At CES 2026 in Las Vegas, AMD CEO Lisa Su issued a stark warning about the future of artificial intelligence, asserting that the industry will require more than 10 yottaflops of computing power to keep pace over the next five years. The term yottaflop, a measure indicating one septillion calculations per second, represents a fundamental leap beyond current computing benchmarks, stressing the urgency for recalibration in the technology sector.
Su’s address underscored the rapid evolution of AI, highlighting that traditional metrics of computing power are becoming increasingly irrelevant. “How many of you know what a yottaflop is?” she queried, before explaining that it equates to a one followed by 24 zeros. Her projection indicates that achieving 10 yottaflops would require a staggering 10,000 times more compute than the industry had in 2022.
In 2022, global AI computing capacity was estimated at approximately one zettaflop, or 1021 operations per second. By Su’s estimation, this figure surged to over 100 zettaflops by 2025. The leap from zettaflops to yottaflops is not linear; rather, it compresses decades of progress into a mere handful of years. “There’s just never, ever been anything like this in the history of computing,” Su remarked, emphasizing the extraordinary pace of change.
For perspective, Su compared her projections against the current computational heavyweight, the El Capitan supercomputer, operated by the U.S. Department of Energy. The El Capitan, which currently leads global rankings, would need to be multiplied by approximately 5.6 million times to reach the 10 yottaflops mark. Even the expansive data centers of major cloud providers are significantly below that threshold.
This shift towards yottaflop-scale computing brings with it more than just technical challenges; it intersects with physical and infrastructural limitations. As the demand for significant computational resources grows, power consumption has emerged as a critical concern. Running large AI models demands substantial electricity, placing pressure on the U.S. power grid. Utilities have begun to warn that generation and transmission capacities are struggling to keep pace with this increasing demand.
Su’s projections imply that scaling compute to yottaflop levels will necessitate a transformation of energy infrastructure. The evolution will likely require new power plants, robust energy grids, advanced cooling technologies, and innovations in efficiency. The challenge surrounding yottaflops encompasses far more than semiconductor manufacturers; it touches on energy policy, industrial planning, and national infrastructure strategy.
Furthermore, there exists an economic dimension to this monumental shift. The financial burden of developing and maintaining yottaflop-scale systems is expected to be immense. As computing capabilities consolidate within a limited number of hyperscale operators, questions surrounding access, pricing, and competition are poised to intensify. Without new shared infrastructure models, smaller companies and research institutions may find themselves sidelined from the most advanced AI technologies.
Su’s keynote at CES positioned AMD as a pivotal player in this new era. She introduced the company’s next generation of AI accelerators, including the MI455 GPU, reinforcing AMD’s commitment to the data-center market. The company is increasingly targeting customers building extensive AI systems, such as OpenAI, as it seeks to narrow the gap with competitor Nvidia in high-performance AI hardware.
The timing of Su’s remarks coincides with a significant shift in the AI landscape, as the technology transitions from experimental phases to industrial-scale deployment. Governments are embedding AI into national strategies, businesses are integrating it into core products, and scientific research is leaning heavily on AI for innovation. This escalating demand for massive computational resources illustrates a landscape that was scarcely discussed a few years ago.
Ultimately, Su’s message at CES served not only as a forecast but as a clarion call. If AI continues its current trajectory, the world will need to rethink how computing power is created, powered, and regulated. The yottaflop, once merely a theoretical concept, is swiftly becoming the new benchmark in the computing landscape.
See also
Intel Launches Core Ultra Series 3 Processors, Powers New TGS-2000 Edge AI Computers
China Investigates Meta’s $100M Acquisition of AI Startup Manus Amid Tech Rivalry
Chiplets and UCIe Standard Enable 10x More Powerful AI Accelerators, Transforming Hardware Design
Interview Kickstart Launches Advanced ML Program to Help Engineers Transition to Top AI Roles
US and China Face Energy Bottlenecks in AI Race: Key Insights and Strategies Unveiled
