Banning Chinese Open Models Won't Protect American Startups. It Will Just Make Them Build on Worse Infrastructure.

On April 29, the House Committee on Homeland Security and the House Select Committee on China sent letters to Anysphere, the company behind Cursor, and to Airbnb. The committees want to know why American companies are running Chinese-developed AI models like Alibaba's Qwen and Moonshot's Kimi in their stacks. It's the opening move of a joint investigation into US corporate use of Chinese AI, and if the reporting on White House deliberations is accurate, the endgame may be bigger: a ban, or at least a six-month delay, on frontier-tier open-weight models, meaning anything that surpasses the GPT-5.5, Claude Opus 4.8, GLM-5.2 class.

I run a startup. I've watched this debate from the seat where the invoices land. And I want to explain why this policy, as described, will hurt exactly the people it claims to protect.

Why companies picked these models in the first place

Nobody at Cursor or Airbnb chose a Chinese model out of ideology. They chose it because of a spreadsheet.

Citi's research puts the average price of Chinese models at roughly 18 cents per million tokens, against a roughly $4 per million average for comparable US frontier models. Look at a specific matchup: Zhipu AI's GLM-5.2 costs $1.40 per million input tokens and $4.40 per million output tokens. Anthropic's Opus 4.8 costs $5 input and $25 output. That's about 6x more expensive on output for capability that, on the benchmarks that matter for coding and agentic work, is close enough that most teams can't tell the difference in production.

If you think a 6x price gap doesn't matter at scale, look at Uber. Five thousand engineers burned through the company's entire 2026 AI coding budget by April. Uber's response was to cap individual AI tool spending at $1,500 a month. That's one of the best-capitalized engineering organizations in the country, rationing tokens like it's wartime. Now imagine you're a 12-person startup where inference is your second biggest line item after payroll. The 18-cent model isn't a preference. It's the difference between a business model that works and one that doesn't.

What a ban actually does, and what it doesn't

Here is the part that gets lost in every hearing: open weights that have already been downloaded cannot be recalled. They are files. They are on hard drives in San Francisco and Singapore and everywhere else, mirrored, torrented, fine-tuned, and quantized into a hundred variants. A ban does not remove them from the internet. It does not remove them from foreign competitors. It does not remove them from anyone who already hit the download button.

What a ban does accomplish is precise and narrow. It stops future legal downloads by compliant US companies. That's it. That is the entire enforceable surface of the policy.

So walk through who is actually affected. A startup in Berlin or Bangalore keeps running GLM-5.2 at 18 cents per million tokens. A US company that ignores the rules keeps running whatever it already pulled down. The only party that changes behavior is the compliant American startup, which now has two options: pay 6x more for closed US APIs, or build on less capable US alternatives. Either way, the American company is now building on worse infrastructure than its foreign competitors, by law.

That is not protection. That is a self-imposed handicap with a national security label on it.

The security concerns are real. The ban still doesn't solve them.

I'm not going to pretend the risk is zero. Booz Allen reportedly found that some Chinese models inject more vulnerabilities into code under certain personas. That's a real finding and it deserves a real response.

But look at what sophisticated companies already do about it. They don't pipe raw model output into production. They run evals. They sandbox generated code. They put trust layers between the model and anything that matters. Every serious engineering organization treats model output as untrusted input, regardless of where the model came from, because American models hallucinate and write insecure code too.

A blanket ban does none of that work. It doesn't harden a single pipeline. What it does is push usage underground, where the companies still running these models have every incentive to hide it rather than secure it, or offshore, where a Delaware startup quietly becomes a Cayman startup with a US sales office. Neither outcome makes American software safer. If Congress wants to address the vulnerability findings, fund evaluation infrastructure and set standards for how model output gets validated. That actually touches the problem. Prohibition doesn't.

The uncomfortable question underneath all of this

Here's the thing nobody in these hearings wants to say out loud. If US labs were price-competitive, there would be no story. No letters to Anysphere, no White House deliberations, none of it. Companies would be running American models because they'd be the obvious choice.

The reason Chinese open models are inside American companies is that they deliver comparable capability at a fraction of the price, and the American alternatives at that price point mostly don't exist. That's not a security failure. That's a market failure, or more precisely, a market working exactly as designed while one side refuses to compete on price. You cannot legislate your way out of being 6x more expensive. You can only make it illegal for your own citizens to notice.

Compete, don't prohibit

I want to be clear about where I stand, because this gets flattened into a pro-China position and it isn't one. I'm pro open-source AI. Emphatically including American open-source AI. The best outcome for the United States is not a world where startups are forced onto expensive closed APIs. It's a world where the best cheap open-weight model is American.

That outcome is buildable. We have the labs, the talent, and the compute. What's missing is the policy will to treat open-weight competitiveness as strategic infrastructure. Fund it. Get frontier-quality American open weights out the door with permissive licenses and pricing that competes at the 18-cent tier, not the $25 tier. Support the companies serving them. Make the default choice for a startup anywhere in the world an American model, not because the alternative is banned, but because ours is better and just as cheap.

Banning Chinese open models won't protect American startups. It will strap them to the most expensive inference on the market while the rest of the world builds on the cheap stuff. If we're worried about Chinese open-weight dominance, and we should be, the answer is to out-build it. That's the American play. It always has been.