Chinese AI Models Cut Into Corporate America as Export Controls Reshape the Race

SAN FRANCISCO — DoorDash has adopted Moonshot AI’s Kimi model, with CEO Andy Fang saying the software was “cheaper” and of “better quality” without degrading code outputs. Cursor, an AI-assisted coding platform, integrated Kimi into its Composer 2 coding agent, while Airbnb and industrial manufacturer Siemens have begun piloting models from Alibaba and DeepSeek.

The shift is occurring as U.S. companies seek lower-cost ways to handle routine technical work. A McKinsey & Company survey found that 20% of corporate leaders consider AI-related operating expenses, particularly token pricing, a major obstacle to scaling the technology.

Chinese developers’ cost advantage emerged after the U.S. Department of Commerce’s Bureau of Industry and Security tightened restrictions in October 2022 and October 2023 on shipping advanced semiconductors, including Nvidia’s H100 and A100 graphics processing units, to China. The controls cut Chinese developers off from the massive computing clusters available to their American competitors.

Rather than halt progress, the hardware limits pushed Chinese laboratories toward highly efficient software workarounds. “Because they had less compute to work with, they found that computationally efficient method instead of just throwing more compute at an inefficient technique, as U.S. labs initially did,” said Brendan Burke, a semiconductor and supply chain analyst at tech research firm Futurum Group.

The technical issue reaches back to “attention,” the mathematical mechanism introduced by Google researchers in the 2017 paper “Attention Is All You Need.” Attention enables a neural network to assess relationships between words or tokens and understand context, but the calculations become exponentially more expensive as documents grow longer. Burke said Chinese laboratories developed algorithms that reduced the mathematical complexity of those calculations by an order of magnitude, allowing their models to summarize highly relevant tokens with far less processor power.

The resulting economics are visible in corporate engineering workflows. Ameya Kanitkar, cofounder of AI measurement platform Larridin, said Chinese models including GLM 5.2 and Kimi 2.6 and 2.7 complete approximately 75% of common tasks at one-fifth the cost of American counterparts. “Frontier U.S. models still have an advantage on the most complex tasks, but Chinese open-weight models are becoming more than capable enough for the majority of everyday enterprise engineering work,” Kanitkar said.

Some companies are applying the models beyond general coding. Thomson Reuters created an internal system called Thomson-1 by adapting Alibaba’s open-source Qwen architecture for document-review work previously assigned to Anthropic’s Claude. Airbnb CEO Brian Chesky described Alibaba’s Qwen model as “fast and cheap.”

Transaction data tracks the movement. The AI index compiled by financial technology platform Ramp found that the share of U.S. businesses paying for platforms using open-source and Chinese-developed models rose from 4.5% in January to 6.1% in July. Hugging Face, an open-source repository, reported that Chinese open-source models represented 41% of its total downloads last year, exceeding the share of U.S.-developed models.

The models’ deployment is also tied to how they are distributed. Chinese developers frequently release “open-weight” versions through registries such as Hugging Face. U.S. enterprises can download the code and host the systems locally or through domestic cloud infrastructure such as Amazon Web Services (AWS), keeping proprietary corporate data in the United States rather than sending it across Chinese servers.

Chinese firms had to optimize architectures such as Mixture-of-Experts (MoE) and advanced attention mechanisms to run on domestic hardware, including Huawei’s Ascend series chips. The White House estimates that the United States contains 74% of the world’s leading computing capacity, supported by billions of dollars in hyperscaler data center investments. U.S. frontier laboratories, with much larger computing budgets, remained “token hogs” because they could operate less efficient exploratory base models.

The rapid adoption has drawn scrutiny from U.S. national security agencies. A joint advisory from the FBI, the National Security Agency (NSA), and the Cybersecurity and Infrastructure Security Agency (CISA) alleged that six Chinese AI companies, including DeepSeek and Moonshot, shortened standard development timelines by buying bulk subscriptions to American AI platforms.

The agencies said that since 2024, the companies had extracted “capabilities worth billions” by systematically scraping outputs from U.S. frontier models and using them to train their own systems through a process known as distillation. They said the practice enabled DeepSeek to artificially reduce the training cost of its flagship models, which the company publically declared cost only $5.6 million to train.

China’s Ministry of Foreign Affairs rejected the joint agency allegations as “groundless” and said the country’s AI progress was the direct “result of high-level scientific and technological self-reliance.”

The dispute has unfolded alongside a narrowing performance gap. A standardized model evaluation by Stanford University earlier this year placed Anthropic’s highest-performing U.S. model only 2.7% ahead of DeepSeek’s leading model.

Still, Mike Finley, chief technology officer of enterprise AI analytics firm AnswerRocket, said Chinese developers depend on American breakthroughs to establish the technical ceiling. Chinese models are highly efficient at execution, he said, but continue to build on foundational discoveries made in the West. “The work they do would simply not be possible without the frontier labs blazing the trail,” Finley said.