Last updated: June 2026
Key Takeaways
- Consumer RAM prices have climbed to roughly three to six times their 2024–2025 lows, and analysts tracking Samsung, SK Hynix, and Micron do not expect meaningful relief before late 2027.
- The shortage has now disrupted Apple's Mac lineup: high-memory Mac mini and Mac Studio configurations were pulled from sale, and the M5-generation Mac Studio missed WWDC 2026 entirely.
- Local AI builds are still viable in 2026 if you buy strategically. Machines with memory already installed, used 24GB GPUs, and right-sizing to the models you actually run are the workarounds that hold up.
If you recently priced a memory upgrade and assumed the listing was an error, it was not. The kit that cost $90 in mid-2025 now sells for several hundred dollars, and the high-capacity configurations that local AI builders care about most have been hit hardest. This guide explains what happened, gives you an honest answer on when prices may come back down, and lays out the build strategies that still make sense while the shortage lasts.
What Happened to RAM Prices: The Short Answer
Consumer DRAM prices rose roughly 300 to 600 percent from their 2024–2025 lows, and they remain elevated in mid-2026. On June 3, Tom's Hardware's daily RAM price tracker recorded the cheapest in-stock 32GB DDR5 kit in the United States at $374.97. One year earlier, that capacity routinely sold for $80 to $120. High-capacity 64GB DDR5 kits that sat near $150 to $200 for most of 2025 now commonly list at $600 or more.
The primary driver is the AI infrastructure buildout. Samsung, SK Hynix, and Micron, the three companies that produce nearly all of the world's DRAM, have redirected wafer capacity toward high-bandwidth memory (HBM) for AI accelerators. TrendForce estimates that AI workloads will absorb about 20 percent of global DRAM wafer capacity in 2026, and the math is unforgiving: one gigabyte of HBM consumes roughly the wafer area of four gigabytes of standard DRAM. Every HBM gigabyte sold to a datacenter is four consumer gigabytes that never get manufactured.
AI demand is not the only factor. Manufacturers were already retiring the older production nodes used for DDR4, which sent DDR4 prices climbing even faster than DDR5 in some configurations. Proposed tariffs on imported components add a further upside risk for US buyers. The clearest signal of where manufacturer priorities now sit came in February 2026, when Micron wound down its consumer-facing Crucial brand to concentrate on enterprise AI customers. The consumer memory market did not just get more expensive; it lost a first-party supplier outright.
For readers who want the one-paragraph version: DRAM is the working memory in your PC, laptop, phone, router, and NAS. HBM is a specialized, stacked form of DRAM used in AI accelerators, and it is far more profitable per wafer. Datacenter buyers sign long-term contracts at premium prices, so fabs serve them first. What reaches retail shelves is the remainder, and in 2026 the remainder is thin.
The Newest Casualty: Apple's Mac Lineup
The shortage stopped being an abstract supply-chain story for Mac users this spring. As MacRumors reported, Apple removed several configurations from its online store, including Mac mini models with 32GB and 64GB of RAM and the M3 Ultra Mac Studio with 256GB of unified memory. Tim Cook has acknowledged that the Mac mini and Mac Studio may be hard to get for months.
The WWDC 2026 keynote on June 8 came and went without new Mac hardware. Apple focused the event on macOS 27 and its updated Siri, and as of publication it has not announced M5-generation Mac Studio or Mac mini models. Current reporting points to a launch later in 2026, possibly October, with analysts warning that memory costs could push Apple to raise prices or trim entry configurations when the new machines do arrive.
This matters for local AI builders specifically because high-unified-memory Macs were the simplest single-box route to running large models. A 256GB M3 Ultra Mac Studio can hold 70B-class quantized models, and larger, entirely in memory. That door is narrowing at exactly the moment open-weight releases like DeepSeek V4 are giving people new reasons to want the capacity.
Why Local AI Builders Get Hit Hardest
For local inference, memory is the binding constraint. Model weights must fit entirely in RAM or VRAM, or the system falls back to disk swapping and token generation slows to the point of being unusable. Our local AI hardware guide covers this in depth; the table below shows what the memory math looks like against the models people are actually running in mid-2026.
| Model Class | Mid-2026 Examples | Approx. Weights (4-bit) | Realistic Minimum |
|---|---|---|---|
| 7–8B | Gemma-class 8B, ZAYA1-8B | ~5 GB | 16 GB system RAM |
| 13–14B | Mid-size open models | ~8–9 GB | 16–32 GB system RAM |
| 30–32B | 30B-class coding and reasoning models | ~18–20 GB | 32 GB RAM or a 24 GB GPU |
| 70B | 70B-class quantized models | ~40–43 GB | 64 GB RAM or 48 GB VRAM |
| 200B+ MoE | DeepSeek V4-Flash (284B) | ~160 GB | 192 GB+ unified memory or multi-GPU |
Figures are weights only. Your operating system, the inference runtime, and the context window all consume additional memory, so treat these as floors, not targets.
The pressure runs in two directions at once. Frontier open-weight releases keep raising the ceiling: DeepSeek V4-Flash is a 284B-parameter mixture-of-experts model, and even with only 13B parameters active per token, the full weights still have to live somewhere. Meanwhile, efficiency-focused models keep lowering the floor. Zyphra's ZAYA1-8B, released in May under Apache 2.0, activates only about 760 million parameters per token and delivers reasoning performance that punches far above its memory footprint. The practical takeaway: a 16GB to 32GB machine is far from obsolete in 2026, because the small end of the model spectrum keeps improving.
One more wrinkle: this is not only a DRAM problem. Graphics memory (GDDR) and flash storage (NAND) are riding the same demand wave, which means GPU and SSD prices are creeping upward too. That shifts the value calculation toward used and renewed graphics cards, which we cover in the build strategies below.
When Will RAM Prices Actually Go Down?
Not meaningfully in 2026. That is the consensus across the research firms tracking the big three manufacturers. The most likely shape of this year is stabilization at elevated levels: the steep weekly increases of late 2025 have largely paused, but pausing is not the same as falling. Analyst forecasts compiled in early June describe prices as more likely to climb than to drop through the rest of the year, and several projections place the earliest credible normalization window in late 2027, with some stretching into 2028. Winbond has indicated that production capacity is effectively booked through 2027, and Samsung's reported contract increases of up to 60 percent have already worked their way into retail.
You will occasionally see headlines about prices dipping. Treat single-digit corrections on individual kits as plateau noise, not a trend reversal. Three things could genuinely change the timeline: new fabrication capacity coming online (most of which is already earmarked for HBM rather than consumer DRAM), a slowdown in AI datacenter spending, or tariff decisions, which cut in the wrong direction for US buyers.
The practical planning assumption: budget for current pricing through the end of 2026, and treat seasonal sale events like Prime Day and Black Friday as the only realistic discount windows. Even then, expect discounts measured against today's prices, not 2025's.
How to Build for Local AI During the Shortage
High memory prices change the order of operations, not the destination. Here is what holds up in mid-2026, depending on your timeline.
If You Need a Machine Now
- Buy memory inside the product, not as bare kits. Mini PCs and prebuilts that ship with RAM installed are frequently priced on component supply contracts signed before the squeeze, which means the memory inside them can cost less than the same capacity bought separately. A machine that includes 64GB is carrying components that would cost hundreds of dollars at retail today. Our mini PC guide for local AI covers the full lineup; two standouts for this strategy are the Beelink SER9 Pro+ with 32GB of LPDDR5X and the MINISFORUM AI X1 Pro 370 with 64GB.
- Shift the bottleneck from DRAM to VRAM. A used or renewed RTX 3090 carries 24GB of GDDR6X, enough to run 30B-class models at 4-bit entirely on the card. Pair it with a modest amount of system RAM and you have sidestepped the worst of the DDR5 market. Renewed 3090s remain the best value per gigabyte of VRAM for local inference.
- Right-size to the models you actually run. If your workload is an 8B to 14B model powering a local agent stack, 32GB of RAM is comfortable and 128GB is panic spending. Put the savings into fast NVMe storage for your model library instead, where capacity per dollar is still reasonable. One caveat: NAND rides the same demand wave as DRAM, so if you know you need the storage, buying sooner is the safer bet than waiting.
If You Can Wait
Waiting is a legitimate strategy as long as you wait with accurate expectations. Watch the quarterly contract price reporting from firms like TrendForce, and use daily retail trackers and price-history tools to set alerts rather than checking listings manually. Recalibrate what counts as a deal: in this market, 10 to 15 percent under the prevailing average is a genuine discount. A return to 2025 pricing is not coming this year, and anchoring to it will keep you waiting indefinitely.
Because prices move weekly, verify the current market before acting on any figure in any article, including this one. The live listings are the ground truth:
Check Current Price: DDR5 64GB Kit (2x32GB)
Check Current Price: DDR4 32GB Kit (2x16GB)
What to Avoid
- Inflated "flash deals." Retailers are running promotions on kits priced two to three times what the same product cost a year ago. A discount badge on a repriced product is not a deal. Check the price history before checkout.
- DDR4 as an escape hatch. DDR4 32GB kits that sold for $60 to $90 in October 2025 more than doubled within months and have kept climbing as production winds down. In some configurations DDR4 now costs as much per gigabyte as DDR5, on a platform with no future. If you are building new, build DDR5.
- Panic-buying capacity you will not use. The table above is the sizing guide. Buying 128GB to future-proof against models you do not run locks today's worst prices into hardware that may outlive the shortage.
- Gray-market sellers. Shortages attract counterfeit and remarked modules. Stick to established brands sold and shipped by authorized retailers, and test new memory on arrival.
Frequently Asked Questions
Why are RAM prices so high in 2026?
AI datacenter demand is the largest driver. Memory manufacturers have shifted wafer capacity to high-bandwidth memory for AI accelerators, which consumes roughly four times the wafer area per gigabyte of standard DRAM. Combined with the wind-down of older DDR4 production lines and tariff uncertainty, consumer supply is structurally constrained.
When will RAM prices go back down?
Analyst consensus points to no meaningful relief in 2026. The earliest credible window for normalization is late 2027, when new fabrication capacity begins ramping, and some projections extend into 2028. Even then, most new capacity is earmarked for AI memory, so a full return to 2025 pricing is unlikely.
Is it cheaper to buy a prebuilt or mini PC right now?
Often, yes. Manufacturers buy memory on long-term contracts, so machines with RAM already installed frequently undercut the cost of buying the same components separately at current retail prices. Compare the full system price against the parts list before assuming a DIY build saves money.
Are SSD prices going up too?
Yes. NAND flash shares the same manufacturing capacity pressures as DRAM, and suppliers have warned of shortages and rising prices. The increases have so far been milder than DRAM's, which is why buying needed storage sooner rather than later is the safer move.
Should I buy DDR4 instead of DDR5 to save money?
No. DDR4 prices spiked alongside DDR5 as manufacturers retired older production nodes, and the per-gigabyte gap has nearly closed in many configurations. DDR4 also ties you to end-of-life platforms. For any new build in 2026, DDR5 is the correct choice despite the prices.
Does the memory shortage affect routers and modems?
Yes. Networking equipment uses the same DRAM and NAND components, and vendors have flagged rising component costs that flow through to retail prices on routers, gateways, and mesh systems. It is one more reason owning your own equipment beats paying an ISP rental fee on hardware whose replacement cost keeps climbing.
