Envision's Tianji is now the best weather model nobody in the West talks about
Three stories from Shanghai's Bund Summit and the Chinese feeds this morning: an energy company's weather model just took the crown on the field's most-cited benchmark, Alipay's assistant started booking real travel, and a Nobel laureate put a number on what AI is worth.
Envision's Tianji weather model entered Google's WeatherBench 2 for the first time this week and landed in the top tier alongside Google's own WeatherNext 2 — beating it on wind speed and precipitation, the two variables the energy business actually cares about. Envision is a Chinese wind-turbine and storage maker that built a "physics AI" forecast system: it forecasts 45 days out at 5-kilometre, 10-minute resolution, and it's trained on global meteorological data plus Envision's own satellite feeds and first-party measurements from wind and solar sites the company operates. Tianji also took first place in the China National Meteorological Center's AI weather model demonstration program and topped the State Grid load-forecasting competition, after scoring 132.34 on the China Southern Power Grid contest in June — 3.82 points clear of second place. The model's own origin story matters less than the pattern: Envision's Tianji entry is now listed on the WeatherBench 2 site next to GraphCast, GenCast, Pangu and Aurora.
The reason this is a story and not a press release is what it says about who is winning applied AI. Weather forecasting was the clean early proof that transformers could do science — Google DeepMind, Huawei and Microsoft Research set the pace, and every serious entry since has come from a lab with a supercomputer. Envision is not that. It is an industrial company that needed better forecasts to trade electricity, and it used 16 years of operating data from its own turbines as the advantage no lab can buy. That's the same shape as the Mazu (MAZU) early-warning system that turned heads at WAIC 2026: the centre of gravity in weather AI is drifting from models that win benchmarks to models that get plugged into grids, and the entities best placed to build those are utilities and equipment makers, not research labs. Keep the obvious caveat — WeatherBench 2's published scorecards still date from 2022, and the claim that Tianji sits in the top tier comes from Envision's presentation at the Bund Summit rather than a refreshed public leaderboard. Verification is thin and worth watching: if Google updates the scorecard and Tianji is genuinely there, this is the first time an industrial Chinese model has led a global AI-for-science benchmark.
Alipay's AI assistant "Abao" can now book you a flight and sign you up for a Marriott account in one sentence, because Fliggy wired its services into it over MCP. Say "next Friday Shanghai to Chengdu, get me the cheap Fliggy fare" and Abao parses the cities, the date and the price preference, calls Fliggy's services through the Model Context Protocol, and returns a result card with a stated reason for the recommendation — tap it and you land in the Fliggy mini-program inside Alipay to pay. If you leave out the date, it asks. Attraction tickets work the same way, the Fliggy Marriott flagship store is connected for membership signup and points queries, and hotel booking is coming. Abao is the renamed, AI-native version of Alipay's assistant that Alibaba has been pushing since last year; this is the first time it closes a commercial transaction on a third party's rail rather than just answering.
What makes it more than a Chinese product note is the plumbing. MCP is Anthropic's open protocol for connecting models to tools, and it has quietly become the default interface for agentic commerce outside the US — which matters because the US debate is still stuck on whether agents should be allowed to spend money at all. Hundreds of millions of Alipay users now have a path from intent to checkout that never opens an app, and the merchant keeps the order. The unresolved question is exactly the one the US is arguing about: when an assistant picks a fare "with a recommendation reason," who is liable if it picks badly, and who captures the margin?
Nobel laureate Philippe Aghion told the Bund Summit that AI could add 1.08 percentage points a year to productivity growth over the next decade — a bigger effect than the IT revolution. The decomposition is the interesting part: automating tasks in producing goods and services gets you 0.68 points, and Aghion calls that a floor; adding AI's effect on generating new ideas contributes at least another 0.4. He then spent the rest of the talk on why the number probably won't land — his own creative-destruction framework says innovation needs winners to be rewarded, but winners then use their position to block entrants, which is what he says happened after 2000 when US firm entry rates fell and productivity growth faded even as superstar firms grew. He warned that the AI value chain's upstream is already highly concentrated, and argued that education, unemployment protection and retraining decide whether the potential shows up.
It's the most useful framing of the productivity debate we've heard from an economist of that rank, because it refuses both poles: the 1.08 is real and large, and it is conditional on competition policy most governments are currently not writing. The concentration point is also the one the industry keeps dodging — the same firms arguing that AI will lift everyone's output are the ones buying the compute, the model labs and the talent.
What to watch: whether Google refreshes the public WeatherBench 2 scorecard with Tianji on it, and whether MCP-based booking in Alipay produces the first large-scale agent-commerce liability fight.
Would you trust an AI assistant to book a flight and pay for it — or is checking out the line where you want to stay in the loop? Tell us in the comments.
