EMILIA GRACE · PROOF-OF-CURTAILMENT

A verifiable demand-response rail for AI compute.

When the grid asks an AI datacenter to reduce load, GRACE proves who authorized it, what was allowed, whether the facility complied, and what should be paid — verifiable by anyone, offline, without trusting the operator’s own logs.

COSA moves the megawatts. EMILIA proves the move was authorized and delivered.

How it works The Gate Request pilot

WHY NOW

~100 GW of flexibility is sitting unused — because no one can verify it.

Duke University’s Nicholas Institute finds the 22 largest U.S. balancing areas could absorb 76–126 GW of new load if it can be curtailed under ~1% of hours — ERCOT alone ≈ 10 GW at 0.5%. But that headroom is only bankable if the curtailment is verifiable enough for a grid operator to count it as capacity. Today curtailment is self-reported and trust-based — baselines gamed, telemetry backfilled, sheds over-claimed. That measurement-and-verification gap is what GRACE removes.

HOW IT WORKS

Authorize → verify → shed → measure → prove → settle.

1

Authorize

A market-authorized party (ISO, utility, aggregator, or facility under the applicable tariff) signs a bounded grid.curtailment order — named human, or quorum for hard cuts.

2

Verify & gate

The facility controller verifies the order offline, fail-closed: posture changes only against a valid, in-scope, unexpired order. Spoofed or stale orders are refused.

3

Shed

The scheduler reduces compute — cache-first inference, deferred batch, capped GPU clocks — preserving life-safety lanes. Power falls. (COSA moves the megawatts.)

4

Measure

An attested meter / smart PDU signs the power telemetry at source, Merkle-anchored so it cannot be backfilled or cherry-picked.

5

Prove

Delivered = baseline − actual, computed from the signed telemetry against the program’s prescribed baseline method (pinned by hash).

6

Settle

A Proof-of-Curtailment Bundle is emitted — order + acknowledgment + attested telemetry + computed kW·h — verifiable offline. The program pays against proof, not self-report.

WHO IT'S FOR

Bankable on every side of the meter.

AI / HPC datacenters & neoclouds

Connect faster (prove flexible load for interconnection), get paid for verifiable curtailment, survive audit without exposing operational logs.

Grid operators / ISOs / utilities

Dispatch flexible compute and confirm delivery without trusting the operator’s self-reported logs. Settlement against cryptographic proof.

Demand-response aggregators

A portable, tamper-evident M&V artifact for large, fast flexible loads — higher-value DR products, fewer settlement disputes.

First mover: an AI/HPC datacenter or neocloud operator with a DR aggregator or utility sponsor — the party that holds the interconnection/payment incentive and a grid counterpart to settle against.

DEMONSTRATION

Available now: a runnable reference. Planned: real hardware.

A runnable reference of the full loop is published and verifies under the production EMILIA verifier — issue a grid.curtailment order, shed, sign attested telemetry, compute delivered kW·h, emit the Proof-of-Curtailment bundle — with the adversarial paths refusing (tampered telemetry → invalid; forged order → refused; replay → refused). The five-minute hardware demonstration adds a multi-GPU node + a smart PDU on a live wattage graph; it requires a host facility.

HONEST POSTURE

GRACE does not invent the baseline. It makes the program’s method un-fudgeable.

The baseline methodology belongs to the ISO/program (CAISO ELAP, PJM CBL, ERCOT). GRACE pins its hash and makes its application tamper-evident against method swaps, telemetry backfill, and input manipulation. EMILIA proves authorization and evidence integrity — a necessary, not sufficient, condition for trustworthy demand response. This is a critical-infrastructure vertical, not the NETL generation prize.

EMILIA Gate Verify a receipt Request pilot