Tokenomics · Environmental
The environmental cost of AI inference.
Cloud vs self-hosted at your volume — kWh, kgCO₂e and water — using DEFRA UK grid factors. The numbers a Scope 2 auditor will want, not "cups of tea."
At your volume — 600 requests/month
Adjust on the Tokenomics calculator; the figures here update with the values you carry over.
| Deployment | kWh / month | kgCO₂e / month | Litres water / month | kgCO₂e / 1M tokens | Source |
|---|---|---|---|---|---|
| Cloud (avg) · scaled from Google's published benchmark | 0.209 | 0.0240 | 0.227 | 0.050 | |
| Self-hosted · 8B · RTX 5070 Ti | 0.865 | 0.1791 | — | 0.373 | |
| Self-hosted · 14B · RTX 4090 | 0.055 | 0.0114 | — | 0.024 | |
| Self-hosted · 70B · RTX PRO 6000 WS | 2.719 | 0.5629 | — | 1.173 |
The cloud vs self-hosted picture
Cloud providers run at scale with optimised hardware and (usually) lower-carbon grids — Google's hyperscale facilities sit on a power mix that's roughly half the CO₂e/kWh of the UK domestic grid. So at this volume, cloud is typically lower-CO₂ than UK self-hosted. That gap closes (and reverses) when:
- You're already on renewables. A self-hosted deployment behind on-site solar or a corporate PPA is single-digit kgCO₂e/MWh, beating most cloud regions.
- You can reuse the waste heat. Inference exhaust into a building loop, swimming pool, or process heater turns 100% of the GPU's electrical input into useful thermal output — the only honest path to "AI-positive."
- You're at sub-batch volumes. Cloud's per-prompt overhead (load balancers, multi-region routing, idle pods) doesn't amortise well below ~10k requests/day. A small on-prem deployment doing 100/day will often beat cloud on absolute kWh.
Scaled to enterprise volume
The numbers at SME volume look small. They're not at scale.
For a typical enterprise AI-chat deployment (50,000 messages/month at ~800 tokens each):
- Cloud: ~2.0 kgCO₂e/month, ~24 kg/year.
- Self-hosted 8B on UK grid: ~14.9 kgCO₂e/month, ~179 kg/year.
That's the figure a sustainability auditor wants for Scope 2 reporting. It's modest compared to the rest of an organisation's IT footprint (a single laptop runs ~50 kgCO₂e/year on continuous use) — but it's material, not negligible, and it scales with volume rather than headcount.
How this fits your wider ESG footprint
AI inference is one slice of Scope 2. The bigger picture — measurable AI alongside sustainable infrastructure decisions and defensible governance — sits in our ESG hub. The methodology here mirrors what we use for portal-side energy reporting in Horizon Portal: primary measurement, not vendor estimates.
Methodology
- UK grid emissions factor: 0.207 kgCO₂e/kWh (DEFRA 2025 average for grid electricity supplied to end users). Update annually with the BEIS conversion factors release.
- Cloud benchmark: 0.24 Wh, 0.03 gCO₂e, 0.26 mL per median 550-token prompt (Google, May 2024). Scaled linearly to your token volume — directional, not authoritative for non-Google providers.
- Self-hosted energy: measured GPU power-draw + 100 W system overhead × hours-per-year × 24/7 × electricity rate. Tokens-per-second from community benchmarks or our own anonymised production telemetry.
