The AI build-out stopped being an abstraction for consumers in 2026, when it started showing up on the price tag of ordinary laptops and phones. This page tracks the transmission mechanism from hyperscaler memory demand to retail device pricing. GPU-side supply is covered separately in the GDDR memory crisis report.
The Numbers
| Metric | Value | Source basis |
|---|---|---|
| DRAM spot price change, year to mid-2026 | Up roughly 700% | Spot market reporting |
| Share of global memory output absorbed by AI data centres | ~70% (estimated) | Industry analysis |
| Memory as share of laptop bill of materials | ~35%, up from 15-18% a quarter earlier | HP |
| Forecast memory price rise by end 2026 | ~130% | Gartner |
| Forecast PC price rise vs 2025 | ~17% | Gartner |
| Forecast smartphone price rise vs 2025 | ~13% | Gartner |
| Forecast device price rise by end 2026 | 10-20% | IDC |
| Conventional DRAM contract prices, Q3 2026 | Up 13-18% quarter over quarter | Contract market reporting |
| Smartphone materials cost impact | Up 15% or more in coming quarters | Counterpoint Research |
Why It Happened
High-bandwidth memory carries much better margins than commodity DRAM and consumes substantially more wafer area per usable part. Faced with hyperscaler HBM demand, the three large memory makers, Samsung, SK Hynix, and Micron, moved cleanroom capacity and capital expenditure toward enterprise parts. Consumer DRAM and NAND were left competing for what remained. This is a deliberate allocation decision rather than a manufacturing failure, which is why it has proved durable.
Where It Landed
Apple raised Mac prices on June 25, 2026, citing higher memory and storage component costs. The MacBook Air moved from $1,099 to $1,299 and the MacBook Pro from $1,699 to $1,999, increases of roughly 18 percent on both. Apple also pushed the M5 Ultra Mac Studio refresh to the fourth quarter and thinned out high-memory desktop configurations, which is covered in the Mac Studio shortage report.
For anyone buying hardware to run models locally, this is the worst possible cost structure. Local inference is a memory-capacity problem before it is a compute problem, so the exact component that determines which models you can run is the one whose price is rising fastest.
When It Recovers
Not soon. Samsung and SK Hynix have both warned the shortage could persist into 2027 and beyond, and SK Hynix's chief executive has said 2027 is expected to be the worst supply year in the industry's history, with demand outstripping capacity beyond 2030. The one moderating signal is demand-side: price increases through Q3 2026 are cooling as consumer buyers reach the limit of what they will pay, which caps how far retail prices can follow component costs.
What It Means for Local AI
Three consequences. First, the cost per gigabyte of resident model capacity is rising for the first time in the modern era, which inverts the assumption that local inference gets cheaper every year. Second, it raises the relative value of sparse mixture-of-experts models, where a large total parameter count still has to fit in memory but active parameters stay small, and of aggressive quantisation formats. Third, it pushes marginal buyers back toward cloud APIs, partially offsetting the local-inference shift that cheap capable open-weight models were driving.
Brand Visibility Implications
Hardware economics set the split between observable cloud inference and unobservable local inference. A memory shortage that makes local hardware more expensive keeps more brand-relevant queries on monitored endpoints for longer. That is a reprieve for measurement, not a reversal, and it lasts exactly as long as the shortage does.
Methodology
Vendor specifications come from NVIDIA, Apple, and model-card publications. Throughput figures aggregate community benchmark reporting from the llama.cpp discussions, the MLX repository, and published independent test suites. Single-stream decode unless stated otherwise. Ranges rather than point values are used wherever independent runs disagree, which is most of the time: quantisation format, prompt length, thermal state, and runtime version each move these numbers by more than the differences being measured. Treat every figure as an order-of-magnitude guide, not a specification. Updated quarterly.
How Presenc AI Helps
Presenc AI tracks brand visibility across both cloud and locally deployed models, so the measurement picture stays complete regardless of which way hardware economics push the deployment mix.