AMD CEO Dr. Lisa Su unveiled Helios, a massive rack-scale computing platform designed to shake Nvidia’s hold on the AI data center market. Helios is a double-wide server rack that weighs nearly 7,000 pounds (about the weight of two compact cars), built to be a “plug-and-play” supercomputer for the world’s largest AI companies. Delivering up to 2.9 exaflops of AI performance, according to Dr. Su, this system is set to be the “world’s best AI rack,” specifically built for training trillion-parameter models.
The system marks a major strategic shift for AMD, moving from selling individual chips to providing a fully integrated “Yottascale” infrastructure. Su has used this term “Yottascale” (10 to the power of 24) to describe the future of computing, and predicted that by 2030, the world will need 100 times more computing power than we have today.
According to their official statement, Helios is powered by the new Instinct MI455X GPUs and Zen 6-based “Venice” EPYC CPUs, offering an industry-leading 31TB of HBM4 memory which is roughly 50% more capacity than Nvidia’s rival flagship.
By utilizing open standards like UALink and Ultra Ethernet, AMD is positioning Helios as the flexible, cost-effective alternative for hyperscalers like OpenAI and Meta who are looking to scale beyond proprietary ecosystems.
The system is on track to launch later this year as part of AMD’s broader goal to grow global compute capacity 100x by 2030.
