> ## Content Index
> Fetch the complete content index at: https://www.edgewisely.com/llms.txt
> Use this file to discover other available public pages before exploring further.

# Google's Bet on Forecasting by the Hour
- URL: https://www.edgewisely.com/googles-bet-on-forecasting-by-the-hour/
- Published: 2026-09-06T04:20:55.000Z
- Updated: 2026-09-06T04:20:55.000Z
- Description: WeatherNext 3 trades stale, six-hour-old physics simulations for live satellite data, forecasting the weather every hour at five times the resolution of its predecessor.
- Author: John Karpentar
- Tags: AI, Deep Tech

**How WeatherNext 3 trades stale physics simulations for live satellite data to forecast the weather every hour, everywhere on Earth.**

Most weather forecasts you've ever seen were already six hours old by the time they reached you. On September 3, Google DeepMind and Google Research shipped a model built to close that gap to sixty minutes.

## What actually changed

WeatherNext 3 is Google's third-generation global weather model, and the pitch is specific rather than sweeping: forecasts every hour instead of every six, at roughly five times the spatial resolution of its predecessor, built from data the model can watch update in real time rather than data that's already stale by the time a supercomputer finishes processing it, according to [Google's own announcement](https://blog.google/innovation-and-ai/models-and-research/google-deepmind/introducing-weathernext-3/?ref=edgewisely.com).

The mechanism is the interesting part. Traditional numerical weather prediction and even most AI weather models train on the output of physics simulations run on government supercomputers — accurate, but carrying a six-hour lag by the time the data is packaged and released. WeatherNext 3 instead ingests raw, live geostationary satellite imagery on an hourly cadence, feeding it into what Google [describes](https://deepmind.google/science/weathernext/?ref=edgewisely.com) as a Functional Generative Network mesh transformer, alongside sparse ground-station observations and two precipitation-specific data sources: NASA's satellite-based IMERG dataset and Google's own radar-based precipitation reanalysis.

The resolution gains are concrete. Key surface variables like temperature and moisture now resolve to 5 kilometers, other surface variables to 10 kilometers, and atmospheric variables like wind to 25 kilometers — versus a flat 25-kilometer grid updated every six hours for WeatherNext 2\. On precipitation specifically, Google reports a Continuous Ranked Probability Score improvement of up to 60% against the IMERG benchmark, 30% against MRMS radar data, and 10% against rain-gauge measurements at early lead times. For forecasts a day or more out, the company says users will see up to 50% more accurate precipitation predictions, with the largest gains concentrated in regions that have historically had the worst forecasts.

The model is also roughly 2.4 times larger than WeatherNext 2 in parameter count, according to [TechCrunch's reporting](https://techcrunch.com/2026/09/03/googles-latest-ai-weather-model-gives-you-no-excuse-to-forget-your-umbrella/?ref=edgewisely.com), and for the first time it's trained to target predictions to specific weather stations — a change Brightband atmospheric scientist Daniel Rothenberg told the outlet moves the model closer to forecasting "what Denver's airport weather station is going to measure," rather than an averaged grid cell that doesn't map cleanly to anything a person or business can act on. Google published the model's architecture in a [technical paper on arXiv](https://arxiv.org/abs/2609.03582?ref=edgewisely.com), alongside the product announcement.

Google is rolling the model into Search, the Gemini app, Google Maps, the Google Maps Platform Weather API, and Earth Engine starting the day of announcement — the same app where Google recently pushed a cheaper, faster [Gemini Flash release](https://www.edgewisely.com/gemini-3-7-flash-google-launch/) into wider use — with raw forecast data also queryable in BigQuery and downloadable via Cloud Storage for developers who want to build on top of it directly. Independent validation comes from Brightband's Operational WeatherBench, a live leaderboard that Google says now ranks WeatherNext 3 ahead of comparable models from Microsoft, Nvidia and the European Centre for Medium-Range Weather Forecasting, as well as ahead of the U.S. National Weather Service's own traditional forecasts.

## Stakeholder read

**For Google's consumer products,** this is a quiet but real distribution advantage: weather is one of the most-queried categories in Search and Maps, and a materially better underlying model shows up in millions of daily interactions without Google having to market a single feature.

**For renewable energy operators,** WeatherNext 3 adds forecasts built for their specific problem — 100-meter wind speed at turbine height, plus cloud cover and solar radiation — letting grid operators match expected wind and solar output to demand more precisely than a generic temperature-and-precipitation model allows. That's the same matching problem startups like [Emerald AI have raised money to solve](https://www.edgewisely.com/emerald-ai-150-million-data-center-grid-flexibility/) from the demand side, by making data centers flex their own power draw around what the grid can actually deliver.

**For forecasters and researchers in the Global South,** the more consequential shift may be reach rather than resolution. Google frames the 5-kilometer grid as particularly valuable in Latin America, Africa and Asia-Pacific, regions that have historically been underserved by high-resolution regional models because running them requires supercomputing budgets few national weather services can afford.

**For rival AI weather ventures,** the competitive picture is more contested than Google's framing suggests. Google calls WeatherNext 3 the first model to directly incorporate raw satellite observations for high-resolution global forecasting, but the startup WindBorne Systems has said its WeatherMesh 6 model has ingested raw observations — sourced from its own fleet of weather balloons — since late 2025\. Google's response, per TechCrunch, is that its forecasts run at higher resolution globally regardless of who used raw data first. Both claims can be true at once: this is a race with more than one credible entrant, not a field Google has to itself — the same lesson that's played out in large language models, where [cheaper, non-American challengers](https://www.edgewisely.com/the-cheapest-frontier-class-model-isnt-american/) have repeatedly punctured assumptions about which lab was actually ahead.

## Why the real-time approach works

The core insight is that weather physics doesn't change, but the freshest available *evidence* about the atmosphere does — and most forecasting systems, AI or otherwise, have been optimized around data that's already old by design. Government supercomputers produce excellent physics, six hours late. Earlier AI weather models learned to mimic that physics, inheriting the lag along with the accuracy.

WeatherNext 3's bet is that a model trained directly on live satellite mosaics can close that gap without sacrificing the accuracy that made physics-based forecasting trustworthy in the first place — and the Brightband leaderboard results suggest that bet is, so far, paying off against both AI and traditional competitors. The tradeoff, as with most real-time systems, is that Google is now dependent on a much larger and more heterogeneous stream of live data feeds working correctly around the clock, rather than a single well-audited supercomputer pipeline.

## The zoom-out

The interesting story here isn't really about umbrellas. It's about what happens when a category that ran for decades on periodic, batch-processed updates — weather, in this case — gets rebuilt around continuous, live data ingestion instead. The same shift is happening in fraud detection, ad bidding, and inventory systems: the winning approach increasingly isn't the most sophisticated model trained on last week's snapshot, but the adequate model that's watching the present moment directly.

For builders outside of meteorology, the transferable lesson is to ask which of your own systems are still optimized around stale-but-clean batch data when a live-but-noisier feed is now available — and whether closing that lag would matter more to your users than another point of model accuracy would.

## Frequently Asked Questions

### What is WeatherNext 3?

WeatherNext 3 is Google DeepMind and Google Research's third-generation global AI weather model, announced September 3, 2026\. It generates forecasts every hour rather than every six hours, at roughly five times the resolution of its predecessor, WeatherNext 2.

### How is WeatherNext 3 different from previous AI weather models?

Instead of training primarily on the output of physics-based numerical weather prediction models, which carry a roughly six-hour data lag, WeatherNext 3 ingests live geostationary satellite imagery updated hourly, along with ground-station data and dedicated precipitation datasets from NASA and Google's own radar reanalysis.

### How much more accurate is it?

Google reports precipitation forecast improvements (measured by Continuous Ranked Probability Score) of up to 60% against satellite-based IMERG data, 30% against radar-based MRMS data, and 10% against rain-gauge measurements at early lead times. For forecasts a day or more ahead, the company says users will see up to 50% more accurate precipitation predictions.

### Where can people access WeatherNext 3?

It's rolling out inside Google Search, the Gemini app, Google Maps, the Google Maps Platform Weather API, and Google Earth Engine. Developers and researchers can also query the underlying forecast data through BigQuery or download it from Google Cloud Storage.

---

*Editor's note — sources: Google (official blog announcement); Google DeepMind; TechCrunch; arXiv preprint.*