AI begins to “see the sky”! Google (GOOGL.US) weather forecasting model redraws the climate with an accuracy of 5 kilometers per hour, opens AI to monetize a new continent

The Zhitong Finance App learned that US tech giant Google (GOOGL.US) has launched its most accurate and advanced global weather forecasting model so far to provide rapid weather forecasts with “unprecedented resolution.” The launch of “WeatherNext 3” marks that the AI application layer is moving from chat, search, and code generation to a production system that directly affects real-world asset scheduling and risk pricing: the model takes real-time geosynchronous satellite data and ground observations as input, reinitializes every hour, and generates predictions at a maximum 5 km spatial resolution, making a significant shift compared to WeatherNext 2's 25-kilometer grid and 6-hour update cycle.

The value of WeatherNext 3 is not only more accurate weather forecasting, but also improving wind power generation forecasting, energy load scheduling, agricultural and raw material supply assessments, logistics route planning, and commodity derivatives pricing, further extending the cutting-edge AI capabilities of global AI application leaders such as Google and OpenAI from consumer-side AI chat tools to high-value industrial decision-making infrastructure.

As cutting-edge AI technology is being upgraded from AI chatbox tools to industrial decision-making core infrastructure for energy, agriculture, logistics, and commodity trading, high-frequency and high-resolution weather forecasting AI systems are expected to transform Google's cloud inference AI computing power and model advantages into enterprise cloud computing service revenue.

Google's weather forecast model that redraws global events every hour

WeatherNext 3 uses artificial intelligence to learn from real-time observation data and uses raw satellite data to generate forecasts every hour. The model enables users around the world to obtain weather forecasts through Google's products.

The model was developed by Google's DeepMind Artificial Intelligence Research Institute and Google Research Institute. It can present temperature and humidity with a resolution of 5 km, other surface variables with a resolution of 10 km, and atmospheric variables such as wind speed with a resolution of 25 km. The resulting weather picture clarity is about 5 times that of the previous generation WeatherNext 2 model, and is far faster than the current weather model's 6-hour data lag.

The company stated in the announcement: “By taking images stitched together from real-time geosynchronous satellite data from around the world, our new model has obtained an informative and continuously updated view of the atmosphere. This allows the model to generate new forecasts every hour, each based on the latest satellite observation data available at the time, with a resolution of up to 5 km.”

Google said in the announcement that with its more clear and accurate weather forecasting capabilities, WeatherNext 3 has sufficient capacity to become an important AI aid for clean energy suppliers, commodity traders, supply chain managers, raw material suppliers, and transportation operators.

Google WeatherNext 3 knocks on the door to monetization of AI applications and begins to drive real-world productivity expansion

High-frequency and high-resolution weather forecasting is expected to drive the spread of AI value from consumer-side traffic to energy transactions, supply chain management, and industrial decision-making. WeatherNext 3 and OpenAI's recently released GPT-6 Astra have shown that computing power is being transformed from high-energy model training assets into production tools in the fields of energy, meteorology, scientific research, and cybersecurity. The long-term return of tech giants continuing to finance and expand computing power ultimately depends on whether vertical applications such as WeatherNext 3 can generate sustainable and strong AI inference token demand and commercial-side monetization revenue.

WeatherNext 3 and GPT-6 Astra together revealed that AI demand is spreading from “training larger models” to “deploying more intelligent systems that operate continuously.” According to Axios, Astra's training used more than 100,000 GPUs; OpenAI confirmed that it was the company's first model to reach the “Critical (Critical)” threshold of cybersecurity capabilities, and was able to discover unknown vulnerabilities and develop new uses under appropriate tools and permissions.

These latest cutting-edge AI developments not only show that cutting-edge training still consumes a high level of computing power, but also that applications such as weather forecasting, software engineering, scientific research, and network security will create a long-term inference load — the measurement of computing power requirements is shifting from the number of models released to the call frequency, runtime length, and actual workflow penetration rate of the intelligent body.

Real world industry demand and AI infrastructure financing data are still accelerating to support the growing demand for AI applications and the strong expansion of AI computing power infrastructure. TrendForce, a well-known market research agency, predicts that the capital expenses of major cloud service providers will increase 98% year-on-year in 2026 and 50% in 2027; the combined share of DRAM and NAND flash memory in their capital expenses will rise from 47% in 2026 to 68% in 2027, and server DRAM and enterprise SSD contract prices are expected to increase cumulatively by about 270% and 235% respectively in 2026.

Although Google Gemini 3.8 Flash maintains the price of 0.75 US dollars per million input tokens and 3.75 US dollars for output tokens, the cost of each benchmark task has increased by about 40% compared to the previous generation due to an increase in the number of tokens exported by about 30% and the number of agent calls. Meanwhile, a report by Goldman Sachs's official research team shows that proxy AI workflows led by AI agents will drive monthly token consumption 24 times to 120 trillion tokens between 2026 and 2030, and determine that there may still be a shortage of chips in the next 12-18 months.

Google's parent company Alphabet supports AI and data center expansion through about 32 billion US dollars of multi-currency bond financing, while SoftBank Group, which is headed by Sun Zheng, issued a record 1 trillion yen retail bonds, showing that the AI competition has entered the stage of supporting computing power construction with long-term debt capital; what investors need to verify next is whether real applications such as WeatherNext 3 can transform huge AI capital expenses into strong cloud revenue, continued expansion of usage, and free cash flow.