A DRAM price jump of 80 to 90 percent in a single quarter would be extraordinary under any circumstance. In 2026, it is simply the cost of keeping up with artificial intelligence.
Memory pricing has entered a regime that supply chain professionals describe with words like "unprecedented" and "never seen" — not as hyperbole, but as a literal description of their experience. By early 2026, DRAM contract prices had climbed approximately 80 percent quarter-on-quarter, with NAND flash rising roughly 50 percent over the same span, according to industry tracker TrendForce. The more recent PC DRAM numbers are even more extreme: 105 to 110 percent quarter-on-quarter, outpacing server DRAM increases of 88 to 93 percent. The memory supercycle is not a forecast. It is here.
The cause is not complicated. AI infrastructure is extraordinarily memory-hungry. A typical 2026 AI node deploys between 192 and 288 gigabytes of HBM per system, alongside additional DDR5 and 20 to 30 terabytes of NVMe, according to EE Times reporting — pushing memory content per system into five-digit dollar territory. The OpenAI Stargate initiative alone reportedly consumes up to 40 percent of global DRAM output, requiring approximately 900,000 wafers per month. The data center has become the dominant DRAM consumer, reaching roughly 50 percent of global consumption in 2025, up from 32 percent five years prior, and projected to exceed 60 percent by 2030. This is not a cyclical surge. It is a structural reallocation of the memory supply chain toward AI workloads, and it is happening faster than fab capacity can respond.
The supply response has limits set by physics. Samsung Electronics expanded its 1c DRAM capacity to target 60,000 wafers per month specifically for HBM4 production by September 2025 — meaningful progress, but incremental against demand measured in hundreds of thousands of wafers. SK Hynix has sold out its entire DRAM, NAND, and HBM capacity through the end of 2026. Micron has completed agreements on price and volume for its entire calendar 2026 HBM supply under binding contracts. The memory industry is not holding back; it is simply producing everything it can, and it is not enough.
The big tech companies spending $650 billion on AI in 2026 — up about 80 percent from last year's record — have responded by locking up supply through those binding HBM agreements. They can absorb price increases because their willingness to pay is set by revenue from AI services, not by hardware margin targets. That leaves everyone else — the server OEMs, the PC makers, the smartphone manufacturers — to fight over what remains.
The PC segment is already showing the strain. HP reported that memory now accounts for roughly 35 percent of the cost of materials needed to build a laptop, up from about 15 to 18 percent just a quarter earlier. That is a doubling of memory's share of the bill of materials in three months. Dell Technologies COO Jeff Clarke described the pace of cost increases as unprecedented — not as a figure of speech. Lenovo Chief Financial Officer Winston Cheng called the cost surge unprecedented and disclosed that the company's memory inventories were approximately 50 percent above normal levels as it moved to build buffer stock ahead of further price moves.
The server market faces a more complicated calculus. AI servers are paying. The question is whether traditional enterprise buyers will accept the new memory economics or delay purchases. The smartphone market, which has historically absorbed component cost increases and passed them to consumers, is taking a different approach. IDC projects global smartphone shipments to decline 12.9 percent in 2026 — the largest-ever decline. When you cannot pass the cost on, you ship less.
The mobile segment is, in the current supply environment, acting as the release valve. Not through price negotiation, but through demand destruction.
The alternative to negotiation is engineering. Google Research published work in early 2026 on a technique called TurboQuant — a training-free compression algorithm that quantizes LLM key-value caches down to 3 bits without measurable loss in model accuracy. In internal benchmarks, the 3-bit version achieved roughly sixfold memory reduction. At 4-bit precision, the method delivered up to eightfold inference speedup on H100 GPUs versus unquantized 32-bit keys. If deployable at scale, it represents the most direct technical solution to the memory constraint: use less memory per inference, not more memory per system.
It is too early to call it a solve. Compression techniques have a ceiling. The demand trajectory — McKinsey projects total data-center capex reaching $6.7 trillion by 2030, with $5.2 trillion for AI-focused facilities — is not slowing. But TurboQuant illustrates where the pressure is pointing: the industry is not waiting for fab capacity to catch up. It is trying to reduce the memory intensity of the workloads that are driving the shortage.
The supercycle has a clear hierarchy of pain. The hyperscalers are insulated by binding supply agreements and revenue models that can absorb cost increases. The PC makers are absorbing a structural shock to their bill of materials. The smartphone manufacturers are cutting volumes rather than raising prices. The question is not whether the memory market will rebalance — it always does. The question is the sequencing: who absorb the cost, who delay purchases, and who simply stop?
Right now, the answer is: everyone except the companies that moved fastest and spent the most to lock in supply. The rest are waiting to see who flinches first.
Micron reported fiscal Q1 2026 revenue of $13.6 billion with a gross margin of 56.8 percent. SK Hynix is sold out through 2026. Samsung is adding capacity where it can. The fabs are running as fast as the physics allows. The demand is not waiting.
