RAM is the single input that decides which local AI models you can run, and it has never been more expensive. Conventional DRAM contract prices jumped 93% to 98% in Q1 2026 alone (TrendForce, 2026), and a 32GB DDR5-6000 kit that sold for about $80 in mid-2025 reached about $432 in early 2026 (Newegg Insider, 2026). Memory makers say relief is years out. For local AI builders the practical answer is to size RAM to the models you will actually run, not to max out, and this guide gives you the numbers to do it.
Why are RAM prices exploding?
DRAM contract price change, quarter over quarter (TrendForce, 2026)
AI datacenters are eating the supply. Samsung, SK Hynix, and Micron have shifted wafer capacity toward high-bandwidth memory for AI accelerators, which starves the consumer DRAM lines. This is a demand reallocation, not a factory disaster, which is why it is lasting so long.
The product-level figures in a SigmaIntel dataset were quoted by TweakTown, but the archived article is currently unreachable and we could not verify the underlying raw data. We therefore omit its exact module prices and percentage changes rather than repeat figures without a checkable primary source.
The one piece of good news: the rate of increase is slowing. TrendForce's Q2 2026 forecast was 58% to 63% growth, down from Q1's near-doubling, and its Q3 projection for server DRAM is 13% to 18% (TrendForce, 2026). Prices are still rising every quarter. They are just rising less steeply.
When will RAM prices come back down?
Not soon, according to the people who make it. Micron CEO Sanjay Mehrotra told investors in June 2026: "We expect tight conditions to persist beyond calendar 2027" (TechSpot, 2026). SK Hynix's CEO went further in a July Bloomberg interview, calling 2027 the industry's worst supply year ever, with demand outstripping supply beyond 2030, as reported by Reuters (2026). The same wire story cites UBS analysts projecting undersupply until at least Q2 2028.
Plan on elevated prices through at least 2027. A "wait for the dip" strategy has no dip on the calendar.
Why RAM is literally your model budget
For local LLMs, memory capacity is the hard gate. Every model needs its weights in memory, plus KV cache for context. The ladder is direct:
| Memory | Model class | What that means |
|---|---|---|
| 8GB | 7-8B | Llama 3.1 8B class |
| 16GB | 9-14B | Qwen3 14B class |
| 24-32GB | 20-32B | GPT-OSS 20B, Qwen3.6-27B |
| 48-64GB | 30-35B comfortable | Qwen3.6 35B MoE with context room |
| 128GB+ | 70-120B class | GPT-OSS 120B |
Get told when a better model fits your MacBook Air M5 16 GB
One email when a new open-weight model is a better fit for your machine, plus the Thursday weekly.
For: MacBook Air M5 16 GB
You'll also get the weekly ModelFit email. By signing up you agree to our Privacy Policy. Unsubscribe from either email anytime, in one click.
Full sizing logic is in our RAM guide, and the calculator shows exactly which models each tier unlocks. The point here: when a 32GB kit costs $432 instead of $80, each rung of that ladder now has a real dollar cost attached. Moving from 16GB to 32GB of DDR5 is no longer a casual upgrade. It is the price of a mid-range GPU.
What should PC builders do?
Buy the tier your models need, and stop there. Speculative "I might want 128GB someday" purchases are how the shortage taxes you hardest. Concretely:
- Running 7-14B models? 16GB to 32GB covers it. This is still the affordable zone, even at 2026 prices.
- Targeting 27-32B models? You need 48GB-class capacity, and at current DDR5 prices that step costs real money on the desktop too. Price the RAM before the GPU.
- Already own DDR4? Keep the platform alive. DDR4 is rising too, but an upgrade within an existing board still beats a full DDR5 rebuild.
- GPU route: VRAM is caught in the same memory squeeze, with reports of planned GPU price increases tied to GDDR supply, though no vendor has confirmed specifics. A used 24GB RTX 3090 remains the classic value play for 20B-class models; our GPU pages rank what runs on each card.
Does this change the Mac vs PC math?
Yes, in Apple's favor, which is strange to write. Apple doubled its RAM upgrade pricing in 2026: a 16GB step that cost $200 now costs $400 (Digital Trends, 2026). Digital Trends pegs the open-market equivalent of that 16GB at about $185, so Apple still charges a premium, but the gap has narrowed dramatically from the days when the same market RAM cost $50.
Meanwhile a Mac's unified memory is all model budget: a 48GB MacBook Pro gives you a ~35GB working budget with 307GB/s of bandwidth, no separate VRAM constraint. We break down the buy decision for the M5 Pro tiers in the $400 question and the timing angle in Mac Mini M4 vs M5: buy or wait. If you were ever going to tolerate Apple's RAM pricing, the year the whole market caught up to it is the year.
FAQ
Why are RAM prices so high in 2026?
Memory makers reallocated production capacity toward high-bandwidth memory for AI datacenters, shrinking consumer DRAM supply. TrendForce measured conventional DRAM contract prices up 93% to 98% in Q1 2026, with a further 58% to 63% forecast for Q2. It is a demand shift, not a temporary disruption.
Will RAM prices go down in 2027?
The manufacturers say no. Micron's CEO expects tight conditions "beyond calendar 2027," and SK Hynix's CEO called 2027 the worst supply year ever in a July 2026 Bloomberg interview, with UBS projecting undersupply until at least Q2 2028. Budget as if current prices are the new normal.
How much RAM do I need for local AI?
16GB runs 9-14B models, 24-32GB unlocks the 20-32B class including Qwen3.6-27B, and 48GB+ runs 35B-class models comfortably. Use the ModelFit calculator or run npx @wecko-ai/modelfit to get the exact pick for a given memory size before you pay 2026 prices for capacity you will not use.
Is a Mac better value for local AI now?
The gap has narrowed. Apple's $400 per 16GB upgrade pricing was always a heavy markup, but with open-market DDR5 around $185 for the same capacity and unified memory doubling as VRAM, a 48GB Mac is now a defensible local-AI buy against a DDR5 rebuild plus GPU. The M5 Pro RAM comparison has the full math.
Sources
- TrendForce, DRAM contract price update (June 2026): https://www.trendforce.com/presscenter/news/20260601-13070.html
- TrendForce, Q3 server DRAM forecast (July 2026): https://www.trendforce.com/presscenter/news/20260709-13140.html
- TweakTown (archived), SigmaIntel Q2 2026 DRAM data: https://web.archive.org/web/20260731091020/https://www.tweaktown.com/news/112358/dram-prices-surged-by-up-to-89-percent-in-q2-2026-destroying-the-consumer-segment/index.html
- Newegg Insider, DDR5 in 2026: https://www.newegg.com/insider/ddr5-memory-in-2026-whats-happening-to-prices-supply-and-speed-tiers/
- TechSpot, Micron CEO on shortages: https://www.techspot.com/news/112894-micron-ceo-expects-memory-shortages-stretch-beyond-2027.html
- Reuters via US News, SK Hynix CEO interview: https://money.usnews.com/investing/news/articles/2026-07-10/sk-hynix-ceo-sees-worst-ever-memory-supply-shortage-in-2027-says-demand-to-outstrip-supply-beyond-2030
- Digital Trends, Apple RAM upgrade pricing: https://www.digitaltrends.com/computing/apples-historically-high-tax-for-ram-upgrades-on-macs-has-now-become-absurd/
Don't miss the next one for your MacBook Air M5 16 GB
This article covers one moment. Get one email when a newer open-weight model fits your machine better, plus the Thursday weekly with what changed.
For: MacBook Air M5 16 GB
You'll also get the weekly ModelFit email. By signing up you agree to our Privacy Policy. Unsubscribe from either email anytime, in one click.
Match this model to a machine that can run it: by RAM tier for Apple Silicon, or by VRAM for an NVIDIA GPU.
See how this changes your recommendation
Run the wizardHave questions? Reach out on X/Twitter