A weather AI company founded less than two years ago has pocketed funding and a satellite at the same time. Tianxu Intelligence Technology (Jiaxing) Co., Ltd. recently completed a Pre-A+ funding round of tens of millions of yuan, exclusively invested by Hou Xue Capital, with the funds primarily going toward construction of the Ningkong constellation, iteration of the energy weather large model, team expansion, and overseas preliminary research. Born only in October 2024, this company is one of the very few AI for Science players in China holding at once a vertical-domain weather large model, its own satellite observation network, and industry infrastructure-level delivery capabilities. Founder Liu Li holds a doctorate in meteorology from the University of the Chinese Academy of Sciences and has spent nearly a decade immersed in the intersection of meteorology and energy.
In Liu Li's judgment, new energy inherently depends on the weather: wind speed and irradiance directly determine power generation. With China's new energy installed capacity share already exceeding 50%, power forecasting has created rigid demand at the tens-of-billions-yuan level, and spot power trading has opened up incremental growth at the hundreds-of-billions level—this is a golden track on the eve of an explosion. Tianxu's approach is to push the large model all the way from the cloud into space.
Its weather large model is not a large language model, but a vertical-domain model with a parameter scale within 10 billion, developed from the first line of code in-house, with energy-oriented modifications made on mainstream architectures, plus a design called the Lorenz framework—named in tribute to Lorenz, the meteorologist who proposed chaos theory. The core idea is to use Transformer as the foundation, modifying the traditional attention mechanism to address phenomena in energy meteorology such as upstream-downstream spatial correlation and temporal drift, locking the target from the very start on forecasting wind speed and irradiance rather than generically reporting temperature, humidity and precipitation.
Compared with traditional physical models that solve systems of mathematical equations on supercomputers taking a dozen-plus hours, the AI large model computes weather more than twenty days ahead in one minute, and many elements are actually more accurate. Since 2025, Tianxu has co-built workflows with the Public Meteorological Service Center of the China Meteorological Administration, and in October of that year jointly released CMA-WSP-AI, China's first energy weather large model, in Zhongshan, Guangdong. It goes a step further with its vertical-domain world model: whereas ordinary weather large models only output meteorological results, it can further simulate how weather changes propagate into new energy output fluctuations, how they shake total grid load, and how they drive spot electricity price movements, achieving the leap from computing weather to computing energy system evolution, providing simulation capabilities for power trading, large base planning, and grid risk rehearsal. In real tests, its 1-kilometer high-precision wind speed prediction outperformed EC-HRES and AIFS on metrics such as RMSE and correlation coefficient, took first place in an extreme weather AI competition, with TS scores improving over traditional baselines by anywhere from 39.6% to 351%.
On the business side, Tianxu has built a working three-tier model of data, platform, and capability building. The bottom layer sells data, with customers covering top power generation groups, meteorological departments, and power trading entities; the second tier is an account-subscription SaaS power trading decision-support and power forecasting platform, bringing recurring revenue; the third tier deploys data, models, and platforms as a whole into customers' intranets, meeting top clients' security and privatization requirements. The three tiers progress in sequence: data is the entry point, the platform drives repurchase, privatization locks in major clients, and the more stations it connects, the more accurate the model becomes, forming a positive data flywheel. Liu Li pinpointed the weakness of the old model: sending data from the generation side to a third party and having it processed and sent back no longer works under energy security requirements—data cannot leave the intranet—so Tianxu builds its models and platform into customers' intranets, letting data and models close the loop within the intranet while helping customers master this independent capability. Its customers now cover the three sides of generation, consumption, and grid, with both State Grid and China Southern Power Grid included—effectively earning an entry ticket to China's energy system's most core clients.
What truly sets this company apart from traditional weather service providers is that it sent its large model into the sky. Liu Li believes algorithmic advantages will eventually be caught up with, and that the data source is the long-term moat, so it launched the Ningkong constellation plan. On July 24, 2026, the weather large model was carried into orbit aboard a satellite and completed daily verification, making it one of the earlier startups in China to put a model into space; the constellation is planned at 18 satellites in the long term, building a global 1-kilometer, 1-hour revisit wind-solar-temperature-pressure-humidity observation network, pursuing the scarce route of AI large model plus own satellites plus on-orbit computing. The subsequent schedule is already set: in 2027 it plans to launch two pilot experimental satellites to verify inter-satellite transmission and on-board computing, shoring up the wind-solar observation shortfall in desert, gobi, and wilderness regions, and later laying out the full 18 satellites to serve domestic markets and Belt and Road regions with a global perspective.
Beyond energy, Tianxu has also extended its reach into low-altitude meteorology and financial meteorology: for the former, it has provided drone delivery weather services with a top client at resolutions reaching the 10-meter level; for the latter, it targets weather derivatives, insurance, and overseas prediction markets, offering a trusted, high-frequency data foundation for pricing and settlement. Against the backdrop that about 80% of industries in the national economy are related to weather, the notion that weather can be priced and traded has been validated overseas but is only just starting in China. When a startup welds a Transformer-modified attention mechanism, its own satellites, and an energy world model into a single pipeline running from space observation to industry delivery, what it clearly wants to be is not just a forecasting tool, but the underlying pricer of global weather risk.