TrendForce forecasts DDR2 — a legacy memory chip last common in PCs in the early 2000s — contract prices up 55–60% in Q2 2026 and another 35–40% in Q3 as one supplier (Winbond) exits and another (ESMT) absorbs the gap at wafer foundry PSMC.
AI training workloads did not ask for DDR2 memory. The chips in question have not powered a mainstream PC in well over a decade. Yet contract prices for that legacy DRAM are forecast to climb 55–60% in Q2 2026, with another 35–40% on the table for Q3, according to TrendForce figures reported by The Register. The buyers absorbing that increase are not data-center operators chasing the latest accelerators. They are the makers of factory controllers, medical monitors, point-of-sale terminals, and other embedded devices whose bill of materials has quietly relied on a stable supply of "retro" DRAM.
The mechanism is a wafer-allocation cascade, and the trigger sits at the very top of the memory stack. AI accelerators need high-bandwidth memory (HBM), and HBM production now consumes mainstream DRAM fab capacity that would otherwise be feeding the DDR4 and DDR5 lines used in servers and PCs. As The Register notes, buyers have turned to legacy DDR2 and DDR3 components to escape a market where mainstream DDR4 and DDR5 prices have already pushed PC costs up by double digits.
But the cascade does not stop at the laptop's DIMM slot. It runs all the way through the silicon supply chain into chips that have no AI use case at all.
The two suppliers that still carry meaningful DDR2 production are reshaping their books. Winbond is winding down DDR2 output and reallocating capacity toward DDR3, DDR4, and LPDDR4, the current-generation standards that command higher margins. ESMT, the other named DDR2 supplier, plans to do the opposite: maximize DDR2 output inside its existing allocation at foundry PSMC to absorb the gap left by Winbond's retreat, per the same Register summary of TrendForce data.
That decision was not made by anyone thinking about industrial PLCs, the programmable controllers that run factory lines. It was made by people reading the same HBM-capex spreadsheets as the hyperscalers and concluding that a wafer of legacy DRAM is more profitable routed to a medical-device customer than to a product line that can move to DDR3 or DDR4. The vendor incentive is clean: exit a sunsetting node, redeploy capital to nodes where the AI cycle is paying premium prices. The cost shows up somewhere else.
The redesigns TrendForce describes sit inside that "somewhere else." Some hardware makers are swapping DDR4 designs for DDR3. A smaller set is going further, replacing DDR3 with DDR2 to capture whichever DDR2 supply remains. Modern consumer CPUs generally lack DDR2 support, so those redesigns are landing where DDR2 has always lived: in embedded controllers, industrial automation gear, medical instrumentation, networking appliances, and other long-life-cycle products that are rarely in the technology press.
This is the part the "retro RAM" framing gets wrong. The story is not nostalgia for an old standard. It is a single foundry-level allocation decision, silicon wafers that could have made DDR2 instead being routed to HBM and mainstream DDR, rippling down to the bill of materials of a blood-gas analyzer or a vending machine controller whose manufacturer has never been part of an AI supply-chain conversation.
The longer-arc capacity story offers some relief, but not on a timeline that matters for buyers absorbing these contracts now. SK hynix has said it will double silicon-wafer output capacity over the next five years, per the same TrendForce figures summarized by The Register. Micron is expecting meaningful new output from its Virginia fabrication plant in 2027 and 2028. Neither will add a single DDR2 wafer. Both will eventually relieve pressure on DDR4 and DDR5, which in turn takes the buyer pressure off the legacy tail.
Until then, the next trigger to watch is whether ESMT's plan to maximize DDR2 at PSMC actually absorbs the Winbond gap, or whether Winbond's exit compresses DDR2 supply fast enough that even the embedded and industrial redesigns cannot keep up. The Register's own skepticism about DDR2 reaching consumer PCs is the strongest tell: the only rational buyers absorbing a 90%+ cumulative contract-price rise across two quarters are the ones whose products cannot move on.