The Physical Cost of Useful AI Work.
Task-normalized energy accounting for multi-step and agentic AI workflows.
Abstract. Invocation-level metrics such as joules per token and joules per response are useful for serving diagnostics, but they can misstate the physical cost of completing a user-valued task when workflows include retries, tool calls, long context, validation and failure recovery. CL-WP-005 defines a quality-gated accounting method in which all measured attempt energy is retained and normalized to accepted task outcomes. This release establishes the method, measurement boundaries, open evidence synthesis, benchmark schema and reproducibility assets. It does not claim a completed Critical Load primary GPU/node benchmark.
Measure energy against work that actually clears the quality gate.
CL-WP-005 defines Energy per Accepted Task (EAT) as the primary reporting unit for multi-step AI workflows. Failed attempts and retries remain in the numerator rather than disappearing from the accounting boundary.
Total measured energy across attempts divided by accepted tasks.
Energy consumed after first attempts is separately reported.
GPU, CPU+GPU, node and facility are never silently mixed.
Quality, energy and latency remain visible rather than collapsed into one score.
Open measurements establish the physical layer; the Critical Load contribution is the accounting and release discipline.
The release builds on peer-reviewed and open measurement work while keeping every external result labelled as external evidence.
| Evidence | What it establishes | Role in CL-WP-005 |
|---|---|---|
| TokenPowerBench / AAAI 2026 | GPU/node/system measurement framework and public inference records | Physical measurement foundation + worked open record |
| ML.ENERGY v3 | 46 models · 7 tasks · 1,858 configurations on H100/B200 | Shows task/configuration dependence of inference energy |
| A-LEMS / 2026 preprint | Energy per Successful Goal and orchestration-overhead accounting | Goal-level evidence; prevents false novelty claims |
| Sustainable Web Agents / AAAI 2026 | Higher energy does not necessarily imply better task performance | Motivates explicit quality-energy comparison |
| NVIDIA NVML / AIPerf | Cumulative GPU energy counters and telemetry collection | Primary campaign implementation path |
External evidence is not relabelled as a Critical Load measurement. The machine-readable evidence register records source, date, evidence class, key fact, URL and Critical Load use.
Quality gates come before energy comparison.
A task is frozen before measurement: input, acceptance rule, scoring method and maximum-attempt policy are declared in advance.
Failures and retries stay inside the energy numerator. Unaccepted tasks are not treated as delivered useful work.
Freeze
Lock task inputs, acceptance criteria, model/runtime and retry policy.
Measure
Record cumulative energy deltas and diagnostic telemetry for every attempt.
Score
Apply the predeclared quality rule without changing the threshold after results are seen.
Aggregate
Include failed attempts, retries, tool use, latency and accepted outcomes.
Compare
Report the quality-energy-latency frontier and matched orchestration baseline.
A GPU measurement cannot be silently promoted into node or facility energy.
| Boundary | Measured scope | Use |
|---|---|---|
| B0 | GPU only · cumulative energy counter or integrated power | Accelerator-attributable comparisons on the same platform |
| B1 | CPU + GPU · component energy measured separately | Host-aware workload comparison |
| B2 | Full node · PDU/BMC/node meter | Preferred operational deployment comparison |
| B3 | Facility · metered and explicitly allocated | Facility-level planning when the wider boundary is actually measured |
The method is usable now, even before the first Critical Load primary leaderboard exists.
Benchmark design, model/workflow routing, reasoning-policy tests, engineering regression, procurement measurement requirements and energy-aware capacity planning.
Universal model rankings, universal agent-efficiency claims, facility-energy claims without facility measurement, or statements that Critical Load has completed its own primary benchmark campaign.
Paper → task registry → attempt log → measurement code → accepted-task metrics → release checks.
The package is structured so the method can be executed on a declared NVIDIA environment without trusting the website or manually rebuilding the data schema.
What is published now—and what remains for v1.0.
Research question, UAWM method, evidence synthesis, boundary taxonomy, task/run schemas, executable measurement wrapper, metric code, validation logic and open release package.
Critical Load-owned repeated GPU/node telemetry, statistical uncertainty from the primary campaign, scoring audit on complex tasks and independent systems/energy + task-domain review.
The publication class is therefore Working Methods Paper v0.9. The pending empirical campaign is a visible release gate, not an omitted limitation.
Cite the method at the strength it actually supports.
Critical Load Research Group (2026). The Physical Cost of Useful AI Work: Task-normalized energy accounting for multi-step and agentic AI workflows. CL-WP-005 / R01 / R03 / CL-P01. Working Methods Paper v0.9, 19 September 2026. DOI pending.
The external measurement foundation remains directly traceable.
The web summary does not replace the evidence register. Core references are linked below; the downloadable CSV records evidence class, key fact, source date and intended use.
| Source | Role | Open |
|---|---|---|
| TokenPowerBench · AAAI 2026 | GPU/node/system measurement foundation | Paper ↗ |
| ML.ENERGY Leaderboard v3.0 | Large inference-energy benchmark landscape | Source ↗ |
| A-LEMS · Energy per Successful Goal | Goal-level energy and orchestration evidence | Preprint ↗ |
| Promoting Sustainable Web Agents · AAAI 2026 | Energy-quality evidence for agentic workflows | Paper ↗ |
| NVIDIA AIPerf telemetry | Implementation reference for GPU telemetry | Docs ↗ |
| CL-WP-005 evidence register | Machine-readable source trail | CSV → |