Fewertokensperresult.FewerkWhpertask.Lesscarbonperdecision.
We publish the math.
A governed 31B open model running under ORCA™ typically uses 5–10× less energy per useful outcome than an ungoverned 400B frontier model doing the same task. That gap comes from two places: the model is smaller, and the governance does the work most models burn tokens to fake. Both effects stack.
This page documents the thesis, the assumptions, the six scales, and the public language we will and won't use. If your numbers contradict ours, we want to see them — apex@apexorca.io.
Thenumberswebuildon
Conservative, sourced where possible, and published with every claim. Replace any of these assumptions with your own and the math still favors the governed pipeline.
| Parameter | Value |
|---|---|
| Frontier (≥400B dense) inference | ~3–5 kWh per 1M tokens (H100/H200 class, 60% util, PUE 1.2) |
| Open 31B inference | ~0.5–1.0 kWh per 1M tokens (same hardware class) |
| Ungoverned vs governed token multiplier | 3–5× for same useful outcome (measured, not assumed — see methodology) |
| US grid average | 0.38 kg CO₂ / kWh |
| Canadian grid average | 0.13 kg CO₂ / kWh |
| Global grid average (IEA 2024) | 0.48 kg CO₂ / kWh |
| US home average | 10,500 kWh / year |
| Gasoline car average (EPA) | 4.6 metric tons CO₂ / year |
Pickyouraudience
Same thesis, six framings. The solo consultant cares about $40 vs $800 a year. The CSO cares about 190 homes and 165 cars. The civilization-scale framing is what makes a Chief Sustainability Officer pick up the phone.
| Scale | Ungoverned (frontier) | Governed (ORCA + 31B open) | Savings equivalent | Who this is for |
|---|---|---|---|---|
| 1.Solo consultant | ~200 kWh/yr · $800/yr · 76 kg CO₂ | ~10 kWh/yr · $40/yr · 4 kg CO₂ | ~95% reduction across energy, cost, emissions | Solo builders, sole proprietors, one-founder AI consultancies |
| 2.SMB (50 agents) | 10 MWh · $40K · 3.8 t CO₂ | 0.5 MWh · $2K · 0.2 t CO₂ | ≈ 1 US home powered for a year · ≈ 1 car off the road | Small-to-mid teams running a modest agent fleet |
| 3.Fortune 500 (10K agents) | 2,000 MWh · $8M · 760 t CO₂ | 100 MWh · $400K · 40 t CO₂ | ≈ 190 homes · ≈ 165 cars off the road | Enterprise deployments, especially where a CSO has to defend AI spend to a board |
| 4.1M governed agents | 200 GWh · 76,000 t CO₂ | 10 GWh · 4,000 t CO₂ | ≈ 18,000 homes · ≈ 16,500 cars off the road | Large platform operators; mid-scale public-cloud agent tenants |
| 5.50M agents (2027 projection) | 10 TWh · 3.8 Mt CO₂ | 0.5 TWh · 190 Kt CO₂ | ≈ 1 million homes · ≈ 780,000 cars — annual output of one 1,000 MW coal plant | Industry-wide projection for 2027 AI agent population |
| 6.500M agents (2030 ceiling) | 100 TWh · 38 Mt CO₂ | 5 TWh · 1.9 Mt CO₂ | ≈ 9 million homes · ≈ 8 million cars — footprint of a small country like Costa Rica | Projected ceiling of the 2030 agent economy |
Projections for scales 5 and 6 use independent industry forecasts for 2027 and 2030 agent populations; we do not claim to know the future, only the arithmetic if the assumptions above hold.
Whatwewillandwon'tsay
Acceptable public language
- “An ungoverned 400B frontier-model agent consumes roughly 5–10× more energy per useful outcome than a governed 31B open-model agent running under ORCA.”
- “If the projected 50 million AI agents in use globally by 2027 all ran under ORCA-style governance on small open models, conservative projections suggest energy savings equivalent to ~1 million homes powered for a year, or ~780,000 gasoline cars taken off the road.”
- “ORCA Agents are, by design, the lowest-energy-per-outcome agents we’ve been able to measure. We publish the math.”
Prohibited
- “Greenest agents on the planet.” (Unverifiable.)
- Any specific kWh / CO₂ / $ number not accompanied by methodology.
- Any competitor-specific claim without measurement.
- Any hyperbole not grounded in the assumptions above.
TheweeklyScaleCardformat
One post per week. Pick one scale. Show homes powered, cars off the road, dollars saved. Link back here for methodology. No hedging, no hype — just the arithmetic.
Template
If 1M agents ran ORCA instead of an ungoverned frontier model: · 18,000 homes powered for a year · 16,500 cars off the road · ~$600M saved in inference spend Same outcomes. Smaller model. Governed pipeline. Methodology: apexorca.io/efficiency
Brand line
“Apex Agents™ are built for outcome-efficient computation. Fewer tokens per result. Fewer kWh per task. Less carbon per decision. We publish the math — not marketing claims.”
This is the only AI product line whose sales copy can lead with a sustainability claim backed by public math. If you're a Chief Sustainability Officer, start here.