Memory and Power Constraints
The AI build‑out is starting to hit physical limits beyond GPUs — memory and electricity are now the bottlenecks. Big tech firms are reportedly seeking multi‑year DRAM deals as shortages tighten memory supply, and analysts note hyperscaler spending is pushing power needs into the spotlight with projects that require massive electricity commitments. (en.bloomingbit.io) (techi.com)
AI companies can still buy graphics processors, but memory chips and electricity are starting to decide how fast new data centers get built. (trendforce.com) A data center needs two things besides processors: memory that keeps data close to the chip, like a workbench beside a machine, and power that keeps thousands of servers running at once. The International Energy Agency said servers account for about 60% of electricity use in modern data centers, while storage and networking add about 5% each. (iea.org) That memory market is tightening. TrendForce reported on April 9 that Samsung Electronics and SK hynix are shifting big customers from one-year deals to contracts lasting three to five years, with reported talks involving Microsoft, Google and Advanced Micro Devices. (trendforce.com) The pressure is strongest in high-bandwidth memory, or HBM, the ultra-fast memory stacked next to artificial intelligence accelerators. Micron said in March that it had completed price and volume agreements for its entire 2026 HBM supply, including HBM4. (investors.micron.com) SK hynix has been signaling the same squeeze. Its January 2026 market outlook said demand for HBM3E and HBM4 would drive an “AI memory supercycle,” and its investor site lists fiscal 2025 fourth-quarter results released on January 29, 2026, after the company had already said earlier production was effectively booked out by artificial intelligence demand. (news.skhynix.com) (skhynix.com) Power is tightening at the same time. The International Energy Agency estimates data centers used about 415 terawatt-hours of electricity in 2024, or 1.5% of global consumption, and projects demand will rise to about 945 terawatt-hours by 2030. (iea.org 1) (iea.org 2) The biggest artificial intelligence sites now look more like utility projects than warehouse leases. The International Energy Agency said a typical AI-focused data center uses as much electricity as 100,000 households, and the largest facilities under construction today will consume 20 times that amount. (iea.org) That is why technology companies are signing unusually large power deals. OpenAI said in July 2025 that its Oracle partnership added 4.5 gigawatts of United States capacity, bringing Stargate projects under development to more than 5 gigawatts and more than 2 million chips. (openai.com) Microsoft made the same turn in 2024, but through generation rather than new buildings. Constellation said on September 20, 2024 that Microsoft signed a 20-year agreement tied to restarting Three Mile Island Unit 1, adding about 835 megawatts of carbon-free power with a target online date in 2028. (constellationenergy.com) The result is that the artificial intelligence build-out is no longer constrained by chip design alone. The next limit is whether cloud companies can lock up enough memory wafers and enough megawatts years before the servers arrive. (trendforce.com) (iea.org)
