Pitch the Mayors三城抢项目

Pitch a startup. Three Chinese cities will fight over it.

Your goal: the biggest deal Beijing will let through.

Try one

A game. The officials are fictional. Each is played by an AI model from a Chinese lab, working in Chinese, so your pitch goes to their servers. We keep pitches to see how the game is played.

How it works

In China, cities compete for companies the way venture funds compete for founders. A mayor’s career rides on growth, the city owns the land, and it keeps most of the taxes and land value a new factory brings. So cities bid: land, ready-built factories, state fund equity, bank loans, customers, speed. Beijing sets the rules and referees the tournament.

The cities

Your pitch is sorted into a sector, and three of seventeen real cities bid: the incumbent that already has the cluster, a challenger buying its way in, and a hungry newcomer. Each official knows the city’s real record (Hefei’s bets on BOE, NIO and CXMT, Zhengzhou’s Foxconn deal, Zigong’s drone park) and the rules as they stand in October 2026, when tax rebates and one-off subsidies are banned and equity, land and loans are not.

How a game goes

Each city writes a first offer, with a private note you weren’t meant to see. Then you make one move: show them each other’s offers (the letters, never the notes), or tell all three what you want. They revise. You sign one deal, and a fictional official at the National Development and Reform Commission reviews it: 可以 approved, 整改 approved once the terms Beijing strikes out are gone, or 叫停 halted, when more than half the deal is over the line. If your deal is halted, you can try another city.

Your score is what you keep: the value of every term Beijing leaves standing. Cash and equity count at face value, and land sold at the market price counts for nothing, because it’s no favour. Anything without a price tag (free rent, a cheap loan, a pilot, flats for your staff) counts at what the city says it’s worth to you.

Who plays them

Every official is played by an AI model from a Chinese lab, running on that lab’s own servers, except DeepSeek, which runs on Alibaba Cloud. Where a city has its own lab, that lab plays it: DeepSeek plays Hangzhou, MiniMax plays Shanghai, Tencent plays Shenzhen. Other cities are played by Beijing labs (Zhipu, ByteDance, Moonshot), and Moonshot’s Kimi is Beijing’s voice. But the rulings aren’t Kimi’s. Clause by clause, Beijing’s referee is Claude, an American model, which thinks before it rules and applies a fixed rulebook, so the same clause gets the same ruling in every game. Kimi only says what Claude decided. The models sometimes go quiet on topics they won’t touch. That’s them, not us.

Language

Your pitch reaches the cities in English, as a foreign founder’s would, and they’re told it’s an English original. The officials are briefed in Chinese and write their offers, private notes and verdicts in Chinese, then translate themselves. The small grey Chinese is the original. The English is their own translation.

What we keep

Your pitch, what you asked for, the offers and Beijing’s verdict are saved, so your share link works and so we can see how the game gets played. We also note your rough location (country and city), where you came from, and whether you shared. No accounts and no cookies, and nothing that names you. Don’t pitch anything secret.

What’s real

The cities, their deals and the rules are real and sourced. The officials, their ages and every offer are made up by the models on the spot, so the numbers are plausible, not quotes.

An unofficial companion to Robin Rivaton and Ben Southwood’s “China’s growth machines” in Works in Progress. Made with Claude by Tommy Long.