SAN FRANCISCO — American artificial intelligence has become an increasingly expensive line item for startups and established firms alike, pushing many to seek cheaper alternatives abroad. Lindy.ai, a San Francisco-based startup that builds AI assistants to manage email and calendars, illustrates the shift. After months of meetings with its finance team, founder Flo Crivello said the company’s largest single expense was not payroll or rent but the top-tier models from Anthropic. “By far, our No. 1 expense was Anthropic,” Crivello said, explaining that cutting that cost was essential for Lindy’s survival. Last month, Lindy migrated 100% of its traffic to the Chinese AI model DeepSeek-V4, a move Crivello called “a very simple business decision.” “It was just 10x cheaper, adding that it had saved the company millions of dollars,” he said. The move underscores a broader trend: artificial intelligence has become one of the fastest-growing, and most expensive, costs for U.S. businesses, forcing many firms to balance the need for advanced capabilities with the desire to control expenses.
Experts say Chinese models trail Western rivals by roughly six to 12 months in capabilities, but they have rapidly expanded access through hubs like Hugging Face, GitHub, and other aggregators. Crivello argues that open-source and cheaper Chinese options have grown into a practical, scalable alternative for many tasks. “The open-source scene right now is absolutely dominated by the Chinese. It’s not even close,” he said. He noted that many founders in the AI space are already evaluating or have migrated to Chinese models.
The cost dynamic isn’t limited to startups. Uber chief executive Dara Khosrowshahi recently described AI budgeting as a constraint on growth during an interview, remarking that the company “blew through our AI budget in a quarter, you know, for the whole year, essentially.” Airbnb’s Brian Chesky has been reported as saying the company relied on Alibaba’s Qwen model last year, which he described as “good,” “fast and cheap.” OpenRouter and other platforms have also broadened access to Chinese models, including Qwen, DeepSeek, MiniMax, Xiaomi, and Tencent.
Crivello compares the choice to car brands: “It’s like the difference between driving a Ferrari and a Honda. You can have the best luxury car, or you can just have a Honda at scale that works.” He adds that many open-source groups are comfortable operating at N-1 where the frontier is still N, implying that the performance gap can be tolerable as costs decline. Still, some companies argue that the newest flagship models may still be necessary for certain tasks, such as deep reasoning or advanced coding.
OpenRouter data show Chinese DeepSeek usage rising from about 9% to almost 20% since January, with rising adoption of models from MiniMax, Xiaomi, and Tencent. Some users opt to self-host open-source Chinese models, but many prefer paid hosting services like Featherless and OpenRouter so that data can remain in the United States. For AI practitioners, cost per token has become a key metric driving deployment decisions. Su-Ortiz, Shanghai-based MiniMax’s global product marketing lead, notes that “a lot of repetitive tasks can be done with a model that’s just as performant but has much lower cost per token.” He and others say companies are shifting from relentless scaling of token usage to optimizing costs by mixing models and routing different tasks to appropriate engines. For routine coding tasks or high-volume work, cheaper Chinese or open-source models can offer substantial savings.
However, cheaper does not always mean enough. Some startups, like Comment.io, emphasize that for certain projects, returning to higher-cost, higher-capability models may be necessary to avoid increased development time and errors. Comment.io’s CEO, Jon Gordner, says saving a few dollars per token isn’t worth it if it creates weeks of additional debugging. He notes that both Anthropic and OpenAI continue to subsidize users with discounted tokens under subscription plans, which can delay the point at which Chinese or open-source models become universally preferable.
A broader view from Ramp’s economist Ara Kharazian suggests U.S. companies will continue to adapt by managing prices and adopting high-quality open-source options to outcompete Chinese offerings. He cautions that American model firms may respond competitively, potentially narrowing the gap as they scale. Still, the trend reflects a fundamental shift in the AI market: growing demand for affordable, scalable AI tools is reshaping how companies approach procurement, development, and deployment of AI capabilities in daily operations.
Overall, the market for AI models remains a battleground of cost versus capability. While many startups lean into cheaper Chinese models for everyday tasks and cost containment, a subset continues to rely on top-tier Western models for critical operations. The dynamic suggests a bifurcated market where price-sensitive tasks are served by cheaper options, while advanced reasoning and specialized functions may keep some customers tethered to higher-cost providers for now. Investors will watch how American model developers respond to this cost pressure and whether subsidies, pricing strategies, and new open-source players will alter the competitive landscape in the coming quarters.
