AMD CEO Lisa Su said that AI compute used for running models now exceeds the compute used to train them for the first time.
This shift indicates that artificial intelligence is moving from a developmental phase into widespread operational use. As enterprises deploy AI services at scale, the demand for inference, the process of a model providing a real-time answer, has overtaken the initial training phase.
Su said these findings on July 20 during the launch event for the AMD Helios rack-scale AI system in Austin, Texas [2]. The Helios system entered full production in 2026 [2].
The transition reflects a broader trend in how companies utilize hardware. While the early years of the AI boom focused on building massive large language models, the current priority is the deployment of those models to millions of end users.
Looking toward future requirements, Su said that AI will eventually require 10 yottaflops of compute power [1]. This massive scaling requirement underscores the growing pressure on data center infrastructure to handle increasing workloads.
Industry projections suggest the global compute market will continue to expand rapidly. The market size is projected to reach $2 trillion by 2030 [3].
AMD is positioning its new Helios architecture to capture this growth by focusing on the efficiency of inference. By optimizing for the running of models rather than just the training of them, the company aims to compete with other major chipmakers in the evolving AI landscape.
“AI compute used for running models now exceeds the compute used to train them for the first time.”
The pivot from training to inference marks a critical maturity point for the AI industry. It suggests that the 'experimentation' phase of generative AI is ending and the 'utility' phase has begun, shifting the hardware priority from raw power for creation to efficiency for execution. For chipmakers, this changes the competitive landscape, as the value proposition shifts toward reducing the cost and latency of running models in production environments.
