A 4% supply-demand gap in DRAM chips has emerged globally as AI infrastructure orders exceed production capacity. Memory prices are climbing toward parabolic levels as hyperscalers from North America to Asia compete for limited chip supplies.
The AMD-Meta partnership for 6GW of GPU capacity shows the scale mismatch. AI training clusters require massive memory configurations, but global DRAM production cannot expand fast enough. New fabrication plants cost $15 billion or more and take 18 months minimum to reach production, guaranteeing capacity arrives long after demand spikes.
Semiconductor indices trade near record highs as investors worldwide bet on an AI-driven supercycle. Israel's Camtek Ltd. projects double-digit revenue growth for 2026 based on its order backlog, expecting Q1 2026 revenues around $120 million with acceleration in the second half. The company's equipment sales signal sustained chipmaker expansion across manufacturing hubs in Taiwan, South Korea, and the United States.
DRAM manufacturing's cyclical nature amplifies the crisis. Chipmakers hesitate to invest in new fabs during downturns, leaving them cash-constrained when boom times return. This boom-bust pattern means production capacity consistently lags behind AI infrastructure buildouts across continents.
GPU shortages compound the memory bottleneck globally. Training advanced AI models requires balanced ratios of processing power to memory bandwidth. Shortages in either component create idle capacity in the other, reducing overall training throughput for AI labs from Silicon Valley to Shenzhen.
The supply constraints pose direct risks to AI development timelines worldwide. Labs planning large-scale training runs face delays acquiring hardware. Smaller AI companies from Europe to Southeast Asia struggle to compete for limited chip allocations against hyperscalers with guaranteed supply contracts.
Industry analysts warn the supply-demand imbalance could persist through 2027. Semiconductor manufacturers are greenlighting new fab construction across Taiwan, South Korea, Japan, and the United States, but the 18-month lead time means relief won't materialize until late 2026 at earliest. Until then, hardware bottlenecks will constrain AI scaling plans across the global industry.