Google's WeatherNext 3 trades the six-hour lag for hourly forecasts
Two stories this afternoon point the same direction: AI is being pushed closer to the physical world it's supposed to describe. Google rebuilt its weather model to learn from live satellite and station readings instead of yesterday's simulation output, and Equinix is turning its colocation floors into a marketplace for inference — the part of AI that has to sit near the people using it.
Google released WeatherNext 3, and the important change isn't the resolution — it's what the model learns from. Most AI weather models, including Google's own WeatherNext 2, are trained on the output of numerical weather prediction: supercomputer physics simulations that carry roughly a six-hour data lag. WeatherNext 3 ingests a live, global mosaic of geostationary satellite imagery and trains directly on sparse ground-station observations instead, which lets it issue a fresh forecast every hour grounded in the most recent picture of the atmosphere. The resolution follows: temperature and moisture at 5 kilometers, other surface variables at 10, atmospheric variables like wind at 25 — roughly five times sharper than the 25-kilometer, six-hour grid its predecessor produced.
Precipitation is where that matters most, because rain is the variable global models have always handled worst. Google says evaluations against baselines show Continuous Ranked Probability Score improvements of up to 60 percent against NASA's IMERG satellite dataset, 30 percent against radar-based MRMS, and 10 percent against actual rain gauges at early lead times — and that users will see precipitation forecasts up to 50 percent more accurate a day or more out. The model also forecasts wind at 100 meters, about turbine height, plus cloud cover and surface solar radiation, and it's already live across Search, Gemini, Maps, Google Maps Platform, and Cloud.
Our take: the underrated consequence is geographic, not technical. WeatherNext 3 gets its biggest gains in places with sparse rain-gauge coverage — Latin America, Africa, Asia-Pacific — because satellite data doesn't care whether a country could afford a regional supercomputing center. High-resolution forecasting has quietly been a rich-country utility for decades; a model that trains on observations rather than simulations is the cheapest route yet to distributing it. Keep the skepticism, though: Google's numbers are its own evaluations, and national weather agencies still blend AI output with physics models before issuing warnings.
Equinix, Nvidia and Together AI are building an inference exchange inside colocation facilities. Announced at an Equinix event in San Francisco, Equinix Inference Exchange launches in the first quarter of 2027 and combines three layers: Nvidia's accelerated computing reference architecture, Together AI's open-model inference platform (roughly 200 open models, and Together is the seller of record that bills end customers), and Equinix's interconnection fabric across 281 facilities in 77 metros. CEO Adaire Fox-Martin called the pitch "neutral by design, open by default." Jensen Huang, appearing by video, argued Equinix's urban locations put compute "close to where the action is, where all the sensors are" — simultaneously close to users and far away from any single provider's cloud.
Why it matters: the bet is that inference, not training, is where the next data center cycle gets decided. McKinsey estimates inference will be half of all AI compute and 30 to 40 percent of total data center demand by 2030, and inference cares about latency to end users in a way training never did. That's a structural advantage for a company whose whole portfolio is small, city-center interconnection hubs — Equinix's xScale facilities are mostly under 100 megawatts while new AI campuses are measured in gigawatts. Investors seem to agree: shares are up 33 percent this year to a $100 billion market cap, the most valuable data center REIT, on revenue of $2.63 billion last quarter (up 16 percent) and $477 million of net income.
Our take is more cautious than the press release. This is Equinix monetizing the one asset hyperscalers can't copy — neutral, dense interconnection in city centers — and it's a genuinely defensible niche. But it ships in Q1 2027, financing terms weren't disclosed, and short seller Jim Chanos has called the colocation REITs "not great businesses" with low returns on capital. The sharper read: Equinix is positioning itself as the Switzerland of inference, and that only pays if enterprises actually want to run models in more than one place.
What to watch: whether national weather agencies start weighting WeatherNext 3 output in official warnings, and whether Equinix names any launch customers before the 2027 date slips.
If Google's forecast beats your national weather service on your own street, which one do you trust? Tell us in the comments.
Sources: Google DeepMind — Introducing WeatherNext 3 · The Verge · TechCrunch · WeatherNext 3 paper · Equinix newsroom release · CNBC · Tech Wire Asia
