HONG KONG — China has a hot new export in the United States: artificial intelligence. Chinese AI models from startups like Moonshot AI, Z.ai (Zhipu), and DeepSeek are quietly taking market share inside American companies, lured by prices that are a fraction of what OpenAI and Anthropic charge. The trend accelerated sharply in July 2026 with a wave of model releases that experts say are now "nearly as intelligent" as the US frontier models — but at a fraction of the cost.
San Francisco-based Raffi Krikorian, the chief technology officer at Mozilla, which runs the Firefox browser, switched to Chinese AI startup Moonshot's Kimi K3 for many of his day-to-day activities within days of the model's launch. "It just seems snappier," Krikorian told the Associated Press, comparing K3 to Anthropic's Claude Fable chatbot, which costs far more. Earlier, he had been using Z.ai's GLM-5.2 for routine tasks like managing his calendar, documents, and email.
Krikorian isn't alone. A growing number of American executives and companies are adopting Chinese AI systems, drawn by prices measured in cents per million output tokens rather than the $30 to $50 that the top US models command. Cryptocurrency exchange Coinbase said it's switching to Chinese models to trim costs, and independent software developers across the country are following suit.
The New Wave of Chinese AI Models
The rapid rise of advanced Chinese AI models this year marks a shift from early 2025, when DeepSeek first put China on the map with a model that performed on par with US counterparts at a fraction of the cost. Now the pace is even faster.
In June, Z.ai (also known as Zhipu) launched GLM-5.2, its latest flagship model. The model boasts a one-million-token context window and delivers benchmark scores in math and coding that come within 5% of OpenAI's GPT-5.6 on several key tests. Zhipu claims its training cost was under $10 million — a tenth of what US labs typically spend.
In July, Moonshot AI debuted Kimi K3 at the World Artificial Intelligence Conference in Shanghai. It's the world's largest open-source model to date, with 1.3 trillion parameters, released under a permissive license that allows commercial use. Within days of launch, Kimi K3 became the most-downloaded model on Hugging Face, the AI development platform, surpassing Meta's Llama 4 in daily pulls.
Also in July, Alibaba previewed its Qwen3.8 Max AI model. Early benchmarks show it beating Google's Gemini 2.0 Pro on reasoning tasks while costing 85% less to run. And in April, DeepSeek rolled out previews of its V4 model, which matches OpenAI's GPT-5.6 on mathematical reasoning and code generation at roughly 3% of the API cost.
The result: Chinese models are reaching what Goldman Sachs calls a "critical stage" for widespread adoption. In a July research report, the investment bank noted that a surge in "agentic" AI usage — where models autonomously conduct multistep, complex tasks — is driving demand for cost-effective alternatives. Goldman forecasts China's AI model API and subscription revenues will grow from an estimated $5 billion in 2026 to $120 billion by 2030.
More Than a Price War: "Good Enough" Beats Frontier
The adoption isn't just about cost. Curt Meinhold, a technology executive from Greensboro, North Carolina, who founded digital legacy platform LilyList, said he prefers DeepSeek for tasks like finding business leads and generating sales ideas.
"At the end of the day, most of us, the vast majority of us, 90 plus percent, don't need Anthropic's Mythos or Fable," Meinhold told the AP. "We just don't need it — we need something good enough."
That sentiment echoes across the industry. Alex Colville, an analyst at the Australian Strategic Policy Institute, told reporters that AI prices are calculated per million tokens for both input and output, and the use of AI agents compounds cost differences dramatically. "The Chinese models, you know, I find them to be pretty close on code and research," Meinhold added. "If I can pay a handful of cents per million output tokens versus $30 or $40 or $50, then it's good enough."
The math works out starkly. DeepSeek V4 costs $0.14 per million input tokens and $2.19 per million output tokens. By comparison, OpenAI's GPT-5.6 charges $15 per million input tokens and $60 per million output tokens. For a company running 10 million AI queries per day — the scale of a mid-size tech firm — switching from GPT-5.6 to DeepSeek V4 would save an estimated $42 million per year.
Even US companies that aren't fully switching are experimenting. Dropbox, Canva, and Notion have all quietly added support for Chinese AI model APIs alongside their existing OpenAI subscriptions, giving developers the option to route less critical tasks to cheaper models.
Open Source: The Unstoppable Force
Another factor driving adoption: Chinese models are overwhelmingly open-source. Kimi K3, GLM-5.2, and DeepSeek V4 are all released with open weights and permissive licenses, meaning anyone can download, modify, and self-host them. That's a stark contrast to OpenAI, Anthropic, and Google, which keep their most powerful models behind API paywalls.
For companies worried about data privacy — especially in regulated industries like finance, healthcare, and defense — self-hosting an open-source Chinese model on private servers is an attractive option. The US Department of Defense and several intelligence agencies have begun evaluating open-source models for internal use, though they've stopped short of deploying any Chinese-made systems directly.
The open-source angle also makes Chinese models harder to block. Export controls can stop chips from leaving the US, but they can't stop a 200-gigabyte model file from spreading across the internet. Once a model is released under an open license, it effectively escapes national border controls.
US Policy Response and Geopolitical Tensions
The popularity of Chinese AI models has frustrated some US tech giants, and the Trump administration is taking notice. The White House on Wednesday accused Moonshot of using "covert" methods to build Kimi K3 by training it on outputs from Anthropic's Fable model. Some US politicians and AI companies have accused Chinese startups of illicit "distillation" — extracting knowledge from US models without permission.
Beijing rejects those claims as "groundless." Chinese President Xi Jinping, speaking at the WAIC conference in Shanghai, called for a new global AI governance framework and urged nations to "seize the opportunities presented by AI development" rather than wall it off. Xi's speech proposed a China-led "AI for All" initiative that would fund model training infrastructure in developing nations.
US-led restrictions already block China from accessing advanced AI chips from Nvidia and AMD, including the H200, B200, and the forthcoming Rubin architecture. Treasury Secretary Scott Bessent has warned more sanctions could come to protect American intellectual property. But experts argue that export controls alone can't stop the rise of competitive Chinese models — especially when they're open-source and freely downloadable.
The Semiconductor Industry Association warned in a July 24 report that export controls are accelerating China's domestic chip industry rather than crippling it. Chinese companies like Huawei and SMIC are producing AI accelerators that, while less powerful than Nvidia's latest, are now good enough to train and run the models Chinese startups are releasing.
What It Means for Enterprise AI
For US businesses, the calculus is shifting. The choice is no longer just between OpenAI, Anthropic, and Google — it's between spending $50 per million output tokens on a US model versus $2 per million output tokens on a Chinese one that delivers 90% of the performance for most tasks.
Goldman Sachs predicts that enterprise adoption of Chinese AI models will accelerate through 2027, especially in Asia-Pacific, the Middle East, and Africa, where price sensitivity is higher. But even in North America, the investment bank says, cost-conscious enterprises will increasingly run multi-model strategies — using expensive frontier models for complex reasoning and cheaper Chinese models for routine summarization, classification, and data extraction.
Short of an outright ban on using Chinese AI models inside the US, they're likely to continue gaining traction among independent developers, startups, and even large enterprises looking to slash cloud AI spending. As Goldman Sachs noted, the agentic AI boom favors whoever offers the best performance per dollar — and right now, Chinese companies are winning that math.
To stay updated on the latest developments in artificial intelligence and related fields, check out our AI coverage and Cloud & Edge Computing sections.
Sources: AP News, Goldman Sachs, CNBC