AMD Bets on Helios Rack and CPU Integration to Close Gap With Nvidia

AMD, a perennial runner-up to Nvidia in artificial intelligence chips, is trying to narrow the gap by selling racks that combine multiple types of semiconductors rather than individual chips.

Chief Executive Officer Lisa Su took the stage at AMD's Advancing AI 2026 event in San Francisco on July 23 and held up the chips used in the company's new Helios rack, calling it “the best rack in the world today.” A rack is a steel system that integrates graphics processing units, central processing units and networking chips. Racks form servers, and servers form data centers.

The centerpiece of AMD's announcement was its move to expand beyond selling chips individually and instead market systems at the rack level. Helios includes 72 of AMD's latest Instinct MI455X GPUs, 18 sixth-generation Venice CPUs and the company's Pensando networking technology. AMD said Helios can process as much as 30% more AI workloads at the same cost as Nvidia's latest Vera Rubin NVL72 rack.

The rack-focused strategy reflects AMD's effort to catch Nvidia in the AI chip market, where it has long remained in second place. The rise of AI agents that perform tasks on their own, rather than chatbots that simply answer users' questions, is creating an opening for AMD because it increases the importance of CPUs, an area where the company is stronger. AI agents require more CPUs to handle complex computing tasks.

Su also brought representatives from AI model developers Anthropic, OpenAI and Meta onto the stage, highlighting AMD's expanded influence. AMD announced a partnership the previous day to supply Anthropic with 2 gigawatts of GPUs, and in February signed a deal to supply Meta with 6 gigawatts of GPUs. OpenAI also signed a 6-gigawatt agreement with AMD and began adopting the chips in the second half of this year.

Kim In-yeop, Silicon Valley correspondent, Hankyung.com inside@hankyung.com