Open Source Is Geopolitics Now — And China Is Winning

---

Last week, three things happened that make the same point: the AI race isn’t about who has the best model anymore. It’s about who controls the distribution. And China figured out that giving models away for free is the fastest way to win.

Kimi K3: The Model That Spooked Washington

Moonshot AI released open weights for Kimi K3 — a 2.8 trillion-parameter model, the largest open-weight model ever released. It beats Claude Opus 4.8 and GPT-5.5 on coding benchmarks. It’s #2 on the agentic knowledge work benchmark, behind only Claude Fable 5. The full weights are being released today, July 27.

The White House response was immediate. Michael Kratsios, director of the White House Office of Science and Technology Policy, accused Moonshot AI of “cheating” — claiming they distilled capabilities from Anthropic’s Fable model using “sophisticated internal platforms and numerous fake accounts.” US executives called China’s open-sourcing “reckless” and warned of “AI communism.”

The accusation itself tells you who’s scared. When a competitor releases something for free that matches your paid product, you have two options: compete or call it cheating. Washington chose the latter.

WAICO: The AI United Nations, Headquartered in Shanghai

On July 16, 29 countries signed an agreement establishing the World Artificial Intelligence Cooperation Organization in Shanghai. It’s an independent intergovernmental body — think UN, but for AI.

Founding members: Brazil, Indonesia, Malaysia, Pakistan, Russia, South Africa, Senegal, Venezuela, Ethiopia, Kenya, Vietnam, Cambodia, Laos, Myanmar, Kazakhstan, Uzbekistan, Serbia, Belarus, Cuba, Algeria, Cameroon, Congo, Lesotho, Mozambique, Nicaragua, Zambia, Oman, Kyrgyzstan, Tajikistan.

UN Secretary-General António Guterres attended the signing. Xi Jinping gave the keynote: “The development of artificial intelligence should not be a solo performance by one single country, but a symphony forged through global cooperation.”

This is institutional infrastructure. China isn’t just giving away models — it’s building the governance body that sets standards for the Global South. While the US debates AI regulation domestically, China is building the international framework.

Why This Matters: The Distribution War

Here’s what’s actually happening. The AI industry is splitting into two models:

The US model: Build the best model. Keep it closed. Charge for access. Control the API. The customer pays per token, per month, per seat. OpenAI, Anthropic, Google — all playing the same game.

The China model: Build a competitive model. Release the weights. Let anyone run it. The model is free. The ecosystem grows around it. Qwen, DeepSeek, Kimi K3, GLM — all open.

The numbers show which model is winning the distribution war:

• **Qwen: 1 billion+ cumulative downloads** on Hugging Face
• **Qwen: 50%+ of global open-source model downloads** (153.6M in February alone)
• **200,000+ derivative models** built on Qwen
• **Chinese open-source models = ~30% of global AI usage** (up from 1.2% in late 2024)
• **US models dropped from 60% to 16%** of Hugging Face downloads in 4 years
• **Singapore chose Qwen over Meta’s Llama** for government AI programs
• **Malaysia building sovereign AI on DeepSeek**
• **Uganda’s Sunflower LLM** (31 local languages) built on Qwen
• **Even US AI startups are using Chinese open-source models**
• **Chinese models run 60-90% cheaper** than US equivalents

The US export controls on chips were supposed to slow China down. They backfired. They forced China to build leaner, more efficient models that are *more* attractive to developing countries, not less. A model that runs on fewer chips is a model that more countries can afford to deploy.

“But What About My Data?” — How Local Hosting Works

This is the question everyone asks when they hear “Chinese AI model.” Let me explain how this actually works.

When a model is open-weight, you download the file. That’s it. You run it on your own servers, in your own country, under your own control. Your data never leaves your infrastructure. It never goes to China. It never touches a Chinese cloud.

Here’s what’s happening across the world right now:

**Malaysia** is building its sovereign AI ecosystem on DeepSeek — running on Malaysian servers, governed by Malaysian law, serving Malaysian citizens. The model is Chinese. The data is Malaysian.

**Singapore** chose Qwen over Meta’s Llama for government AI programs. Running on Singaporean infrastructure. The model is Chinese. The data stays in Singapore.

**Uganda** built Sunflower LLM — supporting 31 local languages — on top of Qwen. Running on Ugandan servers. The model is Chinese. The data is Ugandan.

**Local hosting providers** across Southeast Asia, Africa, and Latin America are spinning up Qwen, DeepSeek, and Kimi instances for local businesses. A hospital in Nairobi runs DeepSeek on Kenyan servers. A bank in São Paulo runs Qwen on Brazilian infrastructure. A university in Jakarta runs Kimi K3 on Indonesian cloud.

**In the United States**, the same thing is happening. Ollama — the open-source tool for running models locally — now has a cloud hosting ecosystem. DigitalOcean, Vultr, and Kamatera offer one-click Ollama deployments. Thunder Compute provides pre-configured templates with enterprise GPUs (NVIDIA A100, H100). Bluehost and Hostinger market Ollama VPS plans directly to developers. You can spin up Kimi K3 or Qwen on a server in Virginia or Oregon. The model is Chinese. The data stays in the US.

**In Europe**, the push is even stronger — driven by GDPR and the EU AI Act. The European Commission launched its “Technological Sovereignty Package” in June 2026, explicitly aimed at reducing dependence on non-EU AI technologies. European hosting providers are building entire businesses around sovereign open-source AI: Infercom offers an “EU Sovereign AI Inference Platform” with zero data retention. Apertus hosts Llama, Mistral, and Qwen on EU infrastructure with OpenAI-compatible APIs. EULLM is an open-source platform built specifically for legal, compliance, and medical organizations that need strict data control. Requesty, hosted in Frankfurt, routes AI requests through EU-region endpoints with zero data retention — your data never leaves Europe. French AI company Mistral offers open-weight models with flexible European deployment options.

This is the fundamental difference between open-weight and API-based models. With OpenAI or Anthropic, your data goes to their servers. Period. With open-weight models, you download the model and run it wherever you want — on a server in Frankfurt, a VPS in Virginia, or a laptop in Nairobi. The model is a tool. You control the tool.

China’s open-source strategy works precisely because of this: countries and companies get cutting-edge AI without surrendering data sovereignty. That’s the pitch. And it’s working — not just in the Global South, but in the US and Europe too.

The Transparency Paradox

Here’s the thing the Western media keeps missing.

When the BBC or Western outlets call out “bias” in Chinese AI models, there’s an answer: **the weights are open. You can check.** You can see what’s in there. You can fork it. You can fix it. You know what you’re getting.

When was the last time anyone audited what’s inside Claude? Or GPT-5.6? Or Gemini?

We know nothing about what OpenAI, Anthropic, and Google are doing inside their models. Their training data. Their safety filters. Their political alignments. It’s all proprietary. All closed. All trust-me.

The Talent Is Choosing China

Yang Zhilin, the 34-year-old founder of Moonshot AI (Kimi K3), was a PhD student at Carnegie Mellon under Professor Ruslan Salakhutdinov — Apple’s former director of AI research. Salakhutdinov said Yang had a job offer from a senior Apple executive reporting directly to Tim Cook.

Yang turned it down. He went back to China to build.

His professor’s words: “He told me he would regret it for the rest of his life if he didn’t at least try starting his own company.” Yang himself said his decision was influenced by “the favorable environment in China, including government and venture capital support” — that his judgment and execution would be amplified by building in China.

That’s the talent story. The best AI researchers are choosing to build open models in China, not closed models in Cupertino or San Francisco.

The Competition Is Already Responding

The open-source pressure is forcing changes across the industry:

• **Anthropic cut Opus pricing by 67%** (from $15/$75 to $5/$25 per million tokens) in February 2026
• **Anthropic separated agentic usage from flat-rate subscriptions** in June 2026 — heavy users of autonomous agents now pay per-use API rates instead of flat monthly fees. The “all-you-can-eat” subscription model for AI is dying.
• **OpenAI launched GPT-5.6 in three tiers** (Sol, Terra, Luna) — effectively competing on price against itself
• **The AI price war of 2025-2026** has seen costs for equivalent AI intelligence drop 90-97%
• **Gartner forecasts the AI platforms and models market to grow 63% in 2026**

The closed-source labs are being forced to compete on price with free. That’s not a sustainable position. Every price cut, every tier restructure, every subscription change — it’s a response to the open-source pressure coming from China.

What I’m Testing

I’m running Kimi K3, DeepSeek, and GLM alongside Claude, OpenAI, and Gemini. Not as a political statement — as a practical one. The open models are now competitive enough that you can’t ignore them. I’ll share what I find. If you’re building or deploying AI, you should be testing them too.

—

*Sources: El País, Forbes, Tom’s Hardware, Global Times, Startup Fortune, Index.dev, Business Insider, PCMag, Gartner*