NVIDIA, AMD Battle to Set Agentic AI CPU Standard

Agentic AI repeatedly shifts between model inference, tool calls, code execution and data retrieval, putting CPUs back on the critical path alongside GPUs. NVIDIA is betting that faster completion of each sequential step will improve responsiveness and accelerator utilization, while AMD argues that enterprises need more cores and threads to run many agents and their supporting services concurrently. The contest pits NVIDIA’s Arm-based Vera design against AMD’s x86 EPYC franchise and its mature enterprise software ecosystem, with the winning metric likely to shape AI-factory procurement.

Bank of America Research said on July 22 that the server CPU addressable market could reach $170 billion by 2030, about four times its current size, while reiterating a Buy rating and $350 price target on NVIDIA. Vera combines 88 custom Olympus cores with 1.2 TB/s of memory bandwidth. AMD answered on July 23 with EPYC 9006, offering up to 256 cores and 512 threads, and says EPYC 9965 and next-generation Venice deliver 2.37 times and 3.30 times Vera’s throughput, respectively, in a modeled 100-kW rack.