AI Water Use Data Goes Missing: EU Complaint, Local Revolts and the Model Race

The complaint, first reported by Politico on 30 September and picked up by The Next Web the same day, argues that Brussels has failed to publish the environmental data it collects from data centre operators. The dossier does not name the complainants or quantify the missing figures, and neither outlet published the filing itself. What both reports agree on is the central claim: the EU holds the numbers and is not releasing them.

That is an awkward position for a bloc that has spent a decade positioning itself as the world's most ambitious tech regulator.

A model problem, not just a building problem

The timing matters because the water question is shifting from the physical plant to the model itself. Two papers posted to arXiv on 29 and 30 September illustrate how the research community is now measuring the cost of inference at a granular level. One, from Gaurav Agarwal and co-authors, describes Decode-Latency Feedback Prefill, a controller implemented in vLLM that resizes prefill chunks based on observed decode latency. On Qwen3-0.6B running in BF16 on a single A100 80GB GPU, three paired 100-request trials cut P99 inter-token latency by 24.8%, 30.1% and 28.2%, a mean of 27.7%. The trade-off is explicit: mean P99 time to first token rose 34.8%. The authors also report a negative result, noting the mechanism does not generalise to Qwen3-8B, Qwen3-32B or a two-GPU tensor-parallel setup.

Efficiency research like this is the unglamorous half of the water debate. It does not produce a headline number, but it determines how much work a given rack does per kilowatt-hour, and therefore how much cooling and water a site needs.

The other paper, from Loay Mualem and six co-authors, introduces GoldiMask, a fine-tuning method for discrete diffusion language models that selects which tokens to reveal as context by approximately maximising a submodular objective. Across three backbones and three training datasets, the authors report the highest average accuracy in most evaluated settings, with gains on reasoning and code generation. They also claim a reduction in decoding iterations on GSM8K and MATH-500 under confidence-threshold parallel decoding.

Neither paper mentions water. Both are part of the same arithmetic that data centre operators and their critics are arguing over.

Fourteen hundred megawatts and a growing queue of objections

India's installed data centre capacity has crossed 1,575 MW and could reach 9 GW by 2030, according to PCC figures reported by Fortune India on 1 October. The same report flags rising power and water concerns. Karnataka's new state data centre policy puts water and energy at the centre of its approval regime, according to Newslaundry, also on 1 October.

Indonesia has already acted. A data centre project in West Java was suspended over water concerns, Asia News Network reported on 1 October.

In the United States, the pattern is local rather than national. A group in Weld County, Colorado, is collecting petition signatures to block a Global AI data centre, according to KUNC News on 1 October. Residents in Adelanto and the High Desert opposed data centre amendments, ie community news reported on 30 September. A township approved a one-year moratorium on data centres, Village Report reported on 30 September.

"A handful of bad actors have done real damage here," IT Pro reported on 30 September, in a piece on how businesses can avoid disruption from the wider data centre backlash.

The industry has started to respond collectively. An AI Infrastructure Coalition pledged to cover power bills and cut water use, Dealroom reported on 30 September, framing the move as midterm backlash politics. Financial Times reported the same day that the AI industry is moving to thwart data centre backlash ahead of the US midterms. Congress is also moving: financialexpress.com reported on 30 September that legislation would put data centre power costs on the operators themselves.

None of this has yet produced a federal US standard on water use. The result is a patchwork of county petitions, state policies and voluntary pledges, which is precisely the environment in which a complaint about missing EU data becomes harder to wave away.

What the model builders are shipping

Meanwhile the model releases keep coming, and each one raises the baseline of compute that data centres are expected to serve. Alphabet launched Gemini 4 Argon on Wednesday, with Google describing it as its most advanced model yet. CNBC reported that the model sets a new record in real-world software engineering, ties for first in cybersecurity, and leads another benchmark covering finance, legal and other professional tasks. Google said Argon is already being used internally to optimise memory at its data centres, freeing hundreds of terabytes of memory without buying additional hardware.

That internal detail is the most operationally interesting number in the launch. The Guardian reported on 1 October that Google is withholding the model from the public for now, releasing it only to a vetted group of cybersecurity experts. Koray Kavukcuoglu, Google's chief AI architect, wrote in a blogpost that "safely releasing frontier capabilities at this level requires a phased approach." A day earlier, Sundar Pichai signed a voluntary AI safety accord at the White House alongside President Donald Trump and other tech executives, the Guardian reported.

On the smaller end, Amazon Web Services released Strands Decider 2B, an open source decision model inspired by TypeSafe's Jev, TechCrunch reported on 1 October. Amazon distinguished engineer Marc Brooker told TechCrunch the class of models makes "a perfect decider for a workflow step." The model is small enough to run locally. Ideogram, meanwhile, released Ideogram 4.5, an image editing model that the company claims only touches the parts of an image the user specifies, with four quality tiers from 0.8 to 22 cents per image at native 2K resolution, according to The Decoder on 1 October.

Small local models and large frontier models pull in opposite directions on infrastructure. A 2B model that runs on a laptop does not need a cooling tower. A frontier model that optimises memory across a fleet does.

The governance gap

Florida Attorney General James Uthmeier filed a motion for a temporary injunction against five OpenAI entities and Sam Altman personally, Tom's Hardware reported on 30 September. The motion asks the court to bar OpenAI from developing any AI models without independent third-party guardrails and approval, among other enjoinments. It is part of a lawsuit filed in June in Highlands County.

That is a safety argument, not a water argument. But it points at the same absence: no regulator in the dossier has published a binding, standardised disclosure regime for the resources an AI model consumes, from training run to inference request to the cooling water that carries the heat away.

The EU complaint, if it proceeds, would test whether the bloc's existing data collection powers are enough. The Indian and Indonesian cases suggest that in the absence of such disclosure, the fight moves to the local planning committee, where the numbers are whatever the developer chooses to put in the application.

Data Center Knowledge published two pieces in late September on the technical side of that gap. One, dated 23 September, argued for moving from efficiency metrics to real-world resilience. The other, dated 1 October, described climate as the hidden variable in every adiabatic water budget. Both are behind specialist audiences, which is part of the problem: the people with the models and the people with the planning votes are not reading the same documents.

What the dossier does not contain is a single agreed figure for how much water an AI query uses. That number is contested, and until the disclosure fight resolves, it will stay contested.