AMD's Answer to the Agentic AI Token Bill: Put Half the Workload on AI PCs and Save Up to 60%

AMD has a message for enterprises watching their inference bills climb as AI agents move from pilot projects into departmental workflows: stop thinking of the cloud as the default destination for every token. Alexey Navolokin, AMD's Asia Pacific general manager, argued this week that "agentic AI is a distributed infrastructure challenge, with the right mix of compute needed across cloud, data center, edge and AI PCs" — and the company backed the argument with its own cost modeling.

The premise is that agents change the economics of AI usage. "As AI moves from occasional prompts to continuous workloads via agentic AI, token consumption can grow quickly, where systems repeatedly reason, call tools, and iterate," Navolokin said. "Each step consumes compute and system overhead, and in cloud environments, usage-based costs can accumulate." He cited Anthropic's State of AI Agents 2026 report, which found 57% of surveyed organizations already deploy agents for multi-stage workflows.

The numbers AMD puts behind the thesis are from its own analysis, and the company says so explicitly. At a medium workload tier of about 5.7 million input tokens and 574,000 output tokens per user per day — which AMD says is meant to represent a knowledge worker actively using an agent harness such as Claude Code, Codex or Hermes — a fleet of 500 AMD AI PCs running a 50% local and 50% cloud configuration would produce projected three-year savings of 40% to 60% compared with a cloud-only deployment, depending on the cloud model used. A fully local deployment would save more, with the initial hardware investment typically breaking even in under 24 months.

The company's single-machine illustration is more eye-catching still. AMD estimates an AI PRO R9700 desktop configuration could handle about 18 million tokens of AI use per day, with electricity costing roughly $64.80 per month. Under its cloud comparison assumptions, the three-year cost of the desktop comes to about $6,533, versus $81,108 for the equivalent cloud-based usage. AMD cautions that the figures depend on workloads, cloud models, hardware utilization, electricity rates and deployment configurations — all of which are AMD's own assumptions, and none of it has been independently verified.

The pitch fits a week in which local AI hardware has been unusually prominent. NVIDIA began selling a 64GB configuration of its DGX Spark desktop on October 23 at $4,999, explicitly marketed at developers running open models and always-on agents on their own desks rather than paying per token; the same launch admitted that memory costs forced it to halve the memory relative to the original 128GB model while raising the entry price above last year's $3,999. AMD, for its part, has been pushing its Ryzen AI platform and a Cisco partnership for enterprise agents that run on-premises under unified security policy.

Navolokin also framed the shift in terms of what agents do to jobs rather than replace them — "AI agents can multiply what employees are able to accomplish" — but the durable argument is about architecture. Frequent, latency-sensitive workloads belong on local or edge compute; large, elastic ones stay in the cloud. Whether the 40-60% savings survive contact with real procurement math is unproven, but the direction of travel is clear: as token consumption grows 14-fold in a year, the marginal cost of a token is becoming a line item that CFOs will negotiate, and chipmakers on both sides of the cloud boundary are only too happy to help them do it.