Most of the conversation around AI and climate is about what AI can do for the energy transition. This episode looks at the other side: the footprint of AI itself, and the engineering choices that make a prompt, a model or an agent workflow cheaper to run.
AI x Climate Tech is a podcast series from Enjins for CTOs, Heads of Data and senior engineers in climate and energy. Each episode is a 30 to 45 minute conversation about concrete system design: the architectures, the trade-offs, and what only becomes visible once something is running in production. That engineering layer is where the two fields meet, and it gets less airtime than it deserves.
Episode 3: the energy footprint of AI itself
For episode three we spoke with Robert Keus, founder of GreenPT. GreenPT hosts open-source models on energy-efficient infrastructure and reports energy use and CO2 back to its customers. On top of that sits a platform layer that decides how much compute a task actually needs.
In this episode we covered:
- How Robert got here, and why GreenPT exists
- The gap between EU and US infrastructure build-out
- Lowering emissions through input compression and output instructions
- The largest source of compute waste in AI systems
- Winning customers with a sustainability story
TL;DR Recap of Episode 3
- GreenPT started from two client questions Robert’s agency could not answer: where does my data go, and what is the footprint of this? The first DeepSeek release made self-hosting realistic, and the search for a transparent hosting provider turned into a company.
- The platform orchestrates between small, medium and large language models, so a task does not automatically land on the biggest model available.
- Input compression removes everything in a prompt that is not needed, without using AI to do it. Output instructions push the model toward concise answers and away from formatting nobody asked for.
- Agent-to-agent traffic is still written for human readers who never show up. Transcripts between agents read like prose, and JSON is structured partly so a person can scan it.
- Depending on the task, these techniques save 20 to 30 percent in tokens. GreenPT has open-sourced a package that applies the same logic inside Claude Code.
- Compression is switched off where full context matters, such as authentication work and security audits.
- Two arguments land commercially: FinOps, because AI spend grows every month, and reporting, because large companies have CO2 figures to publish.
Robert has been an entrepreneur since he was fourteen, starting with computer repairs and later building the internet agency Rise Brothers. The agency already worked on accessibility and sustainability, and in recent years that work moved toward AI features, mostly speech-to-text at the time.
Then clients started asking where their data went and what the footprint of these features was. The team could not answer either question.
The catalyst was the first DeepSeek model, when open-source quality reached a level businesses could actually use. If you can host the model yourself, you control where it runs. Robert and co-founder Kast went looking for hosting providers that were transparent about water use, energy use and CO2 emissions, found one that was open to the idea, and started measuring.
The infrastructure layer: where the model runs
Water is the one number GreenPT does not report. Measuring the water use of an entire data center is difficult, and Robert treats it as a political question more than a technical one: you can decide to build a new data center with techniques that avoid water altogether, which takes time but is largely a solved problem.
Energy is a longer story. The NVIDIA chips the industry runs on were designed as graphics cards for gaming, and processors built specifically for AI are only now arriving from parties like OpenAI and Google. Meanwhile the EU regulates how data centers get built, while operators in the US put gas turbines next to the large ones.
Location becomes a design decision once you own the training compute. Are the GPUs in France, where emissions are low, or in Poland, where they are high? Robert expects fine-tuning to become accessible to anyone with technical knowledge of their own data, which is pushing GreenPT toward the GPU layer underneath the models.
The model layer: fewer tokens per answer
Inside the platform, GreenPT applies two techniques around the model call. Input compression removes everything in the prompt that is not strictly needed, including filler words. Output instructions ask the model to stay concise and to skip tables where a table adds nothing.
The same behaviour is moving into the models themselves. GreenPT recently released two fine-tuned models, Qwen 9B and 27B, where the concise-output behaviour sits in the training layer rather than the prompt. Both were tuned for coding tasks, the use case with the clearest evidence today.
Developers using the raw API get the model untouched, with the efficiency tooling available as something they switch on themselves.
The agent layer: how machines talk to each other
The larger opportunity has moved past the single prompt. Systems are agentic now, with models talking to other models inside an orchestration.
Read the transcript between two agents and it reads like prose. It is written as if a human were asking something of another human, with styling and layout that make the exchange heavier and less clear for the agent receiving it. JSON has a similar origin, structured in part so a person can read it easily.
“There are different ways to bring data to an agent which are far more token efficient.”
Where no human reads along, GreenPT gives different instructions and leans on more efficient structured output. Across tasks that adds up to 20 to 30 percent in token efficiency, and the open-source package makes the same savings available inside Claude Code.
There is a deliberate limit. Push compression too far and answer quality drops, so authentication work and security audits run on full context.
Why customers buy it
Two arguments do the work today. The first is FinOps: companies watch their AI spend climb every month while open-source models such as Kimi 3 and GLM keep improving. The second is reporting: the large enterprises in the pipeline have CO2 figures to publish and want data autonomy alongside them.
Model quality remains the entry ticket. In the early days GreenPT lacked the larger models and lost deals because of it.
