Updated

Updated · MIT Technology Review · Oct 19

Computing Industry Spends $700 Billion on Data Centers as 4 Key Scaling Laws Stall

Updated

Updated · MIT Technology Review · Oct 19

Computing Industry Spends $700 Billion on Data Centers as 4 Key Scaling Laws Stall

1 articles · Updated · MIT Technology Review · Oct 19

$700 billion in 2026 infrastructure spending by six hyperscalers shows computing progress is now being bought largely through bigger data centers, not cheaper transistor shrinkage.
Moore's Law has slowed, Dennard scaling ended around 2004, and Koomey, Kryder and Keck trends are flattening, leaving AI training compute to double about every six months mainly through system-level scale-up.
Current hardware still sits well above physical limits: chips use several hundred thousand times Landauer's minimum energy per switch, optical links deliver about one-millionth of the Holevo limit, and memory density remains nearly 30 orders below Bekenstein bounds.
Those gaps are steering research toward reversible logic, quantum computing, DNA storage and optical interconnects, with the article arguing the next leap will come from building systems up from physics-level limits rather than extending today's transistor roadmap.

Sources

Moore's Law is dying, but physics offers a way out. Will atom-scale engineering arrive before hyperscalers exhaust our power grids?

As AI pushes data centers to consume gigawatts, could synthetic DNA and reversible logic save computing from its own energy crisis?

If the human brain computes at twenty watts, why are tech giants building gigawatt facilities instead of mastering bio-hybrid memory?

1 total