AI hardware & the memory market
AI hardware and the memory market
The price of a desktop memory kit is being set somewhere else entirely. AI accelerators, rack-scale systems and desktop AI boxes are consuming the memory industry's capacity, and consumer DDR4 and DDR5 are what gets squeezed out. This page traces that chain and shows the demand behind it. If what you actually want is whether to buy today, read should you buy RAM now instead.
DDR5 median (desktop)
$15.31/GB
171 kits tracked
DDR5 30-day
+3.7%
mean $/GB
DDR5 90-day
+21.6%
mean $/GB
DDR4 90-day
+5.7%
the outgoing generation, too
How an AI order reaches your checkout
There is no single villain here, just a supply chain with one narrow point in it.
- 01
AI accelerators need HBM
Training and inference silicon is bandwidth-starved, so it ships with high-bandwidth memory (DRAM dies stacked vertically and wired through the silicon) rather than the flat modules a desktop uses.
- 02
HBM is expensive in wafers
A stack packs several dies plus interposer and packaging steps, so each gigabyte of HBM consumes substantially more fab capacity and yield headroom than a gigabyte of commodity DDR5.
- 03
The same fabs make your DDR5
Only three suppliers (Samsung, SK hynix and Micron) make the bulk of the world’s DRAM. Capacity moved to HBM is capacity not making consumer memory, and new fabs take years.
- 04
Consumer supply tightens
Less DDR4 and DDR5 reaches retail against steady demand, so the price of an ordinary desktop kit rises for reasons that have nothing to do with desktops.
The demand side, measured
Every other figure on this page is RamRadar's own. These are not: they come from Epoch AI and describe what accelerator makers spend on memory, which is the demand this whole argument rests on.
Growth in quarterly HBM spend
6.6×
2024 Q1 to 2025 Q4
HBM share of accelerator cost
63.4%
as of 2025 Q4, up from 51.9% in 2024 Q1
What that means
As of 2025 Q4, memory is most of what an AI accelerator costs to build, so memory is what its makers compete hardest to secure.
What AI accelerators spend on memory
Industry-wide HBM spend per quarter, in USD billions. Ends 2025 Q4.
Source: Epoch AI (CC-BY). Modelled estimates, production-volume-weighted across the four largest AI-accelerator designers. Not transaction prices.
Read this chart against the price chart above rather than expecting one line to explain the other. They are different measures on different scales, and drawing them on shared axes would let the choice of scale manufacture whatever correlation the author wanted. Two charts, one reader, no thumb on the scale.
The hardware doing the buying
Four tiers, only one of which competes with you for the same physical parts.
Rack-scale AI systems
Grace Blackwell-class racks
The datacenter buildout that consumes HBM in volume. Each deployment is measured in thousands of accelerators, and it is this tier, not desktops, that sets memory-fab priorities.
HBM3E / HBM4
NVIDIA DGX Spark
GB10 Grace Blackwell desktop
The "AI box" category: a developer machine sized to run large models locally. Notable here because it ships a large unified memory pool of its own rather than sipping from the consumer DIMM market.
128 GB unified LPDDR5X
GeForce RTX 5090
Blackwell flagship
The consumer card most often bought for local AI work as much as for games. Its memory is soldered GDDR7, so it competes for graphics-DRAM capacity rather than for DIMMs directly.
32 GB GDDR7
Local-inference workstations
Large-capacity DDR5 builds
The one tier that competes with you directly: running big models on CPU or offloading layers to system RAM rewards 128 GB and up, which pulls high-capacity DDR5 kits off the same shelves.
DDR5 DIMMs
What RamRadar's data shows
As of October 6, 2026, desktop DDR5 averages $17.33/GB across 171 tracked kits, and DDR4 $7.54/GB across 94 . DDR5 has moved +21.6% over the past 90 days, DDR4 +5.7%. Both generations moving together is the tell: a demand shift confined to one product would not lift the other.
The useful consequence for a buyer is that market-wide advice stops working. When everything drifts upward, "wait for a sale" is not a plan. What still works is judging a specific kit against its own history, which is what the near all-time-low deals view is for.
Common questions
Why are RAM prices going up in 2026?
Because AI hardware is consuming the memory industry’s output. Accelerators use high-bandwidth memory, which takes far more fab capacity per gigabyte than ordinary DDR5, and the same handful of manufacturers make both. Capacity redirected to HBM is capacity not producing consumer DIMMs, so retail DDR4 and DDR5 prices rise even though desktop demand has not changed much.
Do gaming GPUs like the RTX 5090 make my RAM more expensive?
Only indirectly. A 5090’s 32 GB of GDDR7 is soldered to the card and competes for graphics-DRAM capacity, not for the DIMM slots in your motherboard. The larger effect runs through shared fab capacity and through people building high-capacity DDR5 machines to run models locally.
Will RAM prices come back down?
Not quickly. Adding DRAM capacity means building fabs, which takes years, and AI demand for HBM is contracted well ahead. The realistic planning assumption is elevated pricing persisting rather than a fast return to 2025 levels. That is why timing a purchase against price history matters more than usual right now.
Should I just wait to buy RAM?
Waiting is only free if prices fall. In a supply-constrained market the better approach is to buy when a specific kit is near its own all-time low rather than trying to call the market. RamRadar tracks every kit against its own history so you can see whether today’s price is genuinely good.
Is DDR4 safe from this?
No. DDR4 is affected more, not less. Manufacturers are winding DDR4 production down to free capacity for DDR5 and HBM while a large installed base still needs it, so it behaves like a scarce legacy part rather than a cheap outgoing one.
Buy against the data, not the mood
Near all-time-low deals
Kits within 10% of the lowest price RamRadar has recorded for them since tracking began.
RAM Price Index
The full daily $/GB series for DDR4 and DDR5, downloadable as CSV.
Should you buy now?
The buyer's version of this page: what the surge means for your purchase, and the best-value kits we track right now.
DDR4 end of life
Why the outgoing generation is getting more expensive, not less.
Some links below are affiliate links: RamRadar may earn a commission if you buy through them, at no extra cost to you. How we make money
Get price-drop alerts for DDR5 RAM
We track prices hourly. Join the list and we'll email you when kits on this page hit new lows, plus a weekly near-ATL shortlist.