The reason the RAM in your local LLM box costs three times what it did a year ago is not that anyone started buying more RAM. It is that three companies decided the same wafers are worth more as high-bandwidth memory than as the modules on a retail shelf. That decision is now visible in every price list, and it changes how a local build should be specced for the rest of 2026.
The arithmetic: allocation, not demand
Conventional DRAM and HBM draw on overlapping manufacturing capacity. Samsung, SK Hynix and Micron make more than 95% of the world’s DRAM, and each gigabyte of HBM consumes roughly three times the wafer capacity of a gigabyte of DDR5 — a ratio Micron’s own executives have stated publicly. A wafer converted to HBM does not just serve a different customer; it removes several gigabytes of DDR5 from the pool module assemblers depend on.
TrendForce’s numbers put the shift plainly. HBM wafer input among the top three suppliers is estimated at roughly 18% of total DRAM wafer input at the end of 2025, about 22% by the end of 2026, and about 30% by the end of 2027. Bit supply lags that — around 8%, 9% and 13% for the same periods — because stacked memory yields less usable capacity per wafer. The crowding-out effect is not a rumour; it is the model the analysts publish.
The clearest hard evidence is money paid in advance. Customer deposits rose 56% at both TSMC and Samsung in Q2 2026, and Micron’s rose roughly 23-fold the quarter after. Buyers do not prepay for capacity they expect to receive anyway. SK Hynix told investors in October that its HBM, DRAM and NAND capacity is essentially sold out for 2026 under non-cancellable contracts, and Micron retired its consumer-facing Crucial brand in February 2026 to concentrate on data-centre and HBM products. When a manufacturer exits the consumer shelf entirely, the price of what remains on it stops being a consumer price.
What the numbers look like on a shelf in October 2026
A 32 GB DDR5 kit that sold for $90–$100 in mid-2025 now runs $350–$450 at the low end, and the specific kits being tracked are well above that. A Kingston FURY Beast 32 GB DDR5 kit is listed at $656–$694 as of early October 2026 — about $21 per gigabyte — and a 64 GB kit of the same line is listed at $1,517.99. Per gigabyte is the honest way to read this: the same memory, priced three to four times higher.
Contract pricing moves first, retail follows. The benchmark contract price for a 16-gigabit DDR5 device rose to $48 in September from $45 in July and $46.50 in August. Samsung raised its own 32 GB DDR5 modules to $239 from $149 in September — a 60% increase from a manufacturer that cannot insulate itself despite its scale. TrendForce forecasts conventional DRAM contract prices up 10–15% quarter-over-quarter in Q4 2026 and NAND up 15–20%, with PC DRAM revised upward to 13–18% after suppliers’ catch-up hikes for server memory. The pace is moderating because most of the increase is already priced in. The supply stays short.
One detail worth knowing: HBM is no longer the more profitable product per wafer. TrendForce measured HBM wafer revenue falling below that of a DDR5 64 GB RDIMM in Q1 2026, which is why some buyers are evaluating eight-layer stacks instead of twelve-layer ones to contain costs. That does not help a consumer — allocation runs on margin arithmetic a shopper never sees.
DDR4: a dead end that still works right now
DDR4 is being formally retired. Micron issued end-of-life notices for DDR4 and LPDDR4 in June 2025 with final mainstream shipments wrapping in Q1 2026. Samsung stopped taking new DDR4 orders in mid-2025 and closed final shipments by December 2025. SK Hynix, the last high-volume producer, wound down through the first half of 2026. Everything on shelves today is existing inventory plus a trickle from smaller fabs.
That creates an odd situation for a local build. A used AM4 platform with a 64 GB DDR4 kit is still the cheapest way to get a large memory pool for offloading, and it is fast enough for inference — but it exists only while stock lasts. The same Kingston FURY Beast 64 GB DDR4 kit that averaged $119 through 2024 and bottomed at $104.99 in October 2024 is now listed near $1,500. Legacy memory is not cheap memory; it is scarce memory.
Practical consequence: if you already own a DDR4 board, keep the kit. If you are buying a platform now, DDR4 is a dead end with no second generation behind it, and DDR5 pricing is the price of a platform that still has a supply chain.
Why this hits local inference harder than it hits a normal PC
Local inference is memory-bound. Model weights have to sit somewhere, and for anything above a small model the bottleneck is capacity and bandwidth, not compute. That is why memory pricing is an inference-cost problem, not a shopping inconvenience.
The tier structure of a local build follows directly:
- VRAM is the tightest tier. HBM is allocated through 2026 and advanced packaging is the binding constraint — TSMC described its packaging capacity as limiting customers’ growth on its July call. Consumer cards inherit the shortage because the same wafer pool and the same packaging lines feed them.
- System RAM is the second tier, and it is where the price increase is most visible to an individual buyer. A 125B mixture-of-experts engine like Strata runs on a 64 GB DDR4 pool precisely because capacity, not speed, is what makes a large sparse model feasible on desktop hardware. The model works on a 12 GB card because most experts never load. The memory pool is the product.
- Storage is the third tier, and it is rising too. NAND contract prices are forecast up 15–20% in Q4 2026. Offloading to NVMe is a real technique, but the drive holding the weights is also getting more expensive.
At the top of the market the same allocation shows up as a spec sheet. A rack-scale system like AMD Helios carries 31 TB of HBM4 in one rack, and that capacity is what lets it hold a frontier model resident. The buyer of that rack and the buyer of a 32 GB kit compete for the same wafers. One of them is signing a multi-year non-cancellable contract.
A decision framework for the rest of 2026
If you already have the memory, do nothing. A 64 GB pool bought in 2024 is worth several times its purchase price. Selling it to fund a build is the worst trade available.
If you must buy, buy capacity before speed. A slower 64 GB kit beats a fast 32 GB kit for anything that must hold weights resident. Latency tiers matter less than the gigabytes you can actually address.
Quantize instead of buying. A 4-bit quant of a 27B model fits in a pool you already own. The cheapest memory is the memory you do not need to buy.
Do not wait for relief. New fabs are years away. Industry consensus is that meaningful supply relief for consumer DDR5 does not arrive before 2028; a leaked SK Hynix internal analysis puts PC DRAM supply behind demand until late 2028. Waiting for a price drop is a bet against a structural allocation, not a cycle.
Check the per-gigabyte price, not the kit price. A $600 kit at $19/GB beats a $450 kit at $28/GB. Per-gigabyte is the only comparison that survives a shortage.
Gear worth tagging in this market
Prices here move weekly; treat them as an early-October 2026 snapshot.
If you are building on a current AM4 or AM5 board, the Kingston FURY Beast 64 GB DDR4 kit is the capacity-first option — the same part that sold for around $119 through 2024 is listed near $1,500 now, which is the shortage stated as a price. It is desktop U-DIMM, XMP-ready, and the only tier here that is being retired.
For a platform you intend to keep, the G.SKILL Flare X5 32 GB DDR5 CL30 kit is listed at $609.99 — roughly $19 per gigabyte, which is the better end of the current market. It carries an AMD EXPO profile and is a matched kit, so do not mix it with existing modules.
Storage is the tier most people forget is also on allocation. The WD_BLACK SN850X 2 TB is listed at $159.99 against a $189.99 list price, and the Crucial P3 Plus 2 TB at $117.99 is the DRAM-less alternative for a weights cache where capacity matters more than random IOPS. Both are Gen4 M.2 2280.
What this means
The memory market stopped being a consumer market. Three suppliers allocate wafers to the highest-margin product, and the highest-margin product is stacked memory for accelerators that are already sold out through 2026. Every price a reader sees on a RAM kit is the residue of that allocation.
For anyone running models locally, the conclusion is defensive: keep the memory you have, buy capacity over speed when you must buy, quantize before you upgrade, and do not plan on a price drop before 2028. The hardware guides on this site assume a build cost that no longer holds — the best hardware for running local LLMs at each tier is the same silicon it was in September; the memory line on that bill is what changed. Spec a build on gigabytes you already own before specing it on anything you would have to buy now.
Pingback: NVIDIA DGX Spark 64GB: Half the Memory, Higher Price, and the Home Cluster That Actually Matters - GenX AI Tools