Two things happened in August 2026, twenty days apart.
On August 6, Google released a cyclone forecasting model built with the US National Hurricane Center and the UK Met Office and gave the code away. It buys forecasters roughly an extra day of warning. In this field, a day is about a decade of ordinary progress.
On August 26, a piece of glacier broke off above the Trishuli valley in Nepal and fell more than a kilometre. The water that followed killed over 1,400 people in Nepal and Tibet. Four monitoring stations in its path were destroyed by the flood they existed to detect. The last gauge left standing never turned amber: it recorded the river at 6.57 metres against a 7.3-metre alarm threshold, because it had been built to ask how high the water was, not how fast it was rising.
Twenty days apart: the smartest weather machine ever built, and a valley with nothing left to measure it.
That gap is the real story of artificial intelligence (AI) and climate, and almost nobody is telling it.
Where the record runs out
In September, Nature Communications published the arithmetic. Manuel Linsenmeier and Jeffrey Shrader compared forecast accuracy across the world and found that a seven-day forecast in a rich country beats a one-day forecast in a poor one. A farmer in Bavaria can plan a week ahead with more confidence than a farmer in Chad can plan tomorrow. About two-thirds of that gap is geography, since tropical weather is harder to predict. The rest is a choice: poorer countries have fewer instruments, and the ones they own report less often.
AI was supposed to close that gap. It is widening it.
The reason is simple enough. The new forecasting models were never taught physics. They learn by reading decades of recorded weather and copying the patterns they find. Where the record is thick, they learn beautifully. Where nobody was measuring, there is nothing to learn from, and the model does what any student does with a blank page. It guesses, fluently.
Kenyan researchers tested GraphCast, one of the leading models, over Africa. Rainfall and temperature errors came out roughly twice those over Europe, on a continent where 95 per cent of farming depends on rain. When the European Centre for Medium-Range Weather Forecasts audited its own AI system this year, it found the same shape: the gains are real, and they fall most generously on places already rich and already well observed.
The instrument map explains it. Germany alone has more stations meeting the basic international standard than the whole of Africa. Across the continent, only about one in ten required surface stations reports as global models need, and one in five balloon stations. In October 2024, Chad was reporting three per cent. Small island states run multi-hazard warning systems in only 43 per cent of cases, and disaster mortality is nearly six times lower where such systems are strong.
How little of the world’s weather is actually being watched, and who is left without a warning
The one input that cannot be rebuilt
Now the argument missing from every AI debate.
Every other ingredient of AI is getting cheaper. Chips improve each year. Models can be retrained on a whim. Even the computing bottleneck is dissolving: a Cambridge team has built a forecasting system that runs on a desktop, using thousands of times less computing power than the pipelines it replaces, and beating the American national model on many variables with a tenth of the input.
So the hard part is no longer the machine. It is the measurement.
And measurement is the one input that cannot be recovered. A chip delayed is a chip bought next year. A storm that crosses the Sahel unmeasured this June is gone. It is a row missing for good from the training data of every model that comes after. We are not postponing those observations. We are destroying them, quietly, by not making them.
Worse, the loss conceals itself. Offer a finance ministry a free AI forecast and you remove the last budgetary argument for buying instruments. Remove the instruments and nothing remains to check the forecast against. The error does not vanish; it stops being visible. American scientists already warn that AI forecasting is expanding while observation budgets shrink. That logic has already reached poorer capitals. Nepal’s department of hydrology and meteorology watched its budget fall from Rs 765 million to Rs 338 million in three years. The flood of August 26 caused damage its own infrastructure minister put at Rs 200 billion.
There is an asymmetry here worth saying out loud. Under the World Meteorological Organization (WMO)’s unified data policy, countries are obliged to hand over their core observations free and unrestricted. Commercial AI forecasting is built on exactly that data. The poorest meteorological services on earth are donating the raw material for products they will never afford, and nothing flows back to the gauges.
Four fixes, none of them costly
This is fixable, and cheaply. Four measures would do most of the work.
Fund the instruments. SOFF, the UN facility created to close the observation gap in the poorest countries, is asking for $400 million by 2027 and remains about a quarter short. Global capital spending on AI in 2026 is running at some $765 billion. The entire UN target is four and a half hours of that. It should be closed at the next climate conference, not debated at it.
Make the users pay the measurers. Any commercial forecasting product trained on freely exchanged public observations should return a small fixed share of revenue to the networks that produced them. Call it what it is: a licence fee on a public good that stayed free for a century because nobody could sell it. Now someone can.
Pay for operation, not just installation. Observation networks rarely die at purchase. They die three years later, when there is no budget for a technician, a spare sensor or a balloon. Donors love ribbon-cuttings and hate maintenance, which is exactly backwards. Write the financing for decades.
Publish a scoreboard. The WMO should release forecast skill by country income group every year, in public. Inequality that is measured gets argued about. Inequality that is invisible gets inherited.
What a forecast is worth
This has been measured. India, which is buying instruments rather than hoping AI will substitute for them, has the cleanest evidence. A 2020 evaluation by the National Council of Applied Economic Research surveyed 3,965 farming households across 121 districts and found that those acting on all nine weather advisories earned Rs 3.02 lakh a year against Rs 1.98 lakh for those acting on none, a national gain of Rs 13,331 crore a year. India is adding 60 upper-air stations under Mission Mausam. The point holds anywhere: a forecast is only ever as good as the instrument behind it.
What it would cost to close the gap, and what closing it returns
Return to the Trishuli valley on that August morning.
The warning that saved the children of Tribhuvan Trishuli Secondary School did not come from a model. A head teacher took a phone call from someone upstream who had seen the water, and walked some 900 students uphill ten minutes before the school vanished under debris. No algorithm, no satellite, no supercomputer. A person who was looking, and a person who answered.
That is the argument in one morning. The machines are extraordinary and getting better. They can only ever tell us about a world somebody bothered to watch.
Every storm we fail to measure is a page torn from a book the machines have not finished reading, and no money, later, will put it back. The sky is still writing.
The question is whether anyone is still taking notes.
Views expressed are the author’s own and don’t necessarily reflect those of Down To Earth