At CES 2026, Dr Lisa Su linked AI’s future less to breakthrough models and more to the ability to build and scale the infrastructure underneath them.
When Dr Lisa Su talked about AI at CES, she stayed focused on what is already happening rather than what models might do next. AI’s momentum, she said, is evident. The question now is whether the industry can physically support it.
“We are still at the very beginning of the AI era,” Su said, pointing to adoption rates that are moving faster than previous technology shifts. What took decades with the internet is unfolding in years with AI, driving demand for computing power to levels the industry has not had to sustain before.
Global AI compute demand, she said, has grown from roughly one zettaflop in 2022 to more than 100 zettaflops in 2025, a hundredfold increase in three years. Even that, she said, falls short of what is coming. To support AI becoming embedded “everywhere,” the industry will need to scale compute by another two orders of magnitude over the next five years. The implication was straightforward. Progress is now limited less by ideas than by infrastructure.
Advertisement
That perspective changes how AMD is positioning itself. Rather than competing only on chips, Su described a shift toward building systems. “There is no one-size-fits-all solution in AI,” she said, noting that modern workloads rely on tightly integrated CPUs, GPUs, networking and memory. Performance gains increasingly depend on how those components work together, not on advances in any single processor.
That systems-first approach showed up in Helios, AMD’s new rack-scale AI platform. Designed for large-scale training and inference, Helios combines Instinct MI455 accelerators, EPYC CPUs and Pensando networking into a unified rack intended to operate as a single compute unit. Su described this as a response to rack-scale AI, where throughput, latency and energy efficiency depend as much on architecture and interconnects as on raw silicon performance.
Alongside Helios, Su introduced the MI455 GPU, the most complex chip AMD has built to date, with 320bn transistors and support for next-generation HBM4 memory. Paired with the Venice EPYC CPU, the platform is designed to feed data quickly enough to keep increasingly large models running without bottlenecks. AMD says the system delivers up to a tenfold performance improvement over the previous generation across key AI workloads. The larger point, Su suggested, was balance. Faster accelerators lose value if data cannot move through the system efficiently.
Su widened the implications beyond data centers. On personal devices, she said, AI is changing the role of the PC itself. “Your PC is becoming an AI PC,” she said, as local models take on tasks such as summarization, creation and translation without constant reliance on the cloud. That shift affects not only performance, but also privacy, latency and reliability as AI becomes part of everyday work.
The same logic carried across the industries highlighted during the session. In healthcare, Su pointed to AI-driven modeling that is accelerating drug discovery and improving early disease detection. In manufacturing and robotics, she described systems that can simulate and optimize complex processes before anything is built. In each case, the limiting factor was the same. These applications stall when compute cannot scale.
Su also acknowledged that infrastructure alone will not determine outcomes. Data access, energy consumption, workforce skills and open ecosystems sit alongside hardware as constraints. AMD, she said, is investing in education and partnerships to broaden access to AI tools, with the aim of expanding capability rather than concentrating it among a small number of players.
At CES 2026, Su kept the focus on structure and scale, linking AI’s next phase directly to the ability to build and scale compute foundations.
Advertisement
AMD