AMD says AI agents need more CPUs

AMD CTO Mark Papermaster said at the RAISE Summit in Paris that the market has begun to recognize a broader shift in AI infrastructure: agentic systems need far more than GPUs. AMD’s shares have moved from near $200 to above $500 in six months, while its market value is nearing a trillion dollars, up more than 140% in a year.

Papermaster said AMD laid the groundwork by shipping its first leadership server CPU in 2017 and maintaining an annual cadence since. Agentic applications place a heavier burden on CPUs because real workflows coordinate many agents, launch sub-agents for specialized tasks, and manage growing context before GPU-intensive computation begins. One figure he cited put current agent workloads at roughly four times the CPU work.

AMD is also moving beyond standalone chips toward optimized systems. AMD spent $4.9bn to buy ZT Systems so it could tune CPU, GPU, and networking clusters together, then kept the design expertise and sold the manufacturing arm to Sanmina. Its Helios rack-scale system combines 72 of the company’s next Instinct GPUs with server CPUs for large-scale training and inference.

Papermaster tied AMD’s strategy to openness and Europe’s demand for sovereign infrastructure. ROCm is open, Helios uses an open rack standard Meta submitted to the Open Compute Project, and AMD says its chips already power major European systems including Finland’s LUMI and France’s new exascale system. AMD plans to show Helios at its Advancing AI event on 22 and 23 July and teased next-generation parts using TSMC’s 2nm process.