The AI boom has created a memory problem that is now spreading beyond data centres. Prices for conventional DRAM and NAND, used in smartphones, graphic cards, and consumer electronics, have surged as memory makers shift capacity to bolster AI infrastructure.
Chip production has decisively turned towards AI. About 70% of the memory produced worldwide will be consumed by AI data centres, per a WSJ report. That leaves little interest and incentive to produce RAMs.
Market research firm TrendForce expects conventional DRAM contract prices to rise 58–63% quarter-on-quarter in the second quarter of 2026. It also expects NAND prices to rise 70–75%.
The crunch isn’t unknown to the AI ecosystem either. According to Bloomberg, NVIDIA has told major customers that prices for some AI server systems could rise by more than 15% from early 2027. The increase will affect systems using NVIDIA’s Grace Blackwell and Vera Rubin platforms.
Micron also expects the shortage to continue through 2027, with improvement only gradually emerging in 2028.
That creates a wider problem for the technology industry. AI companies are competing for memory as much as PCs, smartphones, and other electronics.
“Memory prices have now increased by more than fourfold and are approaching 5x their historical level,” Prachir Singh, Senior Research at Counterpoint Research, tells AIM.
Consumer electronics brands are now acting to shore up their bottom lines. Amazon has raised prices across Echo, Kindle, Fire TV and eero devices by as much as 60%, blaming higher memory and storage costs.
Meanwhile, after five quarters of growth, global PC shipments fell 4% year-on-year in Q2 2026, according to Counterpoint Research.
AI is Eating the Memory Supply
The root of the problem is high-bandwidth memory (HBM). AI accelerators need large amounts of HBM to move data quickly between memory and compute. According to Minjae Kang, Memory Analyst at SemiAnalysis, the shortage is both a supply deficit and a result of capacity shifting towards HBM.
“There is a genuine deficit underneath DRAM, but the magnitude of the recent price spike cannot be explained by the supply-demand gap alone,” Kang notes.
HBM today consumes roughly 2.5–3x as many wafers as conventional DRAM per bit, according to SemiAnalysis. Hyperscaler long-term agreements and pre-buying by non-AI players have further amplified the price surge.
Counterpoint sees the current crunch as partly cyclical, but says AI has changed the underlying allocation of semiconductor capacity.
“We do not regard this as another short-term DRAM upcycle, although prices will eventually return to normal levels. The allocation of semiconductor investment has changed with consumer electronics now having to share capacity with an AI infrastructure market which has a much higher willingness to pay,” David Naranjo, Associate Director at Counterpoint Research, explains.
Market intelligence firm IDC sees a strategic reallocation of wafer capacity towards AI memory. Every wafer used for HBM reduces capacity for conventional memory in consumer devices.
“As a result, OEMs are primarily raising retail prices and reducing memory configurations, including introducing lower configurations such as 4GB+64GB, to maintain competitive price points,” Singh says.
Higher-margin premium devices give manufacturers more room to absorb higher costs, while mass-market models remain more exposed to weaker purchasing power. That pressure is also changing what manufacturers choose to sell. Smartphone launches above ₹45,000 grew around 65% YoY in the first half of 2026, according to Counterpoint.
PC Gets Caught in the Middle
The problem does not stop at AI servers. PC makers now face higher costs for both DRAM and storage. Omdia expects global PC shipments to decline 12% in 2026 because of memory and storage shortages and higher component costs.
Counterpoint estimates that DRAM and SSDs, which once accounted for about 15% of a typical notebook’s bill of materials, now account for more than 25%. A typical notebook could now carry an additional $75–125 in memory and storage costs, Naranjo says.
AI PCs add another complication. Counterpoint estimates that AI-advanced PCs will account for more than 50% of global PC shipments in 2026. Mainstream laptops have increasingly standardised around 16GB of RAM, while more powerful AI PCs are moving towards 32GB for local models and agentic workloads.
“AI PCs should constitute a significant additional source of demand for PC DRAM, even though they are not the main reason for the present memory shortage,” Naranjo says.
SemiAnalysis agrees that AI PCs are not driving the current shortage, but expects on-device AI to become a larger source of memory demand once supply loosens.
Naranjo adds that for Copilot+ PCs, Microsoft recommends a minimum of 16GB of RAM and a 256GB SSD. He expects price-sensitive consumers to delay upgrades, buy older models, or turn to refurbished devices.
The Squeeze Reaches NVIDIA
Counterpoint also expects AI infrastructure to put pressure on advanced packaging, networking, optical components, and substrates. SemiAnalysis also sees substrates and foundry capacity emerging as potential bottlenecks as AI accelerators become more complex.
NVIDIA’s reported server price increase shows that the memory crunch has reached the AI infrastructure market itself.
The reported increase comes as memory becomes a much larger part of AI infrastructure costs. SemiAnalysis estimates that HBM, DRAM, and NAND together account for around 40% of all-in AI cluster capital expenditure.
Kang says a 40% increase in memory prices alone could raise total cluster costs by about 16%. Custom silicon also does not remove the need for memory. The 15%+ system increase remains rational even after NVIDIA reduces memory content in some newer systems.
NVIDIA CEO Jensen Huang has already warned that supply constraints will persist.
“The whole industry supply chain, everything from wafers to packaging to silicon photonics...is in short supply because the demand is so high. It is going to persist for several years,” he said in an appearance on the Dwarkesh Podcast.
Huang also views scarcity differently. “I think the fact that everything is scarce is fantastic for us,” he said at the Morgan Stanley Technology Conference, arguing that supply shortages force clients to buy from trusted market leaders.
SemiAnalysis points to long-term supply agreements as evidence that major AI customers are securing capacity well in advance.
“We think the memory makers hold strong pricing power over Big Tech and the AI labs,” Kang says.
Why Falling Chip Stocks Matter
The strangest part of the cycle is that memory stocks are falling even as memory makers project revenue uptick. Samsung shares dropped more than 8% on August 24 after investors reacted negatively to its shareholder-return plan. The fall helped trigger a wider semiconductor sell-off in South Korea.
NVIDIA, Broadcom, AMD and Intel fell roughly 3%, Micron dropped more than 6%, SK Hynix fell 5% and SanDisk nearly 7%.
Counterpoint expects memory prices to remain elevated into 2027, with DRAM likely to stay tighter than NAND as AI servers continue taking capacity. “The biggest impact would probably be a structurally longer replacement cycle,” Naranjo adds.
This may change consumer behaviour, retaining PCs and smartphones for longer, while entry-level devices lose ground.
The AI boom started as a race for computing power and is now becoming a race for memory too. SemiAnalysis also cautions that higher HBM stack counts are not necessarily the only way forward, as the industry debates whether adding more memory capacity delivers enough value for its cost.
Your reaction
Discussion
What's your take on this story?
No comments yet — be the first to share your view.