NextFin News - Google DeepMind released WeatherNext 3 on Thursday, an AI weather model that refreshes wind and solar forecasts every hour from satellite imagery instead of waiting for the six-hourly government data cycles that constrain traditional forecasting systems. The upgrade is aimed squarely at intraday electricity markets, where traders and utilities price renewable output hour by hour and where a forecast error discovered late is settled at the system's marginal balancing price, which can swing from deeply negative to strongly positive within a single hour.
The release marks the latest escalation in AI weather forecasting, a field where Google has moved from research curiosity to operational infrastructure in under two years. WeatherNext 3 does not claim to be the most accurate model in every horizon — DeepMind describes it as comparable with other major models on that dimension. What it claims instead is something the power market values more highly in the hours before delivery: a fresh forecast when everyone else is still working with one that is up to six hours old.
The News: Hourly Refresh, No More Six-Hour Wait
WeatherNext 3 forecasts how hard the wind blows at turbine height and how much sunlight reaches a solar farm, then updates those projections hourly from satellite imagery. DeepMind says the new model generates detailed forecasts comparable with other major weather models, but produces them faster and more frequently because it is no longer limited by the six-hourly data cycles that government-run numerical weather prediction systems rely on.
The lineage is short and steep. In November 2025, DeepMind and Google Research introduced WeatherNext 2, which the company described as its "most advanced and efficient" forecasting model, capable of hourly-resolution forecasts out to 15 days and running eight times faster than the prior generation. Each prediction takes less than a minute on a single TPU, compared with hours on a supercomputer using physics-based models, and the system produces hundreds of possible weather outcomes from a single starting point. Overall, WeatherNext 2 surpassed the previous WeatherNext model on 99.9% of variables and lead times from zero to 15 days, Google said.
In August 2026, the lab reported in Nature that its cyclone-focused WeatherNext model delivered three-day forecasts as accurate as what earlier models could manage only for the next two days — an improvement the team equated with roughly a decade of meteorological progress. The same system scaled its ensemble from 50 scenarios a year earlier to 1,000 during the most recent hurricane season, capturing rare but consequential events such as rapid intensification. WeatherNext 3 applies that same philosophy — fast, frequent, ensemble-aware — to the variables that settle a power trader's P&L.
Google Cloud has been explicit about the energy use case. In a post authored by Nelly Wilson, head of AI services for weather and climate at Google Cloud, and Matt Rettagliata, head of energy solutions, the company described WeatherNext Graph as providing deterministic forecasts up to 10 days ahead at six-hour resolution, and WeatherNext Gen as producing ensemble forecasts with up to 50 scenarios over a 15-day horizon. The company said the system produces more accurate results than previous state-of-the-art models on more than 90% of variables and timeframes, and that AI agents built on those forecasts can automatically optimize storage schedules, adjust industrial consumption for demand response, and trigger participation in energy markets.
Why the Intraday Market Is the Real Prize
To understand why hourly matters, start with how European power is traded. Generators and suppliers first commit volumes in the day-ahead auction, then adjust positions in the intraday market as delivery approaches, and finally face the balancing mechanism, where the system operator buys or sells the difference between what was promised and what was actually delivered. The closer to delivery a forecast error is discovered, the fewer options a trader has to fix it, and the more expensive the correction becomes.
Renewable forecast errors are the primary driver of intraday price swings, energy market analysts have noted. The mechanism is not the absolute level of wind or sun — a consistently sunny day is easy to price — but the revision. Energy market analysts put it plainly:
It's the changes that drive volatility rather than volume.
When output deviates from what the market expected, balance-responsible parties must scramble to rebalance their portfolios in the final hours before delivery. That scramble is where volatility lives, and where an hourly forecast earns its keep.
The market being chased is growing faster than the market it orbits. Nord Pool, Europe's largest power exchange, recorded 194.25 TWh traded through its intraday offering in 2025 — 181.40 TWh on continuous trading and 12.85 TWh on intraday auctions — an increase of more than 60% from 2024. Day-ahead volumes across Nord Pool's Nordic, Baltic, UK, and Western and Central European markets reached 1,037.21 TWh. Intraday is still roughly one-fifth the size of day-ahead, but it is expanding at several times the pace.
Market design is moving in the same direction. The 15-minute market time unit was implemented for intraday trading across Europe from March 2025 and for day-ahead trading from October 2025, and the intraday cross-zonal gate closure was cut from 60 to 30 minutes at 14 bidding-zone borders in January 2026, according to ENTSO-E. Europe is rewiring its short-term power markets for finer decision windows; WeatherNext 3 is built for exactly that cadence.
The Mechanism: From Six-Hourly Snapshots to a Continuous Signal
The binding constraint WeatherNext 3 removes is not accuracy but latency and cadence. Traditional numerical weather prediction runs on fixed cycles: global models ingest observations, solve atmospheric equations on supercomputers, and publish new guidance roughly every six hours. Between those cycles, traders work with a forecast that is, by definition, up to six hours stale, and they cannot see a fresh satellite-derived update until the next cycle completes.
AI weather models change the cost curve. Once trained, generating a forecast is a matter of running a neural network on a TPU rather than solving physics equations across a global grid. DeepMind's GenCast produces a 15-day ensemble forecast in eight minutes on a single TPU v5, with every member generated in parallel; traditional physics-based ensembles such as ECMWF's ENS take hours on supercomputers with tens of thousands of processors. The hourly refresh from satellite imagery means WeatherNext 3 can see cloud cover and wind patterns developing between the traditional cycles and reprice them immediately.
That matters most in two situations. First, rapid cloud events — a fast-moving cloud band over a solar cluster in southern Europe — can cut photovoltaic output by hundreds of megawatts within 15 minutes, a move that day-ahead models cannot resolve. Second, wind ramps: a gust front or a lull that arrives an hour earlier than expected can swing a wind portfolio's position from long to short inside a single intraday session. In both cases, the value is not in predicting the weather a week out; it is in seeing the revision before the rest of the market does.
The economic logic is straightforward. A trader who sees a solar shortfall 30 minutes earlier than competitors can buy replacement power before the intraday price gaps higher, or avoid selling into a market that is about to turn short. A trader still relying on a six-hour-old forecast is effectively blind during the window when prices move fastest. The second-order effect is where the story widens: as more participants act on hourly forecasts, intraday bids tighten around the true renewable output, the dispersion of imbalance prices narrows, and the integration cost of variable generation falls. Cheaper integration, in turn, raises the economically viable penetration of wind and solar. An hourly weather model is, in that chain, not just a trading tool but a marginal enabler of the energy transition.
Cyclical Weather, Structural Shift in Forecasting
Is this a cyclical improvement or a structural shift? The weather itself is cyclical — wind droughts end, cloud cover passes, and forecast errors mean-revert over time. But the cadence at which the market ingests weather intelligence is structural. Moving from a six-hourly batch update to an hourly, satellite-driven refresh is a regime change in the information pipeline, not a temporary edge. It will not revert on its own, because the constraint it removes — the government data cycle — was institutional, not meteorological.
Two forces reinforce that judgment. The first is renewable penetration: as wind and solar claim a larger share of generation, the same absolute forecast error moves a larger fraction of the supply stack, so the marginal value of a better intraday forecast rises with renewable share rather than falling. The second is market design: Europe's shift to 15-minute trading intervals and shorter intraday gate closures means the market is paying for precision at finer time granularity. A model that only speaks in six-hour blocks becomes less useful as the market speaks in 15-minute blocks.
The counter-thesis carries weight and is backed by Google's own framing. In its energy blog, the company acknowledges that traditional forecasting still faces "the challenge of predicting extreme events that fall outside historical patterns," and that AI models learn from historical data. Skeptics argue that AI weather models can behave like black boxes and struggle with unprecedented extremes in a warming climate — precisely the conditions under which power markets pay the most for accurate forecasts. Traditional numerical weather prediction, with its physics-based equations, remains more interpretable and more trusted by system operators when the grid is stressed. On that view, WeatherNext 3 is an incremental tool for normal conditions, not a replacement for the physics models that keep the lights on.
That critique holds at the longest lead times, where deep atmospheric uncertainty dominates. It is weaker in the nowcasting-to-intraday window that WeatherNext 3 targets. In that horizon, the relevant skill is not simulating a week of atmospheric dynamics but detecting and extrapolating observed cloud and wind fields — a pattern-recognition task where satellite-trained neural networks have a natural advantage. The falsifying signal is specific: if intraday price volatility attributable to renewable forecast revisions does not decline in markets that adopt hourly AI refreshes, or if system operators continue to rely on six-hourly guidance for balancing decisions through 2027, then the structural claim is wrong and the model is merely a faster way to produce the same information.
The Vendor Economics: Who Sells the Forecast
WeatherNext 3 also lands in a crowded and consolidating market for energy weather intelligence. Specialized vendors have long sold high-resolution wind and solar forecasts to traders and grid operators, with some offering one-kilometer spatial resolution and 15-minute temporal resolution for offshore wind. ECMWF now runs its own AI Forecasting System alongside its physics models, and NVIDIA offers its Earth-2 platform for weather and climate simulation. Research-driven groups provide open solar-nowcasting tools.
The strategic question is whether hourly, satellite-refresh forecasting becomes a widely available cloud utility or remains a differentiated product. Google's distribution advantage is real: WeatherNext technology already powers forecasts in Google Search, Gemini, Pixel Weather, and Google Maps Platform's Weather API, and the model data is accessible through Earth Engine and BigQuery, with an early-access program on Vertex AI. If hourly AI weather becomes cheap and ubiquitous, the informational rent that specialized weather vendors extracted from power traders compresses. That is the same dynamic that followed every previous widening of data access in markets: the alpha migrates from who has the data to who acts on it fastest, and the middlemen who sold the data alone must move up the stack into execution and advisory.
Who Benefits, Who Is Exposed
The immediate beneficiaries are the players whose P&L is most sensitive to intraday weather revisions: renewable generators with large wind and solar portfolios, energy traders running intraday books, and balance-responsible parties that face imbalance charges. Battery storage operators are a second-order winner: more accurate hour-ahead renewable forecasts let storage charge and discharge against tighter, more confident price spreads, and let them offer balancing services with less uncertainty about their state of charge at delivery. Google Cloud's own energy playbook — automated storage scheduling, demand-response adjustments, and market participation triggered by AI agents — is a roadmap for how asset owners will monetize the signal.
The exposed are the intermediaries whose edge came from owning better weather data than their counterparties. If hourly satellite-refresh forecasts become widely available through cloud platforms, the pricing power of specialized weather vendors compresses unless they move into execution, advisory, or asset optimization. Grid operators and regulators face a different exposure: as AI forecasts shape bidding behavior, system operators must decide whether to ingest AI guidance directly into balancing mechanisms or keep it at arm's length as a commercial tool.
Three signals are worth watching. In the short term, watch whether Google announces named utility or trader pilots for WeatherNext 3 — adoption by a major European generator would confirm the commercial case. Over the medium term, watch intraday volume growth and imbalance-price dispersion in markets with high renewable penetration: if hourly AI forecasts are doing their job, dispersion should narrow even as volume rises. Over the long term, watch whether system operators begin to ingest AI weather guidance directly into balancing mechanisms — that would mark the transition from a trading tool to grid infrastructure.
The base case is that hourly AI weather becomes standard plumbing for intraday power trading within a few years, with the biggest gains captured by asset owners who integrate it into automated bidding rather than using it as a dashboard. The upside case is that sub-hourly, satellite-driven forecasting enables materially higher renewable penetration by reducing the balancing cost that currently caps variable generation. The downside case is that model risk in extreme weather keeps system operators anchored to physics-based guidance, limiting WeatherNext 3 to a trading edge rather than a grid upgrade.
DeepMind's hourly model does not predict the weather better so much as it delivers the prediction when the market can still act on it. In power trading, timing is not a refinement of the forecast — it is the forecast.
Explore more exclusive insights at nextfin.ai.
