AI's Next Crisis Isn't Power or Chips—It's a Memory Crunch

The AI sector’s breakneck expansion is about to collide with a physical reality its billions in spending can’t immediately fix. According to Demis Hassabis, the Nobel laureate leading Google DeepMind, the most critical shortage facing developers by next year won’t be processors or electricity. It will be high-bandwidth memory (HBM), the specialized chips essential for feeding data to AI systems.

Hassabis recently told industry attendees that HBM is set to become the scarcest resource in tech’s supply chain. This bottleneck threatens to slow AI progress just as demand for more powerful models accelerates. His warning carries authority; he oversees Google’s massive AI projects and sees the supply chain pressures firsthand.

The problem is one of manufacturing physics, not just finance. HBM is produced by only a handful of firms—SK Hynix, Samsung, and Micron—using intricate processes that stack memory chips vertically. Expanding this capacity requires years and billions. SK Hynix, the leading supplier, is already sold out for 2025, with customers now vying for 2026’s output.

Modern AI models, with trillions of parameters, need to move colossal data volumes at incredible speeds. Without enough HBM, even the most advanced processors stall. Each new generation of AI hardware demands more of this memory, and aggregate needs across thousands of chips are staggering.

The strain introduces geopolitical risk, as production is concentrated in South Korea and the U.S. Any disruption could ripple through global AI projects. This comes as tech giants pour historic sums into data centers. Without reliable HBM supply, those investments could face delays, sparking fierce competition for available stock.

If Hassabis is correct, the industry may be forced to pivot from simply scaling up hardware to pursuing smarter efficiencies in algorithms and software. While some see this as a potential catalyst for innovation, others fear it could hinder Western labs. Chinese competitors, backed by state investment, are racing to build their own memory supply chains, though they currently lag behind.

Memory makers are investing heavily to expand, but new facilities won’t significantly boost supply until 2027 or 2028. For an industry accustomed to relentless growth, the next two years could present an unfamiliar challenge: a hard, physical limit.