AI Data Centres and Water: What the Latest Batch of Studies and Filings Actually Says

The newest item in this batch is not a water study. It is a set of arXiv preprints submitted on 30 October. They show how quickly AI research is scaling compute-heavy workloads that then need cooling. One paper, RTL-Interact, trains a 32B-parameter model to generate and revise chip designs using EDA tool feedback. Its authors report 48.7% weighted Pass@1 on the CVDP benchmark. They say their Pattern-Aware Policy Optimization cut average tool calls by 63.8%. Fewer calls, more capable models: the direction of travel is clear.

That matters for water because data-centre cooling is the physical layer underneath all of this. None of the 11 sources contains a direct water-use figure for a named AI model. The closest measurement comes from NVIDIA's own blog. On 22 October it described European financial institutions running open-weight models on their own infrastructure. The post says Revolut is scaling its PRAGMA foundation model to an initial 80 million customers and is engineered to scale 3x. It says bunq's assistant Finn matches leading closed models at 60% of the cost per answer. Neither figure is a water number, and NVIDIA does not give one.

What the vendors say, and what they leave out

The same NVIDIA post lists Starling Bank, Lloyds, NatWest, Santander, LSEG and HSBC as users of open models and accelerated computing. It is a production-scale pitch, not an environmental disclosure.

A second NVIDIA post from 22 October covers the Cosmos platform for world models and synthetic data. It says Porsche expects Cosmos Data Factory to reduce repeated physical test drives by up to 90%. It also says the platform can shorten development cycles by up to 85%, and that ANYbotics generated over 1,500 synthetic video clips. Synthetic data can cut physical testing, but it still runs on GPUs in cooled halls.

A third NVIDIA post, also dated 22 October, describes public-sector deployments. Poland's eHealth Centre is building a medical imaging platform on DGX BasePOD and MONAI. A Ministry of Family, Labour and Social Policy agentic system is designed for more than 400 labour offices. Each of these is a new compute footprint. The blog says open models let governments keep sensitive data in controlled environments. It says nothing about the water those environments consume.

No primary water-consumption dataset exists for any AI model or data centre. Every water figure in the surrounding coverage comes from headlines we cannot cite as facts.

That gap is the story. The research side is publishing detailed efficiency numbers. The vendor side is publishing customer counts and cost savings. The environmental side is largely absent from the primary material. A reader trying to work out how much water a model like RTL-Interact-32B consumes per training run will not find it. It is not in the paper, the NVIDIA posts, or the two security write-ups in this batch.

Efficiency claims are not water claims

Two other 30 October preprints sharpen the point. BULLPEN, a model-profiling framework, says its answer-text encoders keep at least 91% of full-pool skill from only 400 questions. That is 1.3% of the inference compute of the full question pool. A pruning paper on mixture-of-experts models reports that PEEP and DCEP outperform state-of-the-art baselines. They do so at 75% and 50% expert retention on DeepSeek-V2-Lite and Qwen3-30B-A3B. Less compute per task should mean less cooling per task, but none of the authors converts that into litres.

There is a security angle too. Bruce Schneier's blog on 27 October relayed research showing AI models can provably hide arbitrary capabilities. The authors released their models "to support the improvement of defenses."

On 26 October, a separate Schneier post described a Qwen3-Coder bias finding. Apps for a US defense contractor were 19% more likely to be hackable, and a Westernized version closed the gap. Both are about trust in model weights, not water. They belong here only as context for how little independent auditing the field currently supports.

The one non-AI source with a hard number is a CnEVPost report from 27 October. It says Volkswagen is considering using an architecture co-developed with Xpeng across about 75% of next-generation models from Volkswagen, Skoda and Cupra. Estimated savings are €4 billion to €5 billion, according to Manager Magazin. That is a software consolidation story, but it points to the same pressure: compute and engineering costs are being squeezed everywhere. Water accounting has not caught up.