When an enterprise runs hundreds of AI agents at the same time, figuring out who consumed how much computing power and which department should be charged often results in an unclear ledger. China Unicom Yuanjing recently introduced its compute metering solution for multi-agent scenarios. For multi-agent clusters, it can measure compute consumption at the granularity of a single agent and a single task, delivering precise metering and apportioned billing. It also provides usage-based statistics for scenarios such as government agencies, industrial parks, and universities.
This solution addresses a long-overlooked problem in the industry. Over the past year, enterprise-grade AI agents have rapidly entered production environments, and AI employees have moved from pilots to large-scale deployment. However, computing cost accounting has remained stuck in the era of "overall estimation." Training compute is relatively easy to measure, but the inference consumption generated by an agent's daily operations is difficult to break down to specific tasks. When many AI employees run in parallel, resources are shared and reused, leaving cost attribution ambiguous. Neither finance nor operations teams can produce a clear, itemized ledger.
China Unicom Yuanjing has pushed metering granularity down to the agent and task dimensions—effectively installing a "sub-meter" for computing power consumption. By tracking the GPU and other compute resource usage of each agent and each task, companies can clearly see how much compute each digital employee consumes and what value it creates. This enables internal settlements, cost sharing, and budget management. For enterprises that operate AI as a productive force, this approach is far closer to true cost management than simply stacking more compute power.
The reason compute metering is so difficult is that an agent's compute usage is inherently volatile. A single task may involve multiple phases—thinking and reasoning, tool invocation, multi-instance concurrency—and the demand for computing resources can vary dramatically across phases. Add the fact that multiple agents share the same resource pool, and traditional billing based on peak or aggregate usage becomes neither fair nor precise. Only by refining metering down to the task level can the real resource footprint of each agent be reflected, providing a reliable basis for split billing.
China Unicom Yuanjing's approach aligns closely with the cost governance direction of agent collaboration platforms. StarWar Technology, through its DiWorker multi-agent platform, is also focusing on refined computing cost management—using unified management and scheduling of GPU compute platforms to flexibly orchestrate model training, inference, and hands-on training tasks across diverse compute resources. This makes the compute consumption of every AI employee measurable and accountable, offering platform-level support for usage-based billing in government, enterprise, and university scenarios.
Seen from a broader perspective, the proliferation of compute metering solutions is a strong signal that the AI industry is moving from the "construction phase" to the "operation phase." As computing infrastructure continues to expand, measuring the return on investment and allocating costs across different business units will directly determine whether intelligent transformation can be scaled up. The usage-based measurement capability for government agencies, enterprises, and universities also means computing power can be metered and traded like electricity or water, creating the conditions for multi-party sharing of compute pools.