US AI restrictions push developers toward open-source models

Key Takeaways

• Trump administration ordered Anthropic to block non-Americans from its top models, forcing the company to pull them offline entirely
• Open-source alternatives like DeepSeek and China's GLM-5.2 are gaining market share as closed AI access becomes unreliable
• On OpenRouter, Google, Anthropic, and OpenAI's combined usage share dropped from 55% to 33% between January and June 2025

The Trump administration's sudden crackdown on frontier AI access has triggered exactly what regulators probably didn't want: a rush toward open-source models, including Chinese ones. In early June, the White House ordered Anthropic to block non-Americans from using its most powerful systems. Rather than build complex screening infrastructure, Anthropic pulled the models offline entirely. OpenAI followed a different path, agreeing to let the government approve every customer for GPT-5.6.

The result? Developers and enterprises are scrambling to reduce their dependence on any single AI provider. And open-source alternatives are the obvious hedge.

What exactly happened with Anthropic and OpenAI?

The administration ordered Anthropic to restrict foreign access to Mythos 5 and Fable 5, its most capable models. Faced with the operational nightmare of verifying every user's nationality, Anthropic chose the nuclear option: it shut down access for everyone. OpenAI took a different route with GPT-5.6, accepting government oversight of its customer base.

For developers who had built workflows around these models, the disruption was immediate. Haitham Mengad, co-founder of AI music startup Stems Labs, described losing access to Fable as withdrawal. "It was the first time that I realized... it's almost like a drug," he said.

The episode exposed a fundamental vulnerability in closed AI systems. The company that controls access can revoke it at any moment, whether by choice or government order. Oren Michels, CEO of Barndoor AI, put it bluntly: "If everything you need to do has to be on a specific frontier model, that makes whatever you're building a whole lot less reliable."

Why are open-source models different?

Open-weight models flip the control equation. Developers release the core model files publicly. Anyone can download them, modify them, and run them on their own infrastructure. Once released, neither the original company nor any government can claw them back.

This matters for three reasons. First, reliability. Your AI stack doesn't break because of policy changes in Washington. Second, cost. Open models eliminate per-token API fees. Third, customization. You can fine-tune open models for specific use cases without asking permission.

The timing of China's Zhipu AI releasing GLM-5.2 couldn't have been better for the open-source argument. The model performs nearly as well as top offerings from Anthropic and OpenAI on several benchmarks. AI analyst Andrew Curran noted that GLM-5.2 "is free to download, fine-tune, and run on an enterprise's own servers, putting pricing pressure on frontier labs at the same time that access looks shaky."

How much market share have closed models lost?

The shift is already visible in platform data. On OpenRouter, which routes requests across different AI models, the combined share of Google, Anthropic, and OpenAI dropped from 55% to 33% in just six months. China's DeepSeek now leads by a clear margin.

Enterprise behavior has shifted too. "Maybe a year and a half ago some large company might say we bought Anthropic or we bought OpenAI, and now no one, no one buys only one," Michels observed. Multi-provider strategies are becoming the default.

What about security concerns with Chinese models?

The obvious objection: aren't Chinese AI models a security risk? The argument is weaker than it sounds. Once you download an open model and run it on your own hardware, the company that created it has no access to your data and no control over how you use it.

Mengad dismisses the fears as more "psychological, emotional than rational." That said, the models themselves could contain subtle biases or vulnerabilities introduced during training. The tradeoff isn't zero risk versus high risk. It's one set of risks versus another.

Who's championing open models in the West?

France's Mistral stands largely alone among Western companies in pushing open models as a primary strategy. Meta, which once positioned itself as the open-source AI champion with its Llama series, has stepped back from that stance.

The retreat is notable. Meta's Llama models were arguably the most capable open-weight systems available from a Western company. If Meta pulls back further, China's labs will dominate the open-source frontier almost by default.

Could open models face restrictions too?

Here's the uncomfortable possibility: if governments decide powerful AI is dangerous, they won't stop at closed models. Ethan Mollick, a professor at the University of Pennsylvania and prominent AI commentator, pointed out that "if Mythos-level models are considered risky, China will also not want them to be open."

The logic extends beyond any single government. If frontier AI capabilities are genuinely dangerous, regulators everywhere will want to keep them locked down. Open-source might be a temporary refuge, not a permanent solution.

Frequently Asked Questions

What is the difference between closed and open-source AI models?

Closed models like ChatGPT keep their code and training data proprietary. Users access them through APIs or subscriptions, and the company controls who gets access. Open-source or open-weight models release the core files publicly, allowing anyone to download, modify, and run them independently.

Why did Anthropic pull its models offline instead of screening users?

Verifying the nationality of every user would require building complex identity verification infrastructure. Anthropic apparently decided that pulling the models entirely was simpler than complying with the screening requirement.

Is it safe to use Chinese open-source AI models?

Once downloaded and run on your own hardware, the original company has no access to your data. The main risks are potential biases or vulnerabilities baked into the model during training, not ongoing surveillance.

Which open-source AI models are gaining market share?

DeepSeek from China now leads on OpenRouter. Zhipu AI's GLM-5.2 is gaining attention for matching closed model performance on several benchmarks. France's Mistral offers Western open alternatives.

Will open-source AI models face government restrictions?

Possibly. If governments decide frontier AI capabilities are dangerous, they may attempt to restrict open models too. However, once released, open models cannot be recalled, making enforcement difficult.

Source: Tech-Economic Times / ET