China's AI Labs Are Giving the Model Away. That Is the Business Model.

On July 13, three things happened in Chinese AI within hours of each other, and only one of them was a product launch.

Zhipu AI's founder announced a two-year research plan pointed explicitly at AGI. MiniMax's CEO pledged to take zero salary until AGI is reached, and committed 5 percent of company equity to staff and open-source initiatives. SenseTime fully open-sourced its flagship vision model. All three were reported the same day by Chinese outlet Jiemian.

The easy read from San Francisco is that this is theatre. Zero-salary pledges are cheap to make and cheaper to walk back, and "two-year AGI plan" is the kind of phrase that survives exactly as long as the funding round it was written for. That read is not wrong, but it is incomplete, because it treats the announcements as claims about capability. They are not. They are claims about distribution.

## Giving the model away is a market-entry strategy

Open weights are not charity, and Chinese labs are not confused about that. Releasing a competitive model for free does three things at once, and none of them require you to have the best model in the world.

It buys adoption in markets that cannot or will not pay frontier prices. It builds developer mindshare, which is the only durable asset in a field where the state of the art has a shelf life measured in months. And, most usefully, it commoditizes the thing your competitor is charging for. If a US lab's moat is a closed model behind a metered API, the fastest way to attack that moat is not to build a better model. It is to make a good-enough model free.

This is the same play that has been running underneath GLM-5.2's emergence as a low-cost challenger to US model pricing. Price, not benchmark scores, is the weapon. A model that is 90 percent as good and 100 percent free wins an enormous number of real procurement decisions, particularly outside the US.

## Washington is optimizing for a different variable

Set the July 13 cluster against what was happening in Washington the day before. Reporting on US export controls aimed at frontier systems, including Anthropic's Fable 5 and Mythos 5, described national security concerns, export-control directives and congressional scrutiny of how advanced models are released and to whom.

Read the two together and you get something clearer than either alone. The US and China are not running the same race with different runners. They are optimizing different variables. The American posture increasingly optimizes for control: who can access the frontier, under what license, with what audit trail. The Chinese posture increasingly optimizes for diffusion: how many developers, in how many countries, build on our weights by default.

Both are rational. They just produce very different maps in five years. One produces a small number of extremely capable, tightly governed systems. The other produces an enormous installed base of adequate ones. History is not kind to the assumption that the more capable system automatically wins the market. It frequently does not.

## The showcase is the point

None of this is subtle, and China is not pretending otherwise. The World AI Conference opens in Shanghai on July 17, and the framing around it, per Chinese state-adjacent coverage, leans on "AI for good", governance and hard tech, positioning the city as the country's AI hub and China as a responsible global steward. The city has restricted drone flights across the entire municipality for the duration, which tells you how the event is being treated internally.

"AI for good" plus free weights plus a governance pitch is a coherent export product, and it is aimed squarely at the countries deciding right now whose stack to standardize on. It pairs neatly with the pattern we wrote about last week, where AI dealmaking is increasingly negotiated between states rather than by frontier labs. If you are a government in the Gulf, Southeast Asia or Latin America, a stack that is cheap, open and does not come with an export-control rider attached is a genuinely attractive proposition, whatever you think of its origin.

## Where this thesis could be wrong

Three honest caveats, because the confident version of this argument is the one that ages badly.

First, open weights may already be table stakes rather than an edge. Meta open-sourced aggressively for years and did not convert it into a durable commercial lead. Openness generates goodwill and downloads far more reliably than it generates revenue, and a strategy every serious player can copy is not a moat.

Second, we cannot verify the money. There is no reliable public revenue data on Zhipu, MiniMax or SenseTime, so the claim that openness is converting into a business rests on inference, not on filings. It is entirely possible these labs are buying share they will never monetize, funded by capital that is itself a policy instrument. Note also that the market enthusiasm is not purely domestic in origin: Chinese AI-linked equities rallied this week partly on a teased Meta model codenamed "Watermelon", which is a strange thing to happen if the domestic story were self-sustaining.

Third, a zero-salary pledge is a signal, not a strategy. It costs a founder with meaningful equity approximately nothing.

## What to watch

The measurable thing here is cadence, not rhetoric. Pledges are unfalsifiable. Release schedules are not. Watch whether Chinese labs keep shipping open weights at the pace of the last quarter, or whether the July cluster turns out to be conference-timed marketing that goes quiet after Shanghai. Watch whether US export policy tightens in a way that pushes third-country buyers toward the open stack, which would be the clearest own goal available. And watch the procurement decisions in India, Indonesia and the Gulf, because that is where this actually gets settled, not on a leaderboard.

We track these deals and releases as they land in our deal tracker and trends.

The question worth carrying into next week is not which lab has the best model. It is which labs are willing to give theirs away, and what they expect to collect in return.