Your CFO just approved a Chinese AI model. They don't know it yet.

You allocate, or you build. Either way the question this week is the same. If frontier-class AI is now a free download at a fraction of the price, what exactly are you paying the US labs for?

For months this newsletter argued an unpopular point. The US frontier lead was thinner than the valuations implied, the cost gap was enormous, and closed weights with your data on someone else’s cloud was a fragile model. The world would shift toward cost, control, and open weights, and Chinese open models would spread first across consumers and the Global South, then into Western enterprises.

This week the argument stopped being contrarian. The pace surprised even us.

On July 16, one day ahead of the World AI Conference in Shanghai, Moonshot AI unveiled Kimi K3, a 2.8-trillion-parameter open-weight model that benchmarks neck and neck with the top systems from Anthropic and OpenAI. Weights go public July 27, free for any company or government to run. The market repriced Chinese AI on the spot: Z.ai (Zhipu) fell about 30% and MiniMax about 16% in Hong Kong. One day earlier, Mira Murati’s Thinking Machines, the best-funded Western answer to Chinese open models, shipped its first model, Inkling, on a DeepSeek-V3 architecture and post-trained in part on data from Moonshot’s Kimi K2.5.

Our partner Alvin Foo, a veteran Asia tech investor and one of the largest AI voices on LinkedIn, flew back from the floor at the World AI Conference in Shanghai with a single line for us: “The whole conference was about open. The West is renting you access. China is giving away the engine.” The release that had the room buzzing was Kimi K3 landing open one day before the doors opened, and Beijing leaned all the way in, with Xi Jinping framing AI as a shared global platform rather than any one country’s walled garden. When the largest AI gathering on earth becomes a coming-out party for free, frontier-class open models, the direction is no longer in doubt.

Three takeaways if you skim nothing else:

This past weekend we guest lectured on East/West AI to 30 top CTO-level executives, part of the AI-CTO program run by our partner Dr Felix Arndt, Director at one of the world’s top western universities. The message from the room was blunt, and it is the real story of this edition.

Felix, who works with these operators every week, put the trap in plain terms. Larger Western enterprises see the cost gap clearly, and many of these leaders are already running Chinese and open-source models themselves, on their own laptops, after hours. But at the corporate level their hands are tied. Regulation, national policy, and data-privacy rules keep them from deploying the very tools their competitors abroad adopt freely. Both things are true at once: they know they cannot stay competitive without Eastern open-source AI, and they know they are not yet allowed to run it at scale. That is the trap.

Our message to them was simple. This is the moment to decide and find a way, because the cost gap is not a rounding error. When a competitor runs the same capability at one-fifth to one-thirtieth of the cost, DeepSeek V4 Pro at about 0.44 and 0.87 dollars per million tokens against GPT-5.6 Sol at 5 and 30 dollars, you cannot compete. Not a little. At all.

The second problem we worked was harder and longer term: building a scalable long-term memory for these systems. But the one keeping CTOs up tonight is the first one, balancing cost against data privacy.

So here is the question hanging over every Western boardroom. Do you stay locked into premium US AI platforms, paying five to thirty times more, while the rest of the world moves the other way? Outside the West, governments, enterprises, and startups are adopting Chinese and open-source models at speed, and so are plenty of Western startups, quietly, one layer down in the stack. The non-Western world is not waiting for permission. The only open question is whether the West can afford to keep paying the premium.

/

Pricing as of mid-July 2026. Open-weight models also run on your own servers, so your data never leaves. That is the second half of the CTO dilemma.

Here is the evidence, and why it matters to anyone allocating capital or building on AI.

The West is adopting faster than anyone predicted. Casado was careful with his 80% figure: it is 80% of the startups that use open source at all, not of every startup, and that caveat matters. Even so, the direction is not in doubt. Separately, the US-China Economic and Security Review Commission reported in March that roughly 80% of US AI startups use Chinese open-weight models somewhere in their stack, as a primary engine or a cost-sensitive fallback. Hugging Face data put Chinese-origin open models at about 41% of global open-weight downloads over the past year, ahead of US-origin models near 36%. And a CNBC investigation of OpenRouter traffic found Chinese models peaked at 46% of US enterprise tokens in a single week this summer, up from 4.5% eighteen months ago, and carrying at least 30% every week since February. This is an enterprise default now, not an experiment.

The capability gap closed in public. Kimi K3 is the largest open-weight model in the world at 2.8 trillion parameters, roughly double the nearest open competitor, and it did not arrive alone. Z.ai’s GLM-5.2 already matches Anthropic on agentic coding and security at a fraction of the price. DeepSeek’s V4 Pro sits at 1.6 trillion parameters, MiniMax and others are scaling multimodal systems. It is the field, not any single model, that should worry the frontier. This is a field, not a fluke.

The clearest tell came from the West itself. Mira Murati left OpenAI as its chief technology officer, raised a record 2 billion dollar seed at a 12 billion dollar valuation, and built Thinking Machines to be the Western open alternative. Its first model, Inkling, released July 15, follows the architecture of DeepSeek-V3 and used synthetic training data generated by Moonshot’s Kimi K2.5. Thinking Machines pretrained it from scratch, so this is influence, not a fork. The point still stands. When the marquee Western open lab builds its debut on Chinese blueprints and Chinese data, the argument is over.

The West’s other answer was price, not distance. In a single week in July, xAI shipped Grok 4.5 at 2 and 6 dollars per million tokens, Meta launched Muse Spark 1.1 at 1.25 and 4.25 dollars, its first paid model ever, and OpenAI moved GPT-5.6 to general availability, with the Sol flagship at 5 and 30 dollars. DeepSeek answered by cutting its own prices about 75%. When the strongest players compete on price rather than a capability nobody else has, the market is commoditizing.

This is the trust argument, proven. A closed model is a rental: your data sits on someone else’s cloud under their terms. An open-weight model is an asset: you download it, run it on your servers, and your data never leaves. For a bank, a hospital, or a ministry in Jakarta or Sao Paulo, that is the difference between a pilot and a deployment. It is exactly the dilemma those 30 CTOs put to us. The strategic risk sits one layer up. If Chinese open models become the default foundation the way Android became the default phone system, then whoever governs that layer sets the standards, the defaults, and the switching costs of the next AI stack. At the same Shanghai conference, Xi Jinping said it in the open: AI development “should not be a solo performance by any single country but rather a symphony of global cooperation.” Read that as an invitation to build on China’s layer.

None of this makes China a clean bet. The open weights that solve the privacy problem create a provenance problem: you are running a model whose training data, guardrails, and steering you did not set, and researchers have shown Chinese and Western versions can behave differently. The real read, from a seat in both New York and Bangkok, is that the winning move is a portfolio: open weights for control and cost, closed models where the frontier still matters, and a hard eye on who governs the layer you build on.

What we would do. Reprice the frontier labs for a world where parity is cheap or free. Own the picks and shovels that get paid no matter who wins the model war: inference, serving, evaluation, and security. And watch enterprise and Global South adoption, because that is where the defaults are being set, and right now they are being set in Chinese.

Commandment 6 in our book is Be Patient, Be Bold. As Howard Marks puts it, it is not what you buy, it is what you pay for it. The US frontier labs are a great business and a great story. For two years the market paid for a lead that felt permanent. This week showed how fast a lead reprices when a competitor gives the same capability away and customers quietly switch. Commandment 4, Don’t Be a Dead Fish, is the other half. The crowd is only now arriving at a call we made months ago. The discipline is not to abandon the story. It is to demand the right entry, keep a margin of safety, and let the crowd’s reaction set your price, not your emotion. Position with patience.

We have made this call for months, in public, back when it was contrarian. We are not the only ones. The independent voices who saw it early are being proven right, and we respect their work:

We will be tagging both. New here? Browse our full archive of East/West coverage at www.workoptional.ai.

If you run a larger enterprise caught in this exact bind, wanting Eastern open-source economics without breaking your regulatory or data-privacy rules, write to us. In your email, tell us what you are building and the size of your company, and we will try to help. We would love to share what is working for others and point you to the best practices for a hybrid stack, the right mix of Western and Eastern, open and closed, so you capture the cost advantage without giving up control of your data.

That is the ask: one email to [email protected], and we will do our best to help.

Also worth doing: subscribe to East/West AIpha for the weekly signal, one from the West and one from the East, and forward this to one operator still betting on half the board. The WorkOptional Meetup is expanding to more cities. Join us: https://www.meetup.com/workoptional/

Which Chinese open model are you or your portfolio companies already running: Kimi, Qwen, GLM, or DeepSeek? Reply and tell us. We read every one.

Disclaimer: East/West AIpha is for informational and educational purposes only and does not constitute financial, investment, legal, or tax advice. Nothing in this newsletter is a recommendation to buy or sell any security. The views expressed are our own, may be incomplete, and can change without notice. Do your own research and consult a licensed professional before making investment decisions. WorkOptional.ai, its writers, and affiliates may hold positions in companies or assets mentioned. Past performance is not indicative of future results.