Particle.news
Download on the App Store

AMD’s Lisa Su Projects AI Will Need 10 Yottaflops Within Five Years

The forecast underscores looming power and infrastructure limits.

Overview

  • Speaking at CES 2026, Su defined a yottaflop as one septillion operations per second and said reaching 10 yottaflops would be about 10,000 times 2022’s global compute.
  • She cited estimates that global AI compute rose from roughly one zettaflop in 2022 to more than 100 zettaflops by 2025.
  • Business Insider noted that 10 yottaflops would be about 5.6 million times the performance of the DOE’s El Capitan supercomputer.
  • Reporters highlighted that power-grid capacity, energy build-out, cooling, and data-center infrastructure could be major constraints on scaling to this level.
  • AMD used the keynote to unveil Instinct MI455X GPUs, EPYC Venice CPUs, and Helios rack-scale systems as it positions to serve rising AI infrastructure demand.