Use Cases

Energy and carbon you can attest to

All Use Cases AI & ML Enterprise Neocloud Grid Sustainability Sovereign AI Compliance Investors
Energy & Carbon Attestation

Stand behind the numbers back to the rack

Sustainability and energy leaders are increasingly asked to disclose AI energy and carbon figures — and to stand behind them. The hard part isn't producing a number; it's tracing it. Spark-XC sits above your GPU, scheduler, and facility systems and validates power actions. Mission Control executes. Spark-XC validates.

Spark-XC ties each energy and carbon figure to a measured power delta, the rate and emissions factor it was priced against, and the Power Event Record that proves it. This is attestation, not a savings promise — every figure is a measured delta priced against a stated factor, never a guaranteed saving. Facility correlation reconciles GPU-side action with DCIM, BMS, PDU, and utility data; the tamper-evident evidence chain makes the figures auditable.

  • Every energy and carbon figure traced to a measured power delta — not an estimate
  • Each figure carries the explicit rate and emissions factor it was priced against
  • Facility-correlated against DCIM, BMS, PDU/UPS, and utility data
  • Every figure backed by a Power Event Record, independently replayable
  • Disclosure-ready exports for EED, CSRD, and climate reporting
Anatomy of an Attested Figure
1 · Measured power delta
power_before → power_after, read back from hardware
2 · Rate & emissions factor
the explicit $/kWh and kgCO₂e/kWh it was priced against
3 · Power Event Record
delta + factor on the tamper-evident chain, replayable
An attestation, not a savings claim
Spark-XC reports a measured delta priced against a stated factor. It does not assert guaranteed savings or "zero impact" — it proves what was measured and what assumption it was priced under.
What Every Figure Carries

Disclosure you can defend line by line

100%
Figures traced to a PER
1:1
Explicit rate & emissions factor
Per-action
Facility-correlated
Auditable
Independently replayable
Disclosure Scenarios

Real reporting situations, traced and proven

Scenario · Energy Reporting
Producing an energy report an auditor accepts
An auditor asks the energy manager to substantiate the AI energy figures in this period's report. Instead of a spreadsheet of estimates, each figure resolves to a measured power delta, the rate it was priced against, and the Power Event Record behind it. The auditor can replay any line independently — no need to trust the system that produced it.
Every reported figure resolves to a measured delta and a replayable record.
Scenario · Carbon Accounting
Carbon figure traced to delta + emissions factor
A carbon figure in the ESG disclosure is challenged. Spark-XC shows the measured power delta in kWh, the exact emissions factor (kgCO₂e/kWh) it was priced against, and the source of that factor. The number is presented as a measured delta under a stated assumption — not as a guaranteed reduction — so it holds up under scrutiny.
Carbon stated as measured kWh × an explicit, sourced emissions factor.
Scenario · Regulatory Disclosure
EED / CSRD-style disclosure backed by PERs
A datacenter-reporting obligation under the EU Energy Efficiency Directive — or a CSRD or SEC/California climate disclosure — requires energy and emissions figures the organization can stand behind. Spark-XC supplies each figure with its measured delta, priced factor, and tamper-evident Power Event Record, so the disclosure is defensible from the filed line item back to the rack.
Disclosure figures backed by tamper-evident, replayable evidence.
Scenario · Reconciliation
Reconciling GPU-side energy with facility data
The GPU-side energy figure doesn't match what the facility's DCIM and utility meters report. Spark-XC's facility power correlation captured the rack-level effect alongside each action, so the gap can be reconciled against PDU, BMS, and utility data — turning a disputed number into a traced, shared account across energy and facility teams.
GPU-side and facility-side energy reconciled against one record.
How the Attestation Holds

What makes a figure defensible

Measured Power Delta
Each figure starts from a real, hardware-read change in power — power_before and power_after, confirmed by readback — not a modeled or assumed estimate. The kWh underlying every cost and carbon number is something that actually happened.
GPU Telemetry
Explicit Rate & Emissions Factor
Every cost or carbon figure carries the explicit $/kWh rate and kgCO₂e/kWh emissions factor it was priced against, plus the source of each. The assumption is recorded with the figure — so the number is a measured delta under a stated factor, never a standalone savings claim.
Workload Context
Facility Power Correlation
Spark-XC correlates GPU-side action with DCIM, BMS, PDU/UPS, and utility data, capturing the rack-level effect alongside each action. GPU-reported energy reconciles against what the building and the meter actually saw.
Facility Correlation
Tamper-Evident Evidence Chain
Every governed action is a Power Event Record on an append-only, SHA-256 tamper-evident chain (ARIV). Any modification is detectable, and each record is independently replayable — so disclosed figures rest on evidence, not assertion.
Evidence Chain
Disclosure-Ready Exports
Export tamper-evident Power Event Records for any period, device, or action type — structured for EED datacenter reporting, CSRD, and SEC/California climate disclosures. Each line resolves to its delta, its factor, and its proof.
Evidence Chain
Audit-Aligned by Design
Because each figure is replayable and tamper-evident, the same evidence that supports an energy or carbon disclosure also supports SOC2-style and internal-audit review — without bending the focus away from energy and carbon reporting.
All Paths
Efficiency Metric

The unit that matters: joules per token

Watts and temperature describe the hardware. The figure that decides whether governance actually pays off is energy per unit of work — joules per token. The only question worth answering is whether adaptive governance cuts joules per token beyond what a static power cap already gives you, at equal-or-better throughput.

Spark-XC measures it with an open benchmark designed to survive a skeptical engineer re-running it on their own rented hardware:

Exact energy, not sampled
Reads the GPU hardware energy accumulator (nvmlDeviceGetTotalEnergyConsumption) before and after the run — no power-sampling integration error, no spikes missed between samples.
Fixed work, not fixed time
Every arm runs the same prompts with greedy decoding, and the generated token IDs are hashed to prove the denominator is identical — so it's joules per token, not joules per second of a slower, capped GPU.
Three arms, honest control
Baseline · static power cap · Spark-XC. The win condition is beating the static cap at equal-or-better throughput — not just the uncapped baseline, which only re-proves that a cap saves power.
Validation in progress — no J/token figure claimed yet
Committed hardware results to date cover power, temperature, and utilization (B200 −38.8% power, H100 −26.8% power, utilization unchanged). A joules-per-token-anchored A/B is the active next step — we publish the number when the measured result is committed to the evidence set, not before. The benchmark itself is open in the repository at validation/jpt_bench.py.
Attestation Preview

What an attested figure looks like in the record

Each disclosed energy and carbon figure resolves to a Power Event Record carrying the measured delta, the explicit factors it was priced against, and the proof. The cost and carbon fields are measured deltas under stated factors — not guaranteed savings:

{
  "event_type":        "POWER_CAP_APPLIED",
  "timestamp":         "2026-06-12T02:41:08.220Z",
  "gpu_id":            "GPU-03-NODE-08",
  "validation_path":   "facility_correlation",
  "power_before_w":    700,
  "power_after_w":     520,
  "measured_delta_w":  180,
  "duration_h":        4.0,
  "measured_delta_kwh":0.72,
  "rate_usd_per_kwh":  0.094,
  "rate_source":       "utility_tariff_2026Q2",
  "emissions_factor":  0.371,
  "factor_unit":       "kgCO2e_per_kwh",
  "factor_source":     "grid_region_avg_2026",
  "cost_delta_usd":    0.068,
  "carbon_delta_kgco2e":0.267,
  "basis":             "measured_delta_priced_against_stated_factor",
  "facility_ref":      "DCIM-PDU-08-RACK-12",
  "seq":               61204,
  "prev_hash":         "7c12...a8e4",
  "entry_hmac":        "e93a...1d57"
}
Measured delta, not estimate
Explicit rate & emissions factor
Facility-correlated
Tamper-evident & replayable
Sustainability

Disclose energy and carbon you can defend

Bring a real power action and we'll replay its Power Event Record — the measured delta, the rate and emissions factor it was priced against, and the tamper-evident proof behind it. We work with sustainability, energy, and finance teams to make AI energy and carbon disclosures defensible from the line item back to the rack.

Request an AI Power Event Replay Explore Compliance Use Case