Some forms of RAM tripled in price over the course of 2025. Datacenters are on track to consume 70% of global memory capacity. And in a lab in eastern India, a plant pathologist just had to cut the number of crops she studies because cloud computing bills got too expensive.
A new Nature report published today puts a name to it: "RAMmageddon." The term refers to the global memory chip shortage driven almost entirely by AI infrastructure buildout, and the collateral damage it's inflicting on scientific research worldwide.
How AI Ate the Memory Supply
The root cause is straightforward. Companies like Samsung, SK Hynix, and Micron have shifted manufacturing capacity away from standard DDR5 memory toward HBM (high-bandwidth memory), the specialized chips that power AI training hardware like NVIDIA's GPUs. Each gigabyte of HBM requires roughly three times the silicon wafer space of conventional DDR5. More wafers going to AI means fewer chips for everyone else.
The numbers tell the story clearly. AI workloads will consume about 20% of total global DRAM production in 2026, according to TrendForce. Conventional DRAM contract prices jumped 55-60% in Q1 2026 alone. Micron announced in December 2025 that it was exiting the consumer memory market entirely to focus on AI datacenter customers. Samsung has been filling only about 70% of its DRAM order backlogs.
This isn't a temporary blip. Analysts expect the shortage to persist well into 2027.
The Labs Getting Squeezed
The practical fallout for science is concrete. Pravallika Sree Rayanoothala, a plant pathologist at Centurion University of Technology and Management in Paralakhemundi, India, told Nature that her team reduced the number of crops in their disease-forecasting project. They now break datasets into smaller chunks and model them separately to avoid paying for expensive cloud memory.
Researchers in parts of Africa face an even starker reality. Some travel to wealthier universities abroad, stay for six months to run their analyses on better-funded hardware, and leave with a PDF of results.
These aren't edge cases. Any researcher running genomics pipelines, climate models, or large-scale data analysis needs significant memory. When RAM prices triple and cloud costs spike in lockstep, the labs with the smallest budgets get priced out first.
A Silver Lining, Barely
The shortage is pushing some researchers toward more memory-efficient algorithms and smaller models that can run on constrained hardware. That's genuinely useful work. But framing a supply crisis as an innovation driver feels hollow when the core problem is that three chip manufacturers decided AI datacenter margins were more attractive than serving a broad customer base.
Samsung plans to boost HBM production to 250,000 wafers per month by late 2026, up 47% from current levels. SK Hynix is accelerating its new fab by four months. More supply should eventually ease prices, but "eventually" doesn't help a researcher whose grant cycle runs out this year.
For anyone using AI tools daily, this is worth understanding: the infrastructure that powers ChatGPT, Claude, and every other large model isn't free. The silicon has to come from somewhere, and right now it's coming at the expense of the broader computing market. The AI industry's appetite for specialized memory is real, it's growing, and the people paying the steepest price aren't the ones building the models.