In the AI World, Nobody Is Loyal Forever?

Adapted from a Bloomberg Opinion column
For two years, American AI companies didn’t take Chinese AI models very seriously. They were seen as copycats  good enough to watch, not good enough to worry about. That changed this month, almost overnight.

A Chinese company called Zhipu released its newest AI model, GLM 5.2. The timing could not have been worse for American companies. Around the same moment, the US government temporarily restricted access to Anthropic’s own top models, Fable 5 and Mythos 5. So while America’s best AI tools were briefly tangled in red tape, China’s newest model showed up ready to use  and cheaper too.

Bad Timing for America:
Normally, a regulatory delay is a minor headache. This time, it happened at the worst possible moment. American companies were dealing with paperwork. China’s new model was just… available. For engineers who need to get work done today, that difference matters more than any brand name.

The Engineers Switch First:
Here’s what should worry American AI companies: it’s not casual users switching to Chinese models. It’s serious, technical people.
Well-known tech investor Marc Andreessen says industry insiders now admit GLM can match  and sometimes beat  American AI models, with no real downsides. The CEO of Coinbase says his company is keeping its AI spending flat, even as usage grows fast, simply by using open Chinese models like GLM 5.2 and Kimi 2.7 instead of paying for American ones. A former Meta and Google DeepMind engineer says this model changes everything. These aren’t random users. These are exactly the kind of people whose choices usually predict where the whole industry is headed next.

Open Beats Trusted:
A recent global survey found something interesting: people around the world now believe China leads in AI. But they still trust American AI more.
Here’s the twist, though. Trust isn’t the only thing that matters. Open-weight models  the kind China is releasing  can be checked, changed, and run on a company’s own computers instead of someone else’s servers. That openness is its own kind of trust.
Jim Keller, a legendary chip designer who helped build some of the first iPhone processors, explained why his company switched to Chinese models for coding work. With American models, he has almost no control  he just has to trust the company running them. With open Chinese models, he can adjust them himself and run them on his own machines. He says the switch is about five times cheaper, and much more private.

The Copying Excuse:
American companies have accused Chinese developers of “distillation”  training their models using outputs from American AI. It’s a fair technique to question, but it’s not exactly rare. Even Elon Musk has admitted that his company’s model, Grok, was trained partly using OpenAI’s outputs.

And if copying alone were enough to build a great AI model, everyone would have already done it by now. There’s also a bit of irony here: the same American companies complaining about stolen training data built their own models using huge amounts of human writing and art — mostly without asking permission or paying anyone for it.
Cheaper, and Better Too

This isn’t just about price. Cybersecurity researchers who tested Zhipu’s newest model found it actually performed better than Anthropic’s own coding tool, Claude Code, on genuinely difficult security tasks  and did it for about one-sixth of the cost. When a cheaper tool also performs better, that’s a hard sell to compete against.

Banning It Backfires:
Some in Washington may want to respond by simply restricting these models further. That’s probably the wrong move. Banning open AI models doesn’t make them disappear  it just pushes their use further out of sight, where it’s harder to monitor.
AI researcher Nathan Lambert makes a simple point: if open models get banned now, while only one or two American companies keep making their private models more powerful behind closed doors, we could end up with a much bigger problem than the one we’re trying to avoid. It’s also worth asking how sincere the “safety” argument really is, coming from companies that are also getting ready for big stock market listings and defending the pricing power that comes from keeping their systems closed.

What This Means Next:
This isn’t really a contest over which country’s AI looks more impressive anymore. It’s becoming a real threat to the business plans that a huge part of America’s AI industry and stock market optimism  is built on.
America still has one big advantage: raw computing power. The smarter move now is to build a strong, open AI ecosystem of its own  fund research, train people to secure these open tools properly, and give companies real control over the AI they use  instead of hoping this challenge quietly fades away.

For India’s own booming tech workforce, this shift is worth watching closely too. As companies everywhere start choosing AI tools based on cost, control and openness  not just brand names it will shape which tools the next generation of Indian engineers grows up using.