UK Watchdog Says Departments Cannot Cost Climate Risk, as Weather Models Race Ahead

The peg is a report published on 3 November by the National Audit Office, the UK's public spending watchdog. According to The Register's account of the NAO findings, departments expect higher operating costs and energy charges from climate change and the shift to low-carbon power. Few have a solid picture of the risks or their price.

That is a narrower claim than the headlines might suggest. The NAO did not say departments are ignoring the climate. It said their reporting under Treasury rules, first required for the 2025-26 financial year, is thin on transition risk. It rarely costs anything, even approximately.

The report lands in the same week that a group of European developers showed what the modelling side now looks like in practice. The gap between the two is the story.

What the watchdog actually found

The NAO's report, linked as a PDF in The Register's piece, reviewed the first set of climate-related financial disclosures from large central government bodies. HMT set out the Treasury rules for those disclosures, and separate disclosure requirements have been in place since 2022, the report notes.

Bodies listed physical risks in more detail than transition risks. These included flooding, higher winds and longer droughts. Where they did name transition risks, the most frequent were the financial cost of the shift itself and supply chain disruption. Anticipated costs included switching to lower-carbon energy, new technology and complying with new regulations. They also included fluctuations in the price and availability of materials, water, energy and specialist skills.

Gareth Davies, head of the NAO, said: "UK public services will need to adapt to the impacts of climate change. Parliament needs clear information on the resilience of infrastructure, service provision and finances to climate change, and on how these risks are being managed." He added that the requirement on all large central government bodies to make such disclosures from 2025-26 is "an important first step". The watchdog expects the reporting "to become more useful over time."

"Bodies did not articulate transition risks as well as physical risks in the strategy disclosures. In the disclosures we reviewed, while several transition risks were listed, few specifics were given, unlike physical risks, which were explained in greater detail."

That passage from the NAO report, as quoted by The Register, is the core of the finding. The accounting is incomplete, not absent.

The modelling side is moving faster

On 22 October, NVIDIA's blog described a family of open Earth-2 models that European organisations are deploying for weather and climate prediction. The newest member, StormScope EU, is a regional version of an existing short-term forecasting model. It is trained over European weather and climate data and uses EUMETSAT satellite observations. According to the company's post, it predicts cloud conditions across 16 Meteosat channels at 10-minute intervals.

The post names specific users. AXA is using FourCastNet 3 to study regional cyclone risk beyond the historical record across more than 1,000 simulated seasons. Vaisala Xweather is combining Earth-2 with its own observations, lightning data and forecasting expertise. It is also adapting StormScope with lightning data to extend short-term forecasting lead time. University of Manchester researchers adapted the CorrDiff model for UK air-quality modelling. They trained it on the Isambard-AI supercomputer and run it locally on DGX Spark systems. TotalEnergies is evaluating StormScope-based solar-radiation nowcasts at two photovoltaic sites.

These are vendor claims and pilot projects, not peer-reviewed results. None of them is a substitute for a government department costing its own flood exposure. But they show that local, high-frequency forecasting is becoming an operational tool for insurers, energy firms and transport operators. That is exactly the kind of decision the NAO says public bodies cannot yet support with numbers.

Physics still beats the models on extremes

The case for AI weather models is cost and speed. The case against relying on them for extreme events is accuracy. Earlier work has repeatedly found that traditional numerical models outperform AI forecasts for extreme weather. Those findings are background rather than news.

Recent climate attribution work points the same way. A study reported in late October found that climate change amplified the 2024 European floods more than rainfall alone would suggest. A separate line of research has found physics-based models more accurate than AI at predicting extreme weather. The Earth-2 blog itself is careful to describe CorrDiff as a way to add local detail to coarser data, not to replace the underlying physics.

For a UK department trying to put a number on flood damage to a hospital estate or a road network, that distinction matters. A 10-minute cloud forecast is useful for a solar farm. It is not a cost estimate for a 30-year adaptation programme, and nothing in the NVIDIA material claims otherwise.

Why the numbers are hard to produce

The NAO's finding is not surprising to anyone who has tried to cost adaptation. Physical risks arrive as distributions, not point estimates. Transition costs depend on policy choices that are still being made. The report says bodies expect costs from damage to assets and infrastructure, higher maintenance and repair bills, potential asset upgrades, and increased operating costs including energy. Bodies controlling significant property or infrastructure are particularly concerned about decarbonisation and adaptation costs.

The disclosure regime is also young. The first reports under the Treasury rules covered 2025-26, and the NAO reviewed only a subset. Its conclusion that understanding is poor is a statement about a first attempt, not a final verdict.

That said, the watchdog's job is to say when public money is being committed without a clear picture of risk. On this evidence it is.

The AI research backdrop

Two arXiv papers submitted on 29 and 30 October underline how unsettled the evaluation side remains. One, from Fengxian Ji and colleagues, asks whether marketplace transaction prices can stand in for human visual preferences. It finds that only limited conditions allow it. The other, from Zhifei Hu and co-authors, evaluates whether web world models that predict the consequences of actions actually support safe decisions. It finds that plausible predictions are not enough.

Neither paper is about climate. But both are about the same failure mode the NAO describes: a plausible output is not the same as a decision-grade one. A climate risk disclosure that lists risks without costs is, in that sense, a plausible prediction.

The difference is that the NAO has a statutory audience, and the models do not.