OpenAI and Anthropic Just Joined Forces to Kill Open-Source AI. Here's Why That Should Worry You
OpenAI and Anthropic Just Joined Forces to Kill Open-Source AI. Here's Why That Should Worry You
July 22, 2026. A group of American AI startups sent an urgent letter to the Trump administration: don't ban Chinese open-weight models. The same day, OpenAI and Anthropic — fierce rivals who usually can't agree on lunch — joined hands to lobby for the exact opposite. Their reason? "National security."
Two camps. Same issue. Opposite answers. And the outcome of this fight will determine what models you can use, and what you'll pay for them.
The "Safety" Playbook is a Budget Sheet
I'll say the quiet part out loud.
OpenAI and Anthropic aren't scared of Chinese backdoors. They're scared of free models eating their lunch.
Kimi K3. DeepSeek. Qwen. These Chinese open-weight models are catching up to frontier closed-source systems, and they cost a tenth as much — or nothing at all. I download the weights, run them locally, fine-tune them, deploy them. Zero dollars to OpenAI. Zero dollars to Anthropic.
For companies whose entire existence depends on selling API tokens, this is an extinction-level event. Tom Bedor puts it bluntly in his analysis: "The arguments against open-source AI are very bad." Hold the "national security" fig leaf up to a business-model microscope and it lasts about three seconds.
Your History Lesson: The PGP Encryption Fiasco
Bedor draws a devastating parallel: the U.S. government's failed war on encryption in the 1990s.
- Phil Zimmermann creates PGP encryption. The U.S. government's response? Classify it as "military technology" and open a criminal investigation against him.
When Netscape builds SSL, the government forces them to release only a weakened "international" version. Result? Total policy failure. The crippled version was easier to get, so even Americans used it. Courts eventually ruled that publishing encryption source code is protected speech, and the government backed down.
The real casualties during the restriction period? American companies' global competitiveness.
AI model weights are the same thing as encryption code. They're not missile parts — you can't seal them in a container. A model file downloads from GitHub with a single command. Ban it today, and someone's running it locally tomorrow.
Who Says Open Models Are Just a Chinese Thing?
There's a lazy assumption that only Beijing has skin in the open-weight game. Look around:
- NVIDIA: Jensen Huang calls what they're building "token factories." They don't care whose models run on their chips — more tokens means more demand. They release their own open-source Nemotron suite.
- Thinking Machines Lab: Just launched Inkling, a powerful open model. Their bet: models become commoditized infrastructure. The real differentiation lives one layer up.
- Your company: If you're building products, you need cheap models for simple tasks and granular control over customer-facing features. Closed APIs can't deliver that flexibility.
- Google and Meta: Watching OpenAI's ad product with binoculars. If AI advertising takes off, commoditizing free open-source models to kill ad competition is a textbook countermove.
Banning open-weight models doesn't just hurt Chinese AI teams. It hurts every developer and company that depends on them. That's you.
Why "AI Dumping" is a Broken Analogy
Scott Galloway has a popular argument: China did this with solar panels, steel, and EVs — match quality, slash prices by two-thirds, own the market. AI is next.
Galloway's analogy shatters on contact with software.
Solar panels and steel require physical supply chains. If nobody in your country manufactures panels, you can't open a silicon wafer business. The layers are vertically locked.
Software can be vertically decoupled. A Chinese open-weight model doesn't prevent an American fine-tuning business from thriving. It gives it a free starting point. You take it, modify it, and build on it. That's the core logic of open source: shared foundations, competitive superstructure.
What is the "AI Race" Even About?
Every time I hear "we're losing the AI race," I want to ask: what's the finish line? The best model? The most tokens sold? First to doomsday?
Talking about an "AI race" makes as much sense as talking about an "internet race." We're not in a moon landing competition. This is a transformative technology being absorbed by society. If there IS a race between nations, it's about who integrates AI into their economy faster.
In that framing, free AI models aren't a threat. They're rocket fuel.
Three Security Arguments That Fall Apart
1. "They'll spread Chinese propaganda."
A model has pro-China bias? Great news — it's open. Any American who objects can fine-tune and release a corrected version. In a free market with free speech, a biased model doesn't stand a chance against a debiased one.
2. "They'll embed backdoors."
AI hasn't rewritten the rules of vulnerability markets: good actors patch, attackers exploit. Restricting tools from good actors only helps attackers. If there's a backdoor, the best way to find it is maximum inspection — not lockdown.
3. "China is AI dumping."
Already covered. Software dumping doesn't kill competitors. It creates the surface they compete on.
What You Can Do Today
Nobody knows when this fight gets settled. Here's what doesn't have to wait:
- Audit your dependencies. Which open-weight models does your project use? Prepare 1-2 closed-source fallbacks for each. You'll sleep better.
- If you're on closed APIs, a restriction policy would hand vendors serious pricing power. Lock in long-term pricing while you can.
- Keep perspective. History has a track record: encryption export controls failed. BitTorrent crackdowns failed. Open-source AI restrictions will fail too. The trend is irreversible.
One last thing. This isn't China vs. America, and it isn't open vs. closed source. The real question is: should AI belong to a handful of companies, or to everyone?
History knows which side wins. So do you.
Sources:
✨ 本文由 DeepSeek 生成初稿，Claude 审核润色。