AI compute,
dispatchable by the grid.
Data Joule turns OpenADR 3.0 demand-response events into measured inference curtailment on a Raspberry Pi — and settles each verified kWh on-chain as a Joule Credit on Polygon: Chainlink-verified, supply-bounded by Proof-of-Reserve. Flexible AI load that is measured, auditable, and verifiable.
AI load is becoming grid load.
It needs to become flexible.
Data centers and inference workloads are becoming material grid demand. Adding generation and wires takes years; demand flexibility can make today's capacity work harder during constrained hours.
Most AI compute treats power like a constant. It doesn't have to. Inference workloads can expose graded response: slow down, suspend, or recover based on a grid signal while preserving a measurable service trade-off.
Projected data center power demand growth by 2030
U.S. electricity load share from data centers: 2023 actual to 2030 high estimate
Cold-boot to inference-ready after a Tier 4 hard cutoff — measured on real hardware
How it works
Signal → control action → measured watts → public dashboard
Response ladder
Four progressive tiers — measured on real hardware, not theoretical.
TIER 0Baseline—
ondemand governor, full inference active
TIER 1Throttle−20%
conservative CPU governor, inference continues
TIER 2Power-save−40%
powersave CPU governor, inference continues
TIER 3Suspend−70%
SIGSTOP sent to llama-server process
TIER 4Halt−95%
controlled shutdown path — node goes offline
Joule Credits (JLC)
Each completed demand-response event triggers an on-chain verification request. Multiple independent oracle nodes fetch the physical wattage measurement and reach consensus, and a Proof-of-Reserve ceiling keeps token supply within the verified kWh.
The node is running right now.
Live telemetry from mtl-edge-01 in Montréal. Wattage, tier, and LLM status updated every 5 seconds. Watch a DR event arrive and the load drop in real time.
Why this matters
Demonstrates that AI loads can become controllable grid resources. Several regulators have signaled intent to mandate OpenADR 3 — the standard this project already runs.
Connects power flexibility to SLA impact, making curtailment a measured operating mode instead of a vague sustainability claim.
Open-source testbed for grid-interactive compute, reproducible on commodity hardware and inspectable end to end.
Each verified DR event becomes a Joule Credit (ERC-20) on Polygon — decentralized Chainlink verification of the measurement, plus a Proof-of-Reserve cap so supply can never exceed verified kWh, with an immutable audit trail. The on-chain settlement rails for demand-response.
About this project
Data Joule is an Internet of Energy portfolio project built to demonstrate that AI edge compute can participate in real-time grid flexibility — and that curtailment events can be recorded and settled on-chain with verifiable, decentralized oracle attestation. The full stack — from VTN deployment on a VPS to Chainlink oracle verification on Polygon — was designed, deployed, and tested as a working proof point.
The hardware runs 24/7. The telemetry is measured, not simulated. The OpenADR signals come from a real VTN. When a DR event completes, the measured curtailment is recorded and exposed on-chain for verification — Chainlink's oracle network attests the kWh reduction and a Proof-of-Reserve cap bounds the JLC that can be minted. The measurement is real; minting is gated, by design, until the reserve is independently attested.