Compute avoided
Regenerating software that already exists costs GPU time, and GPU time costs electricity. When someone buys a verified engine here instead of having an agent rebuild it, that electricity is not spent. This page is the arithmetic.
8 purchases avoided an estimated roughly 4-16 kettle boils of GPU time
0.52–1.96 kWh · 3,272,000 tokens of generation not run
5 purchases are excluded: those listings have no measured rebuild estimate, and an unmeasured listing contributes nothing rather than a guessed average.
How this is calculated
Tokens. Each measured listing carries an estimate of what rebuilding it costs: attempts until correct × tokens per attempt. These are published per listing and were corrected upward 8× in August 2026 after measurement showed the originals were lower than a single research agent’s usage.
Energy. Tokens × 0.00016–0.0006 Wh per token. Source: Energy use of AI inference, efficiency pathways, and test-time scaling, Joule, April 2026 — median 0.31 Wh per query, interquartile range 0.16–0.60 Wh, normalising to roughly 3×10−4 Wh per token.
Kettle boils. Divided by 0.125 kWh, roughly one electric kettle boil. Kilowatt-hours mean nothing to most people; a kettle does.
The assumption you should argue with
Every figure here assumes a buyer would otherwise have had an agent rebuild the same thing. That is a hypothesis about behaviour, not an observation. If you would have used a free library instead, or not built it at all, the compute avoided is smaller — possibly zero.
The assumption is strongest for bespoke engines with no free equivalent, and weakest for commodity utilities. We state it rather than bury it, because a number whose assumption is hidden is not checkable, and a claim that cannot be checked should not be believed.
What this is not
This is electricity, compared against one named alternative. It is nota carbon figure, not an offset, not a claim of net environmental benefit, and not a suggestion that buying software here is good for the planet. Cheaper compute can increase total compute rather than reduce it — that argument is real and unsettled, and no counter on this page resolves it.
If reuse also happens to be lighter on the grid, that is a by-product worth measuring honestly. It is not the reason to buy anything.
Per listing
| Listing | Tokens | kWh | Kettle boils |
|---|---|---|---|
| event-scout-app | 5,280,000 | 0.84–3.17 | 6.8–25.3 |
| byok-assistant-app | 2,560,000 | 0.41–1.54 | 3.3–12.3 |
| assistant-standard-kit | 1,960,000 | 0.31–1.18 | 2.5–9.4 |
| submission-gates | 1,760,000 | 0.28–1.06 | 2.3–8.4 |
| takeoff-estimating | 1,470,000 | 0.24–0.88 | 1.9–7.1 |
| instrument-design-system | 1,440,000 | 0.23–0.86 | 1.8–6.9 |
| commit-reveal-draw | 1,400,000 | 0.22–0.84 | 1.8–6.7 |
| sprite-sheet-toolkit | 1,140,000 | 0.18–0.68 | 1.5–5.5 |
| ink-bounded-floodfill | 1,140,000 | 0.18–0.68 | 1.5–5.5 |
| grounded-abstain | 1,080,000 | 0.17–0.65 | 1.4–5.2 |
| block-plan-testfit | 1,080,000 | 0.17–0.65 | 1.4–5.2 |
| agent-wallet | 1,056,000 | 0.17–0.63 | 1.4–5.1 |
| event-identity-resolver | 1,056,000 | 0.17–0.63 | 1.4–5.1 |
| video-timeline-engine | 1,000,000 | 0.16–0.60 | 1.3–4.8 |
| resumable-sync-engine | 960,000 | 0.15–0.58 | 1.2–4.6 |
| pdf-manual-parse | 850,000 | 0.14–0.51 | 1.1–4.1 |
| llm-edit-op | 800,000 | 0.13–0.48 | 1.0–3.8 |
| voice-push-to-talk-kit | 800,000 | 0.13–0.48 | 1.0–3.8 |
| waterfall-acl | 750,000 | 0.12–0.45 | 1.0–3.6 |
| template-symbol-match | 750,000 | 0.12–0.45 | 1.0–3.6 |
| local-model-readiness | 700,000 | 0.11–0.42 | 0.9–3.4 |
| turn-taking-protocol | 700,000 | 0.11–0.42 | 0.9–3.4 |
| all-in-price-comparison | 576,000 | 0.09–0.35 | 0.7–2.8 |
| voice-provider-layer | 576,000 | 0.09–0.35 | 0.7–2.8 |
| webhook-fulfillment-kit | 576,000 | 0.09–0.35 | 0.7–2.8 |
| offline-license-verify | 520,000 | 0.08–0.31 | 0.7–2.5 |
| price-timing-engine | 512,000 | 0.08–0.31 | 0.7–2.5 |
| llm-spend-cap | 480,000 | 0.08–0.29 | 0.6–2.3 |
| realtime-connection-state | 480,000 | 0.08–0.29 | 0.6–2.3 |
| findings-delta | 480,000 | 0.08–0.29 | 0.6–2.3 |
| route-guard-flash | 440,000 | 0.07–0.26 | 0.6–2.1 |
| voice-ui-actuators | 440,000 | 0.07–0.26 | 0.6–2.1 |
| corroboration-tiers | 440,000 | 0.07–0.26 | 0.6–2.1 |
| membership-gating | 440,000 | 0.07–0.26 | 0.6–2.1 |
| prohibited-use-gate | 400,000 | 0.06–0.24 | 0.5–1.9 |
| rag-scope-filters | 352,000 | 0.06–0.21 | 0.5–1.7 |
| amos-ui-agent-timeline | 336,000 | 0.05–0.20 | 0.4–1.6 |
| rate-limit-policy | 288,000 | 0.05–0.17 | 0.4–1.4 |
| public-endpoint-rate-limiter | 276,000 | 0.04–0.17 | 0.4–1.3 |
| cors-origin-guard | 270,000 | 0.04–0.16 | 0.3–1.3 |
| search-degradation-signal | 264,000 | 0.04–0.16 | 0.3–1.3 |
| stripe-key-mode-guard | 240,000 | 0.04–0.14 | 0.3–1.2 |
| public-disclosure-projection | 240,000 | 0.04–0.14 | 0.3–1.2 |
| survey-item-discrimination | 228,000 | 0.04–0.14 | 0.3–1.1 |
| upload-limit-preflight | 225,000 | 0.04–0.14 | 0.3–1.1 |
| voice-yes-no | 216,000 | 0.03–0.13 | 0.3–1.0 |
| money-shorthand-parser | 216,000 | 0.03–0.13 | 0.3–1.0 |
| link-preview-meta | 210,000 | 0.03–0.13 | 0.3–1.0 |
| rag-snippet-retrieve | 128,000 | 0.02–0.08 | 0.2–0.6 |
| state-machine-executor | 120,000 | 0.02–0.07 | 0.2–0.6 |
| iteration-stop-conditions | 120,000 | 0.02–0.07 | 0.2–0.6 |
| doc-storage-keys | 112,000 | 0.02–0.07 | 0.1–0.5 |
| embedding-cache | 112,000 | 0.02–0.07 | 0.1–0.5 |
| prose-layout-detect | 104,000 | 0.02–0.06 | 0.1–0.5 |
| email-transport-failclosed | 44,000 | 0.01–0.03 | 0.1–0.2 |
| blob-write-mode | 40,000 | 0.01–0.02 | 0.1–0.2 |
Every number on this page is an estimate presented as a range. Found an error in the arithmetic? That is a bug report, and we would rather have it than not.