Inference footprint

What the estate consumes to think

Every agent call in the estate is an inference call, and an inference call has an energy cost that no operating report currently carries. This page derives one for 2,778,394 calls a week across 12 service towers and 12 models, and states plainly that it is derived rather than metered.

0 of 30 rows metered

Health

Volume, energy, carbon, cost

Four readings on one week of estate inference. Each is computed from the same call count, so they move together.

Inference calls per week

2,778,394

6,695,929,540 tokens in · 1,066,903,296 tokens out.

Energy per week

3,260 kWh

1.17 Wh per call · roughly 169,541 kWh a year at this rate.

Carbon per week

1,255 kg

At a 385 g/kWh grid factor · roughly 65 tonnes a year.

Inference spend per week

$1.9m

Charged by the supplier. The energy above is consumed on the supplier's hardware, not ours.

Operations

Where the consumption sits

By service tower, then by model. A tower with high volume and a cheap model can consume less than a small tower running a frontier model on every transaction.

By service tower

12 towers · 3,260 kWh a week

Hire-to-Retire Tower

461 kWh

$270k

Record-to-Report Tower

431 kWh

$256k

Source-to-Pay Tower

400 kWh

$230k

Order-to-Cash Tower

328 kWh

$189k

Make-to-Deliver Tower

286 kWh

$165k

AI Operations Tower

265 kWh

$157k

IT Service Tower

230 kWh

$134k

Awareness-to-Lead Tower

216 kWh

$124k

Data and Analytics Tower

192 kWh

$115k

Concept-to-Launch Tower

169 kWh

$97k

Legal Operations Tower

167 kWh

$98k

Plan-to-Perform Tower

115 kWh

$69k

By model

10 models drawing power · 3,260 kWh a week

Helix Guard 3

1,186 kWh

$23k

Abraxas Swift 2

749 kWh

$167k

Abraxas Embed 3

485 kWh

$141k

Abraxas Reason 4

287 kWh

$514k

Helix 8B Instruct

211 kWh

$46k

Abraxas Reason 3

124 kWh

$215k

Helix 70B Instruct

100 kWh

$141k

Abraxas Vision 2

72 kWh

$124k

Corvid Counsel 2

44 kWh

$344k

Corvid Embed 1

2 kWh

$17k

2 retired models on the registry draw nothing and are not listed here.

Actions

What has to change before this can be reported externally

A sustainability disclosure cannot rest on a derived figure. These are the gaps between what this page shows and what an assured report would need.

No metered consumption from any supplier

30 of 30

Every row is derived from token counts and a published efficiency factor. No supplier on the registry currently exposes per-call energy.

One grid factor for the whole estate

385 g/kWh

A single blended factor is applied everywhere. Inference runs in several regions with materially different grid intensity, and that is not modeled.

Training and fine-tuning are out of scope

Not counted

Only inference is counted here. The energy embodied in training the models the estate calls sits with the supplier and is not disclosed to us.

Live observability

Consumption week by week

6 weeks of estate inference energy. The trend is the number that matters: unit cost falling while total consumption rises is the normal shape of a scaling estate.

Estate energy trend

2026-W28 to 2026-W33 · +26.6% over the window · none of it metered

2026-W28

2,576 kWh

992 kg

2,194,931

2026-W29

2,739 kWh

1,054 kg

2,333,851

2026-W30

2,869 kWh

1,105 kg

2,444,987

2026-W31

3,032 kWh

1,167 kg

2,583,906

2026-W32

3,163 kWh

1,218 kg

2,695,042

2026-W33

3,260 kWh

1,255 kg

2,778,394

Derived, not metered. Call volume comes from the gateway; tokens are the week-25 mean per call; energy is tokens multiplied by a published per-million-token figure; carbon is energy multiplied by a blended regional grid intensity of 385 grams per kilowatt hour. No provider on the registry discloses metered energy per request, so the absolute values carry an error band we cannot size. The direction and the relative ranking between towers are meaningful; the absolute kilowatt hours are an estimate.

Is policy and strategy coming to fruition

Can the organization answer a sustainability question about its agents

Three things a buyer or a regulator will ask, each answered from the record above rather than from intent.

Question

What does one transaction consume?

Answerable — 1.17 Wh per inference call, derived, and the derivation is printed.

Question

Is the figure independently verifiable?

Not answerable — 0 of 30 rows are metered. Nothing here can be assured.

Question

Is consumption rising or falling as the estate scales?

Answerable — rising 26.6% across the 6 weeks on record, against a call volume that rose over the same window.

What this page is, and what it is not

A security certification and a penetration test are organizational attestations. The product can house them, track their currency and surface their scope; it cannot manufacture them. Present the trust center as the place they live, never as evidence that they exist.

An energy figure is the same kind of claim. Not one number on this page was metered: every kilowatt-hour is token count multiplied by a published efficiency factor, and every kilogram of carbon is that figure multiplied by a single blended grid factor. It is enough to rank towers against each other and to watch a trend. It is not enough for a disclosure, and it should never be put in one.