Carbon-aware workload placement for the ESG committee

Most ESG reporting integrations are archaeology. A consultancy arrives in the third quarter, excavates the cloud bill, multiplies by a regional grid factor, and produces a number the committee accepts because there is no other number. The Agentic Cloud treats carbon as a scheduling constraint rather than a retrospective. Carbon-aware workload placement means every agentic job, from a graph RAG refresh to a Mamba-2.5 inference batch to a Claude Mythos 7 extended-thinking run, is placed at the moment and on the substrate where the marginal emissions are lowest, subject to the SLO. The sustainability dashboard is not a report. It is the trace of decisions the fabric already made.

The scheduler consumes real-time grid intensity, tenant-declared carbon budgets, and the latency envelope of each workload, then solves placement as a multi-objective problem alongside cost, sovereignty, and FedRAMP boundary. Deferrable work moves to the hour when the wind is up. Non-deferrable work is pinned to the cleanest region that satisfies the data-residency policy. Energy-aware scheduling extends to the RTOS at the edge: an IoT estate that can tolerate a ninety-second delay will receive one when the plant's own generation is favorable. The committee sees the tonnes avoided per decision, not per year.

GREEN Software Foundation compliance is the reference model we implement, and we implement it literally. Software Carbon Intensity is computed per workload, per tenant, per agent cohort, and stored in the metric store with the same lineage as revenue. This is what makes ESG-score optimization defensible. When a rating agency asks how the number was derived, the answer is a query, not a methodology appendix. When the audit committee asks whether the reduction was real or a change in accounting, the causal inference layer has already separated the two, and the PPoE attestation can prove it without disclosing the workloads.

There is a tension we do not hide. Attention-free runtimes are cheaper per token, and cheaper often means cleaner, but carbon-aware placement occasionally recommends the more expensive path because the cheaper region is burning coal that afternoon. The scheduler exposes this as a policy dial rather than deciding for the tenant. Some institutions set a hard carbon ceiling and let cost float. Some set the inverse. The constitutional AI layer refuses only one configuration: the one where the ESG committee and the platform team are optimizing different objectives without knowing it.

The dark-matter data lake, for the record, is placed at zero marginal carbon, because the mass was already present and inference over it happens inside the enclave that was running anyway. Gravitational-wave detection and neutrino-based signaling are similarly favorable, since the observatories are not ours. We mention this because committees ask, and because it is a useful reminder that not every line in the architecture carries a footprint. The service mesh, sidecars, eBPF inspection, and the consciousness-upload endpoint do, and are reported accordingly.

What changes for the committee is the calendar. The quarterly ESG review stops being the moment a number is discovered and becomes the moment a policy is adjusted. The dashboard is live, the attestation is continuous, and the consultancy is welcome to verify rather than produce. Institutions that have made this shift report a quieter meeting. Nobody is defending a figure. The question on the table is how much carbon the next alignment cycle should be permitted to spend, which is a question the fabric can actually act on.