American AI is expensive. Some startups are turning to cheap Chinese models

A growing number of startups are quietly swapping out American artificial intelligence systems for cheaper alternatives built in China, as the cost of running AI tools becomes one of the fastest-rising line items on their budgets. The shift reflects a broader reckoning inside young companies that adopted generative AI early and are now confronting just how expensive it is to keep using it at scale.

For many young companies, AI has moved from an experimental perk to a core operating cost, embedded in customer support, coding tools, content generation and internal workflows. As usage climbs, so do the bills, since most leading systems charge based on how much text or data a company processes rather than a flat subscription fee. That usage-based pricing can turn a modest pilot project into a significant monthly expense once a product scales to thousands or millions of users.

Chinese developers have spent the past year and a half releasing AI systems that rival top American models on many tasks while costing a fraction as much to run. Firms such as DeepSeek and Alibaba have published rival systems, some as open-weight software that outside developers can download, modify and operate on their own servers rather than paying per query to a single provider. That approach strips out much of the markup built into subscription-style access to proprietary systems from companies like OpenAI, Google and Anthropic.

For cash-conscious founders, the appeal is straightforward: comparable performance at a lower price extends a startup's runway and improves margins on AI-dependent products. Some companies described treating model selection less like a loyalty decision and more like a commodity purchase, routing tasks to whichever system delivers acceptable results most cheaply, switching providers as pricing and capability shift.

The trend is not without friction. Businesses weighing Chinese-made models must also consider where data is processed and stored, given heightened scrutiny in Washington over Chinese technology and concerns that sensitive user information could be exposed to foreign servers or content restrictions embedded in models developed under Chinese regulation. Some startups are addressing that by running open-weight Chinese models on their own infrastructure rather than sending data to servers based in China, a workaround that preserves the cost savings while sidestepping some data-transfer concerns.

The cost gap has not gone unnoticed by the dominant American AI companies, several of which have introduced cheaper, lighter-weight versions of their flagship systems in response to competitive pressure. That dynamic mirrors a wider pattern in the AI industry, where rapid improvements in cheaper, openly available models are pushing prices down across the board even as demand for AI tools keeps climbing.

The competition is unfolding against a backdrop of U.S. export restrictions meant to limit China's access to advanced chips used to train the largest AI systems. Those controls have not stopped Chinese labs from producing software that performs competitively on many everyday tasks, underscoring that the race over AI costs is playing out at the software layer even as hardware remains a geopolitical flashpoint. How durable the price advantage proves, and whether regulatory or security concerns slow adoption further, remains an open question as more startups test the waters.

Source: NPR World