[Edaily Reporter Bang Sung Hoon ] As artificial intelligence (AI) costs soar, U.S. companies are switching from the cutting-edge models of OpenAI and Anthropic to cheaper “OpenWeight” models. This trend is spreading beyond Silicon Valley to sectors such as finance, logistics, and telecommunications, and it is drawing attention because Chinese companies hold the lead in this market.

According to the Financial Times (FT) on the 27th (local time), the research platform AlphaSense reported that the number of times executives mentioned “open-weight” or “open-source” models during earnings calls and investor conferences from August to September this year surged sixfold compared to the same period last year. Not only tech companies but also PNC Financial Services, logistics firm CH Robinson, and Germany’s Siemens have recently discussed adopting these models.

Changes are also evident in the usage of intermediary services that allow users to select and use various AI models in one place. The share of open-weight models among all tokens processed by U.S. cloud provider Versel’s “AI Gateway” jumped from 7% in December of last year to 56% in August of this year. On another intermediary platform, OpenRouter, Chinese open-weight models dominated the top 10 models by processing volume.

Open-weight models are those that make public the parameters that determine how the AI responds. Since companies can download them, run them on their own servers, and modify them, they do not have to pay per token (the unit of data processed by AI), and their performance is often comparable to that of closed-source models.

Chinese companies have dominated this market. Firms like DeepSeek and Zipu generate revenue by selling tokens directly at prices far lower than those of OpenAI and Anthropic. Some companies distribute models for free to lock users into specific hardware or to sell advertising. France’s Mistral, as well as U.S.-based NVIDIA, Reflection AI, and Thinking Machines Lab, have also joined the development race.

The biggest reason is cost. Vinay Kuruvilla, Chief Technology Officer (CTO) of Tinder—a dating app owned by Match Group—said, “Our AI spending, which was around $1 million (approximately 1.36 billion won) annually in January, ballooned to $10 million (approximately 13.6 billion won) in July. “We don’t want to see it increase tenfold again,” he said. Tinder has begun routing some queries from non-technical users to OpenWeight models. He added, “State-of-the-art models like OpenAI’s Astra and Claude Fable are already smart enough to handle 90% of our work. Once OpenWeight catches up, we may no longer need them.”

AT&T is currently running about 40% of its AI workloads on open models and plans to increase that to 70% within a year. Andy Marcus, Chief Data and AI Officer, said, “At a scale of processing 45 billion tokens a day, cost is extremely important. If we can operate much more cheaply while maintaining accuracy, we’ll make the switch,” adding, “Open models are getting better and better, so our options are expanding.” AT&T is fine-tuning open models using its own data to achieve performance equal to or better than proprietary models for specific tasks.

However, cost isn’t the only reason for the shift. Data center provider Digital Realty operates an OpenWeight-based internal chat service on its own infrastructure for reasons of data sovereignty and security. It mixes open and closed models depending on the sensitivity and complexity of the task.

Scott Wallace, Global Senior Director of Solutions Architecture, emphasized, “We’re putting our own data—which we wouldn’t put into state-of-the-art models—into these private models. We absolutely, absolutely, absolutely never put customer data into state-of-the-art models.”

Analysts suggest this trend could threaten the revenue growth of Anthropic—which is preparing for an initial public offering (IPO) with a target market capitalization of over $2 trillion (approximately 2,718 trillion won)—and OpenAI, which has postponed its IPO until next year and is seeking investment at a valuation of $1.2 trillion (approximately 1,631 trillion won).

This is why investors are closely watching how companies allocate their AI spending. For now, these two companies account for the majority of AI spending, and their massive computing demands are driving the data center construction boom that underpins the U.S. economy. However, as the price war with Chinese companies intensified, both firms released lower-cost versions of their flagship models last week.