nytimes.com - China AI open weight US


China’s A.I. Has a Winning Formula. America, Don’t Panic.

By Jason Hsu, The New York Times, July 23, 2026

…
Last week, the Chinese start-up Moonshot A.I. unveiled Kimi K3, which the company says is the world’s largest open-weight A.I. model and rivals the performance of Anthropic and OpenAI’s best models to date. Alibaba’s Qwen is now the most downloaded A.I. model family on earth. Developers have used it to build more than 100,000 specialized models of their own. On OpenRouter, a service that lets developers send their work to whichever A.I. model they choose, Chinese open-weight models accounted for more than 20 percent of the units of text processed over a number of weeks in 2025 and more than 40 percent over a few weeks this spring; in 2024, such models’ weekly processing of text was as low as about 1 percent…

This had been China’s plan since 2017: push Chinese models into every market, shape the standards others adopt and pull the world toward the Chinese hardware the models run best on. The country ran the same sequence with solar and telecom equipment and, most notably, critical minerals — building the world’s dependence on cheap rare earths before restricting exports once that dependence was real…

Chinese models have come to dominate in places like Nairobi and São Paulo because China made them free, packaging the model, the hardware and the financing together. …

OpenAI and Anthropic’s business is the “frontier”: the hardest, highest-value work, offered to customers that want the best model available and will pay for it. That is not the business an open-weight model takes, where finances or data-sensitivity factor in more heavily. For many U.S. customers, the frontier model keeps the hard work moving, and something cheaper — and easier to tinker with — sits underneath it… [end quote]

“Frontier” AI is expensive to create and expensive to run. But the vast majority of practical end-uses are repetitive and use only a tiny fraction of the real capability of frontier AI. These “distilled” AI programs can be run in-house on a large customer’s own servers or in the cloud cheaply or even free if the distilled AI was developed from open source AI and owned outright by the customer.

This is a real threat to the entire high-cost frontier AI model.

Will the few paying customers who are willing to pay for the gigantic capabilities of frontier AI cover the immense investments in infrastructure? Or will the CFOs and CIOs choose distilled AI for 95% of their work and reluctantly pay out for the 5% that requires those expensive frontier AI tokens?

A rapidly growing ecosystem of specialist AI engineers, open-source boutique consultancies, and independent AI labs is already doing precisely that—enabling enterprises to bypass traditional hyperscalers and frontier API providers for the vast majority of their AI workloads.

Gemini has much more detail but it boils down to this:
There will always be a need for frontier AI for the most sophisticated multi-step AI applications. But a customized distilled AI has at least two tremendous advantages: It’s a one-time capital expense (instead of infinite rental like SaaS) and it’s confidential if it’s run in-house (either literally on in-house servers or in the cloud). The hardware needs will be much less expensive as well as the ongoing expense for “inference” (that’s jargon for using AI to solve a problem).

This strikes directly at the business model of the hyperscalers and all their high-end suppliers.

Wendy