Moonshot AI’s latest model had become one of the AI sector’s most closely watched releases before it was even announced.

In mid-July, online researchers who track signs of new AI models began circulating expectations about a major launch from Moonshot AI, the Beijing-based artificial intelligence startup behind the Kimi large language model. The company’s next release was widely expected to be significant.

On July 16, Moonshot introduced Kimi K3, described as the largest open-source model ever released. The company said the model could perform near the level of Anthropic’s Fable 5, widely regarded as one of the most powerful publicly available models, while costing far less to use. Moonshot’s official benchmarks consistently placed K3 among the top three AI models, and one independent benchmark from Arena.AI ranked K3 as the best model then available, ahead of Anthropic. Such rankings matter because enterprise buyers and developers often use benchmark performance, pricing, and deployment flexibility together when deciding which models to test in production.

The release marked a major moment for Moonshot and its founder, Yang Zhilin, a 34-year-old alumnus of Tsinghua University and Carnegie Mellon University. Yang, a Pink Floyd fan, based the startup’s Chinese name on his favorite album, The Dark Side of the Moon.

The launch also unsettled investors because it challenged the assumption that U.S. firms could preserve a large advantage in the global AI race mainly by spending more than Chinese competitors on computing power. Asian markets fell after the release. The Philadelphia Semiconductor Index, which is heavily weighted toward chip companies, declined 1.6%. Nvidia lost nearly $600 billion in market value and briefly ceded its position as the world’s most valuable company to Apple.

Many observers, including Anthropic CEO Dario Amodei, had not expected a Chinese AI lab to release a model approaching the leading U.S. offerings for at least another six months. Tesla CEO Elon Musk had suggested that such a development might occur by the first quarter of next year. Kimi K3 changed that timeline. Chinese AI models have become capable and inexpensive enough that some U.S. startups and Fortune 500 companies are quietly integrating them into operations to control rising AI costs.

The progress has come despite U.S. export controls intended to restrict China’s access to the world’s most advanced chips. “The AI ecosystem in China is probably much better than people thought,” said Paul Triolo, a partner at DGA–Albright Stonebridge Group.

Chinese developers have operated under significant constraints for years. The U.S. began cutting China off from top-tier AI processors, including Nvidia chips, in 2022. By limiting sales of advanced chips used to train and run AI models, as well as tools that could be used to manufacture them, Washington sought to weaken China’s technology sector and preserve the U.S. lead in AI.

DeepSeek, a Hangzhou-based lab connected to a Chinese hedge fund, challenged that strategy in early 2025. It drew broad attention in the AI industry by launching its V3 and R1 models, which matched the performance of U.S. counterparts. DeepSeek said it trained the models on a very small budget by using efficient engineering, programming techniques, and mathematical optimizations. The models showed that Chinese AI developers could continue advancing even when using less capable hardware.

DeepSeek emerged as an early leader in China’s AI race, but a rapid sequence of model releases soon shifted attention to other companies. In June, AI startup Z.ai gained prominence with the release of its GLM-5.2 model, which was particularly strong in coding and creative design. Its newly listed stock had risen more than 1,100% through mid-July. At times, Z.ai’s market capitalization exceeded 1 trillion Hong Kong dollars, or $127.6 billion, valuing a company that generated $106 million in revenue last year at roughly the same level as BYD and Starbucks. Moonshot’s Kimi K3 launch then contributed to a 40% decline in Z.ai shares over two days.

China’s consumer-internet giants are also entering the frontier-model competition. Meituan’s LongCat-2.0 model used as much data as DeepSeek’s V4 and performed at levels comparable to OpenAI and Anthropic releases from February. Meituan, best known as a food-delivery platform, also said it trained the system entirely on Chinese-made processors rather than U.S. chips.

“The idea that Meituan could train a 1.6 trillion-parameter model on domestic hardware would have been inconceivable in October 2022,” Triolo said, referring to the month when the U.S. introduced its AI export controls.

AI users around the world are adapting to an environment in which Chinese models are competitive with U.S. models on capability and often much cheaper. Chinese models now account for much of the activity on OpenRouter, a marketplace that allows developers to access multiple AI models and providers through one interface. At one point in mid-July, six of the top 10 models on OpenRouter—and all of the top five—came from Chinese companies: Tencent, Xiaomi, DeepSeek, MiniMax, Moonshot, and Z.ai.

Those rankings primarily reflect developer usage, but adoption has also appeared in mainstream companies. Last year, Airbnb CEO Brian Chesky said the company was using Alibaba’s Qwen for customer service. Cursor, the AI coding startup, has said Moonshot AI’s Kimi provided the foundation for Composer 2, its coding model. In a June social media post, Coinbase CEO Brian Armstrong said the crypto platform had cut its AI spending in half by encouraging more employees to use Kimi and Z.ai’s GLM models.

The cost gap is substantial. One million output tokens, or roughly 750,000 words, cost $50 when using Anthropic’s Fable model. The same number of output tokens cost about $0.87 on DeepSeek-V4-Pro and $4.40 on Z.ai GLM-5.2. Kimi K3 is relatively expensive by Chinese standards at $15 per million output tokens, but still below Fable’s $50 rate. According to Moonshot AI and OpenRouter, Chinese AI models accounted for 57% of tokens used by U.S. firms on OpenRouter during one week in July. For companies using AI in customer support, coding, search, or internal automation, token pricing can become a material operating cost because each prompt and response consumes billable model capacity.

DoorDash is also sending coding work to Chinese models. Chief technology officer Andy Fang said the company delegates “lower-level work” to Kimi, producing “better quality [at] cheaper cost.”

Several factors help explain the lower prices. Power is cheaper in China than in many parts of the U.S., partly because China has invested in power generation and transmission, making it easier to expand data center capacity. New U.S. data centers, by contrast, often encounter political resistance because of concerns about pressure on electric grids and water usage.

Chinese AI companies are also prepared to accept lower profit margins as they compete for market share and try to establish their models as de facto standards. U.S. export controls may have contributed to this pricing dynamic as well. Because Chinese labs lack access to the most advanced AI processors, they have had to extract more performance from less capable hardware.

“Labs are so compute-constrained, capital-constrained, and talent-constrained that a lot of them are being cautious in how they use their resources,” said Grace Shao, an AI analyst and author of the AI Proem newsletter.

Recent Chinese AI models are increasingly compatible with cheaper domestic processors. “For the money [a Chinese AI company would] spend on an Nvidia chip, they can buy 10 local chips from Huawei or other local chipmakers,” said George Chen, a partner at the Asia Group.

Chinese firms have also embraced open source. Almost all Chinese companies release their models under permissive licenses, allowing users to download and fine-tune them for free and run them on local hardware, including in the U.S. In that scenario, the relevant costs are “GPUs and energy,” said Ameya Kanitkar, cofounder of Larridin, an AI measurement platform. When Anthropic sells its models, it also factors in the cost of research and development, he said. Open-source access also gives companies more control over where models run and how they are adapted, though it does not eliminate the need to manage security, compliance, and performance testing.

Geopolitical concerns remain. U.S. officials are increasingly concerned about American companies using Chinese AI models, citing security considerations and fears that Chinese developers could continue undercutting U.S. developers on price. Congress is examining the use of Chinese AI models by U.S. companies including Airbnb and Cursor. In a statement, Airbnb said it used only “a limited number” of Chinese models, all of which were open-source and run through “approved U.S.-based service providers.” Cursor did not respond to Fortune’s request for comment. Lawmakers are also considering whether the U.S. should do more to support its own open-source models.

Chinese developers are using those concerns to strengthen their own position. Z.ai announced its GLM-5.2 model just days after U.S. officials temporarily cut off access to Anthropic’s Fable and Mythos models for some users outside the U.S. and for foreign nationals. “Frontier intelligence should not belong to only a few people, nor be subject to withdrawal by a handful of rules at any moment,” Z.ai wrote in an accompanying social media post.

Z.ai is courting governments seeking sovereign AI systems or models that can run on domestic hardware, with local control over data and upgrades. Demand has increased as U.S. policy has become more protectionist and less predictable. “The decision to restrict the latest models really backfired,” Chen said. “If you think about Singapore or India, there’s growing uncertainty around U.S. AI policy. One day, you’re told you can use the latest model; the next day, no foreign citizens can use it. How can any country deal with that kind of uncertainty?”

Beijing also sees a soft-power advantage in broader global use of Chinese AI. At an AI conference in July, President Xi Jinping pledged to “uphold openness and win-win cooperation” in the technology. AI “should not be a solo performance by any one country,” he said, “but a symphony of global cooperation.”

The article appears in the August/September 2026 Asia issue of Fortune under the headline “China’s AI firms are catching up to U.S. labs—and beating them on cost.” It was originally featured on Fortune.com: https://fortune.com/2026/07/26/china-moonshot-deepseek-zai-kimi-challenging-us-ai-cost/