AI on renewable power.

Where your prompts are answered

All AI inference for Union AI runs on Scaleway Generative APIs, inside the European Union. Scaleway keeps every one of its AI servers in a single data centre, DC5, which it built for this kind of load. So the figures below are not an average across a mixed fleet: they are the figures for the building your prompts actually go to.

What the numbers mean

PUE (power usage effectiveness) is the total energy a data centre uses divided by what the servers themselves use. At 1.25, every kWh of computing costs 0.25 kWh extra for cooling and power delivery. At the global average of 1.55 that overhead is 0.55 kWh, more than twice as much.

Cooling without air conditioning is where most of that difference comes from. The AI data centre blows filtered outside air through the servers and only adds evaporative (adiabatic) cooling when it is warmer than 28°C outside. Scaleway puts the saving on cooling energy at 30 to 40% compared with the rest of the market.

WUE (water usage effectiveness) is litres of water per kWh. Most data centres use cooling towers, which evaporate large volumes all year round. This one does not, and only uses water on hot days.

You do not have to take this from us. The operator publishes the PUE of the AI data centre live, in real time, with the trailing twelve-month figure next to it.

What we add on top

The data centre is Scaleway's work, not ours. What we control is how much computing each answer takes, and there we made three choices.

• No idle GPUs. We run serverless: the models sit on shared hardware that serves many customers, and we pay per token. There is no dedicated graphics card waiting, powered on, for our next request.
• Models that only wake the part they need. Our models are mixture-of-experts models, which use only a slice of their weights per word. Standard, the everyday tier, activates about 3 billion of its 35 billion parameters for each token.
• Heavy only when it is needed. Everyday questions, summaries and rewrites go to the lighter tiers. Pro, the largest model, is there for the hard problems, not for every message. Which model sits behind each tier is on the AI transparency page.

Methodology: energy per answer

Under every answer in the chat you see a small label like ≈ 0.2 Wh. That is an estimate, not a measurement. Scaleway, which runs the models for us, does not report energy per request, so we work it out from what we do know: how many tokens the answer used, which tier answered, and how efficient the data centre is.

Everything you need to check it is below: the formula, every figure, a worked example and the sources. Where we are unsure we round up, so if the estimate is wrong, it should be wrong on the high side. Method version 1, reviewed 28 September 2026.

The formula

• Tokens are the counts the model provider reports for the answer, the same numbers we meter for billing. Nothing is guessed from the length of the text. A token is roughly three quarters of a word.
• The input and output figures are fixed per tier, in watt-hours per 1,000 tokens, for the graphics card and the rest of the server. They are in the table below.
• 1.25 is the PUE of the AI data centre (Scaleway, 2024 data). It adds the energy for cooling and power delivery on top of the servers.
• 51.9 g CO₂e per kWh is the average for the electricity grid the data centre draws from, over the whole life cycle (ADEME Base Empreinte, electricity 2024).

Where the figures per tier come from

Nobody publishes measured energy for the exact models behind our tiers, on the hardware and load Scaleway runs them at. So each figure is built in four steps from the closest published measurements.

1. What the graphics card uses per word it writes. From the ML.ENERGY Leaderboard, which measures GPU energy per output token on NVIDIA H100 cards. We take the figure for a moderately busy server (batch size 64); busy production servers use 2 to 5 times less per token. Where a model has not been measured, we use 0.10 joules per billion active parameters, the top of the range that similar models show.
2. The rest of the server. A GPU figure leaves out the processor, the memory and machines kept ready but idle. In Google's measured breakdown those add about 57% on top, so we multiply by 1.6.
3. Reading your prompt. Input is read in parallel and costs far less per token than writing the answer: measurements put it between 1:10 and 1:35. We count an input token as one fifth of an output token, which keeps a margin of at least two times.
4. The worst case per tier. If a tier's model is unavailable, a fallback model answers. Each tier carries the highest figure of its model and its fallbacks, rounded up, so a fallback never makes the estimate too low.

The figures per tier

The input and output columns are for the server alone, before the data centre overhead. The last two columns are one medium message of 900 tokens in and 500 out, with the overhead included.

A tier's figure is the highest of its own model and its fallbacks, so two tiers can show the same number. Heavier tiers use more because their models activate more parameters for every word they write.

Worked example: one medium message on Standard

A medium message is 900 tokens in (your question plus the conversation so far) and 500 tokens out.

In the chat that shows as ≈ 0.2 Wh and ≈ 0.008 g CO₂e. The same message on Pro, the heaviest tier, comes to about 1.7 Wh.

Does that make sense?

• Epoch AI estimates about 0.3 Wh for a ChatGPT (GPT-4o) query with 500 output tokens, and calls that figure fairly pessimistic.
• Google measured a median of 0.24 Wh per Gemini text prompt, everything included.
• Microsoft researchers put the median for large models on H100 servers at 0.31 Wh per query.

Standard activates about 3 billion parameters per word, so landing below those figures is what you would expect. Pro activates far more and lands above them: that is where our rounding up shows most.

Carbon: two statements, kept apart

1. The electricity is 100% renewable. Scaleway supplies the data centre with renewable electricity backed by Guarantees of Origin. A Guarantee of Origin is a certificate that an equal amount of renewable power was fed into the grid, not a cable from a wind farm to the building.
2. The CO₂e figure uses the grid average anyway. Counting the certificates would put it near zero. We show what the same electricity would emit at the average of the grid the data centre is connected to: 51.9 g CO₂e per kWh over the whole life cycle. That is higher than the grid operator's own figure, so it errs on the high side.

What the estimate does not cover

• No meter per request. Every figure is an estimate from published measurements of similar models on similar hardware.
• Shared hardware. The models run on graphics cards shared with other customers. How busy they are changes the energy per token by 5 to 10 times, and we do not know Scaleway's load or exact card types. We assume a moderate load on H100-class cards.
• Newer models. None of the models behind our tiers has a published energy measurement yet, so their figures come from their predecessors or from their size.
• Very long prompts. The figure per input token is flat. A prompt of hundreds of thousands of tokens costs more per token than that, so for such a prompt on Long context the estimate is likely low.
• Training and manufacturing. We run open-weight models that others trained, and we do not know what that training used. Making the servers, the cards and the building is not in the figure either.
• Only the model. Our website, database, payments and your own device are not included.

When Scaleway publishes energy per request, or better measurements of these models appear, we replace the estimate and raise the method version.

Where you see it

• Under each answer in the chat, with the CO₂e and a link to this section.
• For the whole conversation, at the top of the chat.
• Per month, in Billing: every chat answer and API call since the 1st.

Sources for the estimate

Every figure in this section comes from one of these:

• ML.ENERGY Leaderboard v3.0, H100 energy per output token (data of 2026-02-16)
• Chung et al., Where Do the Joules Go? Diagnosing Inference Energy Consumption (2026)
• Elsworth et al., Measuring the environmental impact of delivering AI at Google scale (2025)
• Epoch AI, How much energy does ChatGPT use? (2025)
• Oviedo et al., Energy Use of AI Inference, Efficiency Pathways, and Test-Time Scaling (Joule, 2026)
• Delavande, Pierrard and Luccioni, Small Talk, Big Impact: The Energy Cost of Thanking AI (2026)
• Scaleway Impact Report 2025 (PDF)
• ADEME Base Empreinte, electricity 2024, average consumption mix

What we do and do not claim

• ✓True: Every AI inference runs in a data centre on 100% renewable electricity, inside the European Union.
• ✓True: That data centre has a PUE of 1.25 and is cooled without air conditioning (Scaleway, 2024 data).
• ✓True: Every chat answer shows an estimate of its energy use and CO₂e, worked out with the published method above.
• ✕Not true: “Carbon-neutral AI.” Renewable supply is backed by Guarantees of Origin, not by a wind farm wired to the building, and manufacturing the servers has a footprint of its own.
• ✕Not true: “The models were trained on green energy.” We run open-weight models that others trained. We cannot verify how that training was powered, so we do not claim it.
• ✕Not true: “Our whole stack is green.” These figures cover inference. Website hosting, the database and payments run at the other parties on our subprocessors page, under their own numbers.
• ✕Not true: “Measured energy per message.” The figure under each answer is an estimate from published measurements, rounded up, not a meter reading.

Sources

All data centre figures are Scaleway's own, for 2024, from the Scaleway Impact Report 2025 (PDF). The global PUE average of 1.55 is the Uptime Institute figure that report cites. Why European hosting matters beyond energy is in European sovereign AI.