At a small lunch in San Francisco’s Presidio, OpenAI CEO Sam Altman told reporters that the United States is misjudging the scale of China’s advances in artificial intelligence. He doubts export bans on chips will slow Beijing’s work. For now, questions remain about whether Washington can keep pace.
KEY POINTS:
- Altman says U.S. may underestimate China’s AI capabilities.
- Export controls alone are unlikely to stop Beijing’s progress.
- OpenAI shifts strategy, releasing open-weight models to stay competitive.
Concerns over China’s pace
“I’m worried about China,” Altman said. He described the race as more than a matter of who is ahead. Inference capacity, research, and product development all play into the picture. China, he argued, may be able to move faster in certain areas.
Over tapas with journalists, Altman explained that the contest between Washington and Beijing is tangled. He insisted it will not be solved by counting chips or headlines. The real question is whether U.S. policies can adapt quickly enough.
Limits of export controls
Despite tougher rules on semiconductors, Altman is skeptical. He suggested that blocking shipments of GPUs may not matter if China builds its own facilities or finds other paths. “You can export-control one thing, but maybe not the right thing,” he said.
The Biden administration tightened restrictions. Later, former president Donald Trump went further, halting advanced chip sales entirely. But last week, a carve-out allowed “China-safe” chips to return to market under unusual terms. Nvidia and AMD must now give the government 15 percent of China revenue.
The result looks fragile. U.S. companies still depend on Nvidia and AMD. Meanwhile, Huawei and other Chinese suppliers are racing to fill the gap. It’s unclear if cutting off supply is really slowing the momentum.
Open source and competition
China’s rise has also shaped OpenAI’s thinking. The company once kept models behind APIs. Now it has begun releasing open-weight systems. Altman said the spread of Chinese open projects like DeepSeek forced his hand. “If we didn’t do it, the world was gonna head to be mostly built on Chinese open source models,” he admitted.
Earlier this month, OpenAI published two text-only models, gpt-oss-120b and gpt-oss-20b. They are designed for developers to run locally and adapt. Unlike full open-source projects, the training data and code remain private. But the parameters are public, marking a major shift since GPT-2 in 2019.
Meta had led on openness with its Llama models. Yet CEO Mark Zuckerberg has since hinted at pulling back. OpenAI is instead moving in the opposite direction. The company hopes wider access will grow its developer base and keep talent from drifting to Chinese platforms.
Mixed reaction
Not everyone is impressed. Some developers called the new models limited, pointing out missing features found in OpenAI’s commercial tools. Altman didn’t deny it. He said they were designed mainly for coding agents that can run on local machines. “If the kind of demand shifts in the world, you can push it to something else,” he said.
The Bottom Line
Altman’s warning is clear: China’s AI work is moving faster than many in Washington believe. Export bans may not hold it back. OpenAI, facing the same reality, is opening up its models to stay in the race.