Token-based pricing is forcing a rethink of enterprise AI infrastructure, and IDC’s Ashish Nadkarni argues the shift runs deeper than a billing model change. As agentic AI workflows replace single-prompt interactions, token consumption is becoming the primary lens for measuring compute, memory, storage, and network efficiency. The emerging AI factory model<\/a> treats the entire infrastructure stack as an optimization target, not just the GPU layer, with the goal of eliminating token waste the way a manufacturer eliminates idle time on an assembly line.<\/p>\n The question of whether your organization is burning tokens efficiently is already a live cost question, even if your finance team hasn’t framed it that way yet. Enterprises running agentic workflows, where a model hands off to a retrieval system, evaluates results, and loops back through reasoning steps without human checkpoints, are the ones most exposed. If you have agentic pipelines in production and no telemetry (the system data showing how each stage consumes resources) tied to business outcomes, you’re flying blind on one of your fastest-growing cost lines.<\/p>\n The IDC framing here is directionally correct, though Nadkarni’s analyst role gives him an incentive to frame the problem as broadly as possible, which tilts the argument toward large-footprint infrastructure overhauls rather than targeted fixes. The more precise point is that token waste isn’t uniformly distributed. The recurring failure mode is poorly tuned orchestration, where agents retrieve information they don’t use, or loop through reasoning steps that don’t change the output. That’s fixable at the pipeline level before it requires rearchitecting your data center. Organizations that audit agentic workflows for retrieval and reasoning efficiency first will find the infrastructure optimization conversation much cheaper to have.<\/p>\n If your next GPU procurement cycle or data center refresh is within 18 months, the AI factory framing deserves a seat in that planning conversation now. The vendors who win that cycle will be the ones who can demonstrate token throughput per watt and per dollar, not just raw model performance. That’s the benchmark worth demanding from your infrastructure partners before the contracts renew. I’d revise that view if agentic workflows plateau in enterprise adoption, but every signal right now points the other direction.<\/p>\nWhat this means for your business<\/h2>\n