Artificial intelligence (AI) expenses are rapidly increasing for businesses worldwide. Companies now look to cut costs by opting for more affordable Chinese AI models.
Startup Decisions Driven by Expense
Flo Crivello leads Lindy.ai, a startup based in San Francisco that provides AI assistants for email and calendar management. Initially, the company relied on the high-end AI models from Anthropic. However, financial meetings revealed Anthropic’s models were the largest expense, even surpassing payroll for over 24 employees and rent. Thus, Lindy.ai transitioned entirely to the Chinese model, DeepSeek-V4, which reduced expenses by a factor of ten and saved millions.
AI: Both a Necessity and a Burden
AI is becoming crucial for businesses, yet it remains costly. Some firms now transition away from American models to cheaper Chinese AI alternatives. While U.S. organizations like Anthropic, OpenAI, and Google lead AI development, Chinese models lag by six to 12 months in functionality. Nevertheless, China’s strength lies in open-source models that are free to download and modify.
“The open-source scene is dominated by the Chinese,” noted Crivello. Many entrepreneurs are considering switching to Chinese models, owing to escalating AI costs.
AI costs aren’t unique to startups. Uber CEO Dara Khosrowshahi, on the Invest Like the Best podcast, highlighted how quickly their AI budget was exhausted, prompting them to reconsider their strategies. Information about Uber’s use of Chinese models wasn’t confirmed. Airbnb, meanwhile, has relied on Alibaba’s Qwen model, which provides efficient performance at a reduced cost.
Chinese Models’ Practicality
Despite political concerns, Chinese AI models are commonly accessible on platforms such as Hugging Face, GitHub, and numerous aggregators outside China. Featherless, a San Francisco firm, provides 30,000 AI models, highly sought after by companies despite not being top-tier.
Eugene Cheah, Featherless founder, compares using these models to the choice between driving a Ferrari or a Honda. The latter often suffices without exceeding budget constraints. “As the capability gap closes,” noted Cheah, “the significance of the difference diminishes.” Use of China’s DeepSeek has grown from 9% to almost 20% since January, with models from MiniMax, Xiaomi, and Tencent also increasing.
Some firms prefer downloading these models directly, while others opt for data security with services like Featherless and OpenRouter.
Token Economics
AI models operate on tokens, measuring units of AI output. Victor Su-Ortiz from MiniMax stressed cost per token as the critical factor in model selection.
Companies now focus on utilizing models with low token costs for repetitive tasks, making Chinese models attractive for such applications. Nevertheless, top-end models might better suit tasks requiring advanced analysis.
Mixed Receptions and Industry Adaptations
For some firms, such as Comment.io, Chinese models aren’t yet viable. CEO Jon Gordner emphasized that while immediate savings are alluring, quality matters. His company benefits from subsidized models from Anthropic and OpenAI, offering tokens at a discounted rate.
Ramp’s lead economist, Ara Kharazian, anticipates American companies must innovate to meet business needs currently unfulfilled by their offerings. While Chinese model rise reflects business demands, he believes U.S. model companies will eventually counter competitively.
However, Gordner hints that American firms may face price increases as public offerings draw near. Both Anthropic and OpenAI have submitted confidential IPO paperwork.
NPR acknowledges Anthropic as a financial supporter.
Leave a Reply