere between controller and memory, and the system can log them to flag a failing module.

At PowerDedi, the EPYC and Xeon configurations use DDR5 ECC memory. The Ryzen configurations use non-ECC DDR5: AMD’s Ryzen 9000 processors support ECC only where the motherboard does, and our Ryzen platforms ship without it.

ECC is worth insisting on for databases, storage servers, virtualization hosts running other people’s workloads and blockchain nodes; Anza suggests it for Solana validators. Non-ECC is a reasonable trade for stateless web tiers, build and CI machines and game servers, where a rare crash is recoverable and the data of record lives elsewhere.

Each PowerDedi configuration shows a “market” price next to ours: the median public list price of the same CPU, RAM and storage at major EU and US hosts with their standard 1 Gbps port, excluding setup fees and VAT. We re-check that comparison periodically, most recently in September 2026, and our own prices depend on the same component costs as everyone else’s. Current figures are on the pricing page.

Sources

• DRAM prices to continue rising in 4Q25 — TrendForce, 24 September 2025
• Tight DRAM supply to boost DDR5 contract prices — TrendForce, 29 October 2025
• Memory makers prioritize server applications, driving price increases in 1Q26 — TrendForce, 5 January 2026
• Memory price outlook for 1Q26 sharply upgraded — TrendForce, 2 February 2026
• Price rally drives 4Q25 DRAM revenue up 29.4% — TrendForce, 26 February 2026
• AI server demand to drive memory contract price increases in 2Q26 — TrendForce, 31 March 2026
• Server DRAM contract prices expected to rise 13–18% QoQ in 3Q26 — TrendForce, 9 July 2026
• Micron stops selling memory to consumers as demand spikes from AI chips — CNBC, 3 December 2025
• Standardization and price adjustment of our server products — Hetzner, 15 June 2026
• Resource consumption — PostgreSQL documentation
• Buffer pool and enabling automatic configuration for a dedicated server — MySQL 8.4 Reference Manual
• The /proc filesystem and Pressure Stall Information — Linux kernel documentation
• ZFS on Linux — Proxmox VE wiki
• Agave validator requirements — Anza
• Error correction code (ECC) in DDR memories — Synopsys
• AMD Ryzen 9 9950X specifications — AMD