DDR5 prices quintupled in a year and homelabs are footing the AI bill

August 2026. 500%. $3,399. The most boring component in a homelab just became the most expensive. A Tom’s Hardware study published in mid-August 2026 shows DDR5 prices climbed 500% in twelve months, and up to ten times the lowest prices ever tracked: a 128 GB DDR5 kit now sells for $3,399, while a kit that was $72 last year now runs $392.

The cause is not a classic chip shortage. It is AI: hyperscalers are absorbing nearly all global DRAM production to feed their training data centers, and end buyers — self-hosters included — are left with the scraps, at a premium.

The numbers that tell the surge

The Tom’s Hardware data, cross-checked with PCPartPicker, is unambiguous. A standard 64 GB (2×32 GB) DDR5-5600 kit that sat under $200 last summer now tops $1,100 — a five-fold multiplier on a line item that used to be predictable.

The contagion reaches the previous generation too. With DDR5 out of reach for most builders, demand spilled onto DDR4, whose kits are up 120% to 180% depending on the model. The “I’ll stay on DDR4 because it’s cheaper” reflex no longer works: DDR4 has become the pressure-release valve of a strained market.

The trend is not American. The German site ComputerBase tracks the same phenomenon in Europe and reports, the same week, an average rise of 345% compared to September 2025, with hard drives and SSDs up more than 125% over the same period.

Why DRAM became a precious metal

The study’s most striking detail is a comparison: mainstream DRAM is now worth more than half the price of gold per kilogram. That is not a columnist’s metaphor; it is a scarcity indicator — memory has become a strategic raw material, on par with a metal.

The mechanism is well known. The manufacturers — SK Hynix, Samsung, Micron — have redirected production capacity toward HBM, the stacked memory that accompanies AI accelerators. Every wafer devoted to HBM is a wafer taken away from consumer DDR5. The result is a market where classic DRAM supply contracts while AI demand explodes.

Hyperscalers are not waiting. According to the study, they have already locked in almost all global DRAM capacity for 2027, paying advance deposits to guarantee supply. Kwak Noh-jung, the SK Hynix CEO, warned that 2027 would be the worst year in the industry’s history for memory supply, with demand durably outstripping production.

What it changes for a homelab

For the self-hoster, the cost equation is inverted. A Proxmox server, a NAS, or a Jellyfin machine consumes memory for virtual machines, ZFS cache, and containers — a resource you once sized generously, almost without looking at the price. That era is over.

The practical consequence is direct. Sizing now happens at the minimum: you buy exactly the capacity you need, and favor configurations that leave room to grow. Two high-capacity sticks beat four low-capacity ones — not for technical reasons, but because they leave two slots free for a future addition, at a moment when every gigabyte costs more.

The same logic applies to platform choice. Older DDR4 motherboards gained a reprieve in value: their sticks are rising too, and moving to a new DDR5 platform means buying memory at the peak.

What to actually buy

The advice is not just “buy earlier.” Three practical rules change a homelab’s bill during a surge.

Size to the need, not to comfort. A Proxmox server running a dozen containers and two virtual machines fits comfortably in 32 to 64 GB. ZFS cache loves memory, but at $1,100 for 64 GB, every extra gigabyte of ARC costs a premium: a cache SSD — also up, but less — beats over-provisioning RAM.

Prefer two sticks, not four. Two high-capacity sticks leave two slots free for a future addition, and ease XMP/EXPO stability on two channels. On a four-slot board, filling all four at the market peak means locking yourself out of any extension without reselling everything.

Do not flee to DDR4 by reflex. DDR4 is up 120% to 180% through spillover: it is no longer the cheap refuge, and it freezes the platform on end-of-cycle CPUs. If the goal is to last three years, DDR5 remains the rational choice, at the current acknowledged price.

One more consideration for server builders: ECC memory — the kind a real NAS or Proxmox host wants — carries its own premium on top of the general surge, and its retail supply is even thinner. Budget for ECC up front, because sourcing it later, at the peak, is the most expensive way to build a server.

The trap of waiting

The natural reflex in the face of a surge is to wait for the drop. That is exactly the trap the study dismisses. For prices to fall, the AI market would need a major contraction — in other words, a brutal correction in global financial markets. Absent that scenario, the manufacturers themselves see no turnaround before 2027, at best.

In short: waiting is a bet on an AI crash. That is not a purchasing plan; it is a macroeconomic wager. The rational position is the opposite — if capacity is needed now, it gets paid for now, on a two-to-three-year horizon, without hoping to resell higher or rebuy lower later.

A shortage that is not a classic cycle

Memory has always been a cyclical market. In 2017-2018, a comparable surge already doubled prices, driven by the DDR4 transition and smartphone demand. The cycle then turned, and prices receded. Those betting on an identical scenario today are missing a difference of nature.

The 2017-2018 cycle was supply-driven — a technological transition bottleneck, absorbable once factories caught up. The 2026 surge is demand-driven: hyperscalers are buying DRAM for AI, and they are paying in advance by locking in 2027 capacity. A transition resolves itself; a structural demand exceeding supply does not, short of a turnaround in the AI market itself.

The ComputerBase signal completes the picture: SSDs and hard drives are up more than 125% over the same period, a sign that the tension is spreading through the entire storage chain. For a homelab, that means the extra cost is not limited to RAM: a server’s “storage + memory” budget is rising as a whole, and sizing must be done across both line items at once.

Verdict

If you are building or expanding a homelab now, buy the capacity you will need for two to three years, favoring two high-capacity sticks over four low-capacity ones — you keep headroom without rebuying at the peak. Do not count on DDR4 as a refuge: it has surged too, and it offers no lasting price advantage.

If your need is not urgent, wait, but with eyes open: the drop will only come from an AI contraction that no one — from manufacturers to hyperscalers — anticipates before 2027. The era of cheap, plentiful memory is over, at least for now — and the homelab, like the rest of the market, must relearn to size it as a scarce resource.

References

• Memory prices climb 500% in 12 months — Tom’s Hardware, August 2026
• Self-Host Weekly (21 August 2026) — selfh.st, accessed August 23, 2026
• ComputerBase — European memory price tracking, accessed August 23, 2026