AMD executive makes case for ‘agent computers’

As artificial intelligence (AI) shifts from occasional prompts to always-on agents, the cost of relying entirely on cloud computing could become a growing concern for developers, creators and businesses.

In an opinion piece, Alexey Navolokin, general manager for APAC, argues that the rise of AI agents could create a new category of computers designed specifically to run AI workloads locally.

Unlike conventional PCs used to occasionally access AI services, so-called “agent computers” would run models continuously for tasks such as coding, research, document processing, content generation and workflow automation.

Navolokin said cloud AI remains important, particularly for advanced models and demanding reasoning tasks. However, he argued that repetitive or high-volume workloads could increasingly be handled locally, potentially offering greater privacy, lower latency and more predictable costs.

The argument centers partly on economics. The piece estimates that an AMD Ryzen AI Halo system handling about 6 million tokens daily could, under its assumptions, avoid up to $750 a month in comparable cloud API costs and break even in around six months.

A higher-throughput Radeon AI PRO R9700 configuration, meanwhile, could reach break-even in about three months under the same comparison assumptions. Navolokin acknowledged that actual results would depend on factors including utilization, electricity prices, model choice and workload.

Local AI could also give creators more freedom to experiment with image, video, music and 3D generation without relying on cloud credits or monthly usage limits. The piece stresses, however, that humans would remain responsible for creative direction, selection and final output.