Every month we estimate the energy and carbon of all of Merdial’s AI activity and publish the result here: the good, the uncertain and the part we still have to fix. Latest month: September 2026.
An estimate, stated as a range
AI providers do not publish the energy each request uses, so no software company can measure this exactly. We estimate it from our own metered usage and published research, and we show a low, a central and a high figure rather than one number that looks more precise than it is.1. September 2026 at a glance
| Metric | Value |
|---|---|
| Energy, all AI activity | 0.75 kWh (range 0.24–1.56) |
| Carbon, location-based | 0.38 kg CO₂e (range 0.12–0.79) |
| Independent cross-check (Google’s measured median prompt) | 0.57 kWh |
| AI requests metered | 2,373 |
| Call audio transcribed | 155.1 minutes |
| Share of carbon on renewable-matched clouds | 9% |
| Largest source | DeepSeek, our heavy-lane model (China): 79% of carbon |
2. Efficiency
| Measure | This month | Why it matters |
|---|---|---|
| Requests on small models | 35% | Each task goes to the smallest model that can do it. |
| Input served from cache | 44% | Repeated instructions are reused, not recomputed. |
| Reasoning tokens | 0 | Reasoning is off by design. It multiplies compute without improving our tasks. |
| Requests that had to run twice | 371 | When a model fails, its backup re-runs the request, so energy is spent twice. We are working to reduce this. |
| Models trained | 0 | Merdial trains no models, so there is no training footprint. |
3. How we estimate it
- Usage: the input and output tokens of every AI request, from our metering ledger, plus the minutes of call audio we transcribe.
- Energy per token: 0.0001–0.002 Wh per generated token, from published measurements. Small models sit at the low end and large models at the high end. Input tokens count at one tenth, and cached input at zero.
- Whole data centre, not just the chip: AI accelerators account for 58% of serving energy (Google, 2025), so we scale up by 1.72.
- Speech-to-text: estimated from typical GPU transcription speed, because no vendor publishes it.
- Carbon: each provider’s electricity grid. The US is 384 g and China 560 g of CO₂ per kWh (Ember, 2024). These are location-based figures, so they don’t claim credit for renewable purchases.
4. What is included
| Included | Not yet included |
|---|---|
| All generative-AI requests (scoring, live coaching, summaries, learning), including our own testing. Speech-to-text on call audio. | Database and web hosting, staff devices, travel, and the carbon embodied in hardware. |
5. Where our compute runs
| Provider | Used for | Energy |
|---|---|---|
| AWS (behind Supabase and Vercel) | Hosting | 100% of electricity matched with renewables in 2023 and 2024. Data-centre PUE 1.15. |
| OpenAI on Microsoft Azure | Live coaching, backup scoring | Microsoft has committed to a 100% renewable supply from 2025. |
| DeepSeek | Post-call scoring (primary) | No renewable disclosure. China’s grid averages 560 g CO₂ per kWh. |
| TypeSafe AI, Deepgram | Live checks, speech-to-text | Not disclosed. |
Our weak spot, stated plainly
Most of our estimated AI carbon comes from one provider, on a coal-heavy grid, with no renewable disclosure. Moving that workload to a renewable-matched provider is under review. When it changes, this page will show it.6. What we do about it
| Step | What it means | Status |
|---|---|---|
| Avoid | No model training. No AI on calls with no human contact. Pre-recorded answers where no model is needed. | In place |
| Reduce | Small models, no reasoning, caching, compact embeddings, and serverless infrastructure that scales to zero. | In place |
| Fix waste | Fewer requests that have to run twice. | Tracked monthly |
| Switch | Move the largest workload to renewable-matched infrastructure. | Under review |
| Offset | Verified carbon removal for what remains, with a 10× safety margin, published here. | Planned |
7. Every month so far
| Month | Energy (kWh) | Carbon (kg CO₂e) | Renewable-matched |
|---|---|---|---|
| September 2026 | 0.75 (0.24–1.56) | 0.38 (0.12–0.79) | 9% |
Updated every month
These figures come from the same monthly report our AI Accountable Owner reviews, and the same archived files. If we change a factor or find an error, we correct every affected month and say so here. Questions about the method: team@merdial.com.Sources
- Google, “Measuring the environmental impact of delivering AI at Google scale” (2025): arXiv 2508.15734
- Epoch AI, “How much energy does ChatGPT use?” (2025): epoch.ai
- Energy per token in LLM inference: arXiv 2603.20224
- Ember, Global Electricity Review 2025: ember-energy.org
- Amazon Sustainability Report 2024; Microsoft datacenter sustainability commitments.