One open schema for per-request AI energy, cooling, carbon and water: the measurement layer that turns "sustainable AI" from a claim into a fact.
Anyone can say their AI is sustainable. Almost no one can prove it — because there's no shared, independent way to measure energy use, water draw, and emissions across the AI and cloud infrastructure the whole industry runs on.
The Stardust telemetry standard is that shared way. It's the first of the four parts of Stardust, AIforE's open, end-to-end AI sustainability operating system, and the foundation the other three build on. It's published as RFC 0001, a working draft open for public comment.
It's built on OpenTelemetry and aligned with the Software Carbon Intensity specification (ISO/IEC 21031). The specification text is licensed CC BY 4.0, and the schemas and reference code Apache 2.0.
Measure once, in one shared format, and every part of the system can use and check the same numbers.
You are here. The open schema and data pipeline that records what each AI request consumes, at the depth each reporter can see.
In development. Benchmarks will emit the same telemetry, so energy per token and useful output per kWh will be measured on real requests and compared like for like across models, serving frameworks and hardware, with methods and raw results published.
Any router that claims savings, including AIforE's own reference router, must report its choices, its own overhead and any quality loss in this schema, so right-sizing savings can be checked by anyone.
An organization's measured emissions, water and energy are matched like for like to verified providers through the AIforE carbon marketplace, so every offset or credit traces back to a measured footprint.
AIforE intends to run an open, independent certification program for the energy, carbon and water performance of AI systems. We're building it in public, in stages, so the criteria rest on published methods and real measurements, not marketing claims.
One vendor-neutral telemetry schema for AI inference energy, carbon and water.
Will measure energy per token and useful output per kWh, with published methods and raw results.
Models, platforms and data centers are measured openly before any pass/fail thresholds exist.
Set by an independent Certification Committee after public benchmarking and comment.
Certification decisions will be made by an independent committee under AIforE's Conflict of Interest & Certification Independence Policy (draft). AIforE's own reference software, including Stardust and the Efficiency Control Plane, is never eligible for certification and appears on leaderboards as “AIforE reference, unranked.”
The same electricity has a different cost and a different carbon footprint depending on the hour. Prices move with wholesale markets, demand-response events, and time-of-use tariffs, and the grid's carbon intensity swings with whatever is generating at that moment. So wherever time changes the cost, our telemetry records the timestamp and the grid region alongside the number, rather than a single annual average.
The grid signals come from real-time sources such as Electricity Maps (generation mix and carbon intensity) and WattTime (marginal emissions, meaning what actually responds when demand changes). We use both because average and marginal intensity answer different questions.
Once the hour is measured, it becomes something engineers can act on. Batch training runs, embeddings, and other asynchronous workloads can wait for the hours when the grid is cleanest or when renewable generation would otherwise be curtailed, at no cost to anyone waiting on an answer. We're building the telemetry so an operator can see what a workload would have cost, in dollars and in carbon, had it run at a different time.
Before a chip computes anything, it has already cost something: mining, fabricating silicon wafers, and manufacturing high-bandwidth memory (HBM). AI accelerators are also replaced quickly, often within two to four years, which spreads that upfront cost over a short working life. Ignoring it makes a workload look cleaner than it is.
Our lifecycle measurement already accounts for upstream hardware manufacturing. We're extending it so embodied carbon is attributed per workload across the hardware's useful life, with the system boundary and confidence recorded each time. Some of these figures can't be traced today, so every value carries a confidence tag, and we'll work directly with AI infrastructure operators to close the gaps.
Our Recycled pillar started with heat and water. Hardware belongs there too. We're working toward a lifecycle standard for data-center hardware: how long it runs, whether it's redeployed or refurbished, and how it's recovered and recycled at the end. Recycling of electronics and other materials is already tracked as its own impact type, with recovery attested by certified recyclers.
This is early work with real gaps, and it will be shaped with the operators who run the equipment.
The goal is a standard that works anywhere in the world, not a proprietary format. Our telemetry is designed to align with OpenTelemetry, so it flows through the observability tools operators already use, and it maps to existing standards rather than competing with them, including the Software Carbon Intensity specification (ISO/IEC 21031), the GHG Protocol, and ISO/IEC 30134 data-center KPIs. It's published for public comment as RFC 0001, and AIforE proposes new attributes upstream to OpenTelemetry and the Green Software Foundation under vendor-neutral names. It's developed in the open as Project Stardust.
Building out the data pipeline and telemetry standards across major cloud/AI providers.
Publishing the benchmarks in the open, so any company — or any other nonprofit — can build on them.
The research and validation work behind AIforE's planned independent certification program.
This is infrastructure for the whole field, not just one product — support here underwrites the standard everyone else will eventually measure against.
AIforE is a 501(c)(3) tax-exempt nonprofit organization. EIN: 42-2272946. Donations are tax-deductible to the extent allowed by law.