Memory chips were supposed to get cheaper. Instead, prices for DRAM have climbed relentlessly through 2026. The culprit sits in plain sight. Artificial intelligence demands vast quantities of specialized high-bandwidth memory. That shift starves conventional supply lines.
Producers now funnel wafers toward premium HBM stacks for AI accelerators. Every line converted means fewer standard DRAM chips for PCs, laptops and servers. Buyers outside the hyperscale club pay more. They wait longer. Some simply scale back.
The HBM Squeeze Reshapes Production Priorities
TrendForce data paints a stark picture. HBM will eat up 22% of total DRAM wafer input at the top three makers by the end of 2026. That share jumps to 30% a year later. (Yahoo Finance) Bit supply follows. Nine percent in 2026. Thirteen percent in 2027. Those wafers produce less usable output. Yields hover in the 25% to 35% range for advanced HBM4 variants versus over 80% for standard DRAM. The math doesn’t favor consumers.
Samsung, SK Hynix and Micron hold the keys. They command roughly 89% of the market. Samsung at 38%. SK Hynix 29%. Micron 22%. All three have locked in long-term contracts with hyperscalers. Nvidia alone soaks up huge volumes. OpenAI struck a deal with Samsung for hundreds of thousands of raw wafers a month. That removes supply from open markets entirely. (Himanshu Balani blog via recent analysis)
But wait. Conventional DRAM prices still rose 13% to 18% quarter-over-quarter in Q3 2026 according to the latest TrendForce report covered by Tom’s Hardware. NAND climbed 10% to 15%. Those gains come slower than the 50% to 60% spikes seen earlier in the year. Consumers finally hit a wall. They balk at sticker shock on new PCs and phones. Demand softens. Yet AI servers keep pulling forward. x86-based AI systems and inference clusters show no signs of slowing through 2027. (Tom’s Hardware)
Micron’s CEO Sanjay Mehrotra delivered the clearest warning. The HBM market will expand from $35 billion in 2025 to $100 billion by 2028. That milestone arrives two years ahead of earlier projections. “Demand will outstrip supply substantially for the foreseeable future,” he told analysts in late 2025. His company’s entire 2026 HBM output already sits under contract. No spare capacity exists. New fabs in Singapore and Taiwan won’t hit meaningful production until 2027 at the earliest. A major New York facility stretches to 2030. (IEEE Spectrum)
Intel CEO Lip-Bu Tan echoed the sentiment at a recent Cisco AI Summit. “There’s no relief until 2028.” He spoke from direct experience. Data center operators scramble to secure every available module. Spot prices for standard DRAM jumped 80% to 90% in late 2025. Contract prices followed with 90% to 95% gains in Q1 2026. The second quarter added another 58% to 63%. Momentum persists even if the rate of increase moderates. (IEEE Spectrum)
And the capital spending tells its own story. Industry outlays climb from $53.7 billion in 2025 to $61.3 billion this year. Most of that money flows into HBM production lines, advanced packaging and yield improvements rather than broad capacity expansion. New plants take 12 to 24 months to reach full output. Even then, much of the added wafers feed AI. UBS analysts expect supply and demand to approach balance only in the second quarter of 2028. Morgan Stanley projects memory could represent 40% of total AI infrastructure spending by 2030. The numbers compound.
PC builders feel the pinch hardest. DDR5 modules for consumer systems now carry what some call an “AI tax.” System prices rise. Shipments of notebooks and desktops weaken. Smartphone makers pass costs along where they can. But affordability limits have started to bite. Inventory for SSDs builds in some channels, moderating those prices faster than DRAM. Still, no meaningful relief arrives for memory-intensive builds anytime soon. Manufacturers prioritize higher-margin enterprise contracts. They have little incentive to flood the consumer channel.
Recent discussions on X highlight the tension. Korean-language analysis shared widely notes that 3Q contract prices are set to rise another 18% to 25%. Fourth quarter adds 8% to 10%. The cycle peaks sometime in the second half of 2028. New capacity from SK Hynix in early 2027, Micron mid-year and Samsung in July could finally tip the scales. Until then supply growth for standard DRAM bits hovers around 15% annually while underlying demand runs at 22%. The gap widens. HBM4 development accelerates the distortion. Base dies move to 12-nanometer processes. SK Hynix taps TSMC’s 3-nanometer technology for future variants. Those steps improve performance but further complicate production economics.
Hyperscalers reserve supply years in advance. They pay premiums gladly to stay ahead in the training race. Traditional buyers lack that leverage. Smaller tech firms face an “existential crisis” according to some analysts. They cannot secure enough memory at any price. OEMs such as Dell, HP and Acer hunt for alternative suppliers including Chinese players like CXMT. Results remain mixed. Quality and volume fall short of the big three.
So what breaks the pattern? Yield improvements on HBM stacks offer one path. Better coordination between memory designers and GPU architects helps squeeze more performance from each wafer. Process refinements accumulate over time. None of these deliver quick fixes. New fabs finally coming online in 2027 and 2028 represent the first real inflection. Even then, forecasts suggest supply will cover only 60% of demand at the end of 2027. Prices stay elevated. The industry has entered a structural upcycle unlike past commodity swings. AI intensity per server keeps rising. Memory now sits at the center of infrastructure budgets rather than the periphery.
Investors noticed. Memory stocks soared hundreds of percent in some cases. Yet volatility remains. Any slowdown in big tech capex could trigger a sharp correction later in the decade. For now the data points to sustained tightness. DRAM prices aren’t falling. They are finding a new, higher equilibrium built on insatiable compute demand. PC enthusiasts and enterprise IT departments alike will budget accordingly. The memory wall didn’t disappear. It simply moved and grew thicker.
WebProNews is an iEntry Publication