On July 23, 2026, at its Advancing AI 2026 event, AMD announced that its Helios rackscale system is in production and slated for deployment at gigawatt scale by leading AI companies. Each Helios rack combines 72 AMD Instinct MI455X GPUs with 18 sixth-generation AMD EPYC CPUs and AMD Pensando networking handling front-end, scale-up, and scale-out connectivity. AMD claims Helios delivers up to 30 percent more tokens per dollar than the leading competitive solution.
The accompanying CPU launch is the sixth-generation EPYC “Venice” line, with up to 256 cores and 512 threads with simultaneous multithreading enabled, support for up to 16 channels of 12.8 GT/s MRDIMM memory, PCIe Gen 6 connectivity, and boost frequencies up to 5 GHz on select parts. On the accelerator side, AMD says the MI455X delivers 34 times higher token throughput than the previous MI355X, while the MI430X targets scientific computing with up to 288 TFLOPS of hardware FP64 performance and the MI350P claims up to 4.2 times more tokens per second per dollar.
AMD also extended the announcement into edge and robotics with Ryzen AI Embedded X100 processors and the Kria AI system-on-module and robotics developer platform, framing the whole portfolio as full-stack compute for the agentic AI era.
The claim to watch is tokens per dollar rather than raw FLOPS. AMD is not arguing that it has the fastest chip; it is arguing that a customer running inference at scale gets more output per dollar of capital and power. That is the metric hyperscalers actually optimize, and it is the wedge that gives buyers a credible second source in a market where one vendor has set pricing largely unopposed.