The AI infrastructure paradox: over-ordering and under-building

What happened
Memory used in AI infrastructure is in very short supply, causing prices to nearly double in early 2026 as manufacturers focus on server memory for AI applications. At the same time, the number of announced data center projects is far greater than what can realistically be built, with grid connection requests in Britain tripling in under a year due to data center proposals. Major DRAM producers like Samsung, SK Hynix, and Micron are directing more production to high-bandwidth memory for AI, leaving less available for other uses. New memory fabrication plants take years to build and require manufacturers to be confident that current demand will last.
Key details
- Conventional DRAM contract prices rose about 93% to 98% in Q1 2026.
- SK Hynix warned that high-bandwidth memory shortages may last until 2030.
- Samsung, SK Hynix, and Micron control around 90% of the DRAM market.
- Grid connection applications for data centers in Britain increased from 41GW to 125GW in less than a year.
Why it matters
Cloud providers, data center developers, and smaller memory buyers face higher costs and tougher competition for limited memory supplies. Companies planning new data centers must navigate a market where many announced projects may never be built, and only those with secured power, financing, and equipment are likely to proceed. Memory manufacturers and infrastructure planners must make long-term investment decisions based on commitments rather than speculative demand, affecting how quickly new capacity can come online and who gets access to it.
