China Wants Intelligence to Cost Zero

On Sunday night a Chinese company gave away something that cost hundreds of millions of dollars to build.

Moonshot AI published the full weights of Kimi K3, the model we talked about two weeks ago (here), which took the number one spot on global coding benchmarks. Free. Anyone can download it. They also released the software tooling to run it, which almost nobody wrote about, and which turns out to be the most important part.

Yesterday was also the IPO day of Chinese memory chipmaker CXMT, which closed +465% higher on its first day of trading, officially making it the most valuable company listed in China. The rally sent the stock's market cap from $85.5 billion at IPO to nearly $500 billion today.

On the other side, American and Korean chip stocks had their worst day of the year. Over $500 billion in value gone.

In this piece, I want to dive into what markets are so violently digesting this news and how to make sense of it.

In 2007 Google faced a problem. Apple had the iPhone, and if phones became the way people used the internet, Apple would control the door to Google's business.

Google's answer was to give away an entire mobile operating system. Android went free to any manufacturer who wanted it. Building a competing OS became pointless, because how do you charge for something your rival hands out?

The price of mobile operating systems collapsed to zero, and Google made a fortune becoming the default. Search on every Android phone. Maps, Gmail, the Play Store, ads everywhere.

That's the playbook China is now running on AI, at national scale.

It's nothing new, and the USA has been trying to take measures to prevent it for a while now. In March, the US-China Economic and Security Review Commission published a report on it, and their summary was direct: China wants to commoditize this entire technology.

The report describes two parts.

The first part is in the top layer of digital intelligence: Models. Chinese labs are constantly releasing open-source frontier models for free: DeepSeek, Qwen, GLM, now Kimi K3. Initially, this didn't really influence OpenAI or Anthropic because they were far behind in terms of advancement, and most of the AI race was about who had the best model. But the latest open-source release, Kimi K3, is very comparable to the American leaders and even better at many coding tasks. The point is every free release makes it harder for anyone to charge for AI; it fragments the model layer, and the price of intelligence falls toward zero worldwide.

The second part is physical. Because of competition, monopolies like Anthropic and OpenAI no longer have pricing power, and margins are squeezed down the supply chain to the biggest bottlenecks. Demand for AI remains the same, so the winners become infrastructure and compute providers and owners. Open-source models like Kimi K3 can be deployed by anyone who owns or rents compute, datacenters, factories, warehouses, robots, power grids… That deployment generates real operational data from the physical world, the kind no chatbot company can collect, and the data makes the next open-source model better. Which then gets released for free. And around it goes.

Here's why that second loop matters so much. China generates roughly twice as much electricity as the United States and produces about 30% of the world's manufactured goods, compared with America's 15%. Factories and power are where they're strongest.

So the strategy makes sense. If AI stays expensive and locked inside a few American companies, value pools in software, which is America's home turf. If AI becomes free, value moves to whoever can deploy it into the physical world fastest. That's the game China would rather play, so they're spending real money to change which game everyone is playing.

Back to Sunday night, and the software tooling I mentioned.

Alongside the model, Moonshot released something called MoonEP. In simple terms, it's plumbing: software that helps big AI models run efficiently across many chips at once. Deeply useful, completely free, and the kind of thing engineers around the world adopt within weeks because it saves them months of work.

Now look at which chips it supports.

Zhenwu PPU (under review, coming soon).

Zhenwu is a Chinese AI chip.

So the free software that developers everywhere are about to build on is being quietly wired to run on Chinese silicon. Free tools today, and the tools happen to work beautifully with hardware Beijing would like you to buy tomorrow. Android didn't just give away an operating system either. It gave away an operating system preloaded with Google.

Now the strangest part of the week.

On July 24, in the middle of the selloff, Nvidia published an open letter arguing that America must protect open AI models rather than restrict them. Seventy-plus companies co-signed. Nvidia and AMD, who fight over every chip socket. Google, Meta, and OpenAI, who are in a knife fight over models. Mozilla and Andreessen Horowitz, who agree on almost no policy question ever.

Why would rivals unite like that?

Because after Kimi K3, Washington started discussing restrictions on open models, including America's own. The logic in Congress sounds reasonable at first: open models help China, so slow them down.

The industry thinks that would be a catastrophic own goal. Free models spread everywhere, and whoever's models the world builds on ends up controlling the tools, the standards, and the hardware those models are tuned for. If America restricts its own open models while China floods the world with theirs, developers everywhere default to the Chinese stack. China loses nothing. America loses the entire foundation layer.

Each signature has an obvious motive. Nvidia wants free models because they turn every company on earth into a chip buyer, instead of five giant labs who can squeeze Nvidia on price. Meta's whole AI strategy is open models. Google and Microsoft rent out computing power to everyone who runs them. OpenAI signed too, which surprised people, though it ships open models now and the letter also defends a training technique it relies on.

One notable absence: Anthropic didn't sign.

The headline that did the most damage said China had begun mass-producing its own chipmaking machines. Within hours, it was community-noted: those are prototypes in testing at SMIC; mass production is targeted for 2027 at the earliest; output is tiny; performance is still well behind ASML.

CXMT's listing spooked memory investors on oversupply fears. CXMT makes none of the high-end memory AI servers need, and has no access to the advanced machines required to make it. The profitable end of memory is untouched.

Half a trillion dollars repriced on a strategy that hasn't shipped. That tells you how much of this trade is positioning.

Strip everything else away and one thing stands out. The world's two rival superpowers now agree on where AI value ends up.

One government is working to make AI models worthless as a matter of national strategy. The other's entire technology industry just signed a document arguing that America should embrace the same thing. Nobody with real money at stake is defending the idea that profits stay in the models themselves.

Which leaves everything underneath. The chips, the memory, the electricity, the data centers, the power equipment, the robots. Everything that has to be manufactured, energized, and physically deployed. That's been the thesis in this newsletter since day one, and this week it picked up state backing on both sides of the Pacific.

The market briefly voted the other way this week. Software stocks rallied while chips crashed. That disagreement is what to watch.

Personally, I’m waiting for the next two days of earnings calls before changing any positions. These reports could be key for the AI infrastructure, memory, and semiconductor sectors, with names like SK Hynix, Samsung, Microsoft, and Meta providing important signals on HBM demand, AI spending, and the broader chip cycle.

This is not financial advice. neym is an independent research newsletter. The author holds positions in assets discussed. Do your own research.