"U.S. companies turning to open AI models to cut costs"

[Financial News] U.S. companies are turning to low-cost open-weight AI models as they seek alternatives to Anthropic and OpenAI to reduce their enormous cost burdens, the Financial Times (FT) reported on the 27th, local time.
The trend is spreading across the U.S. industrial sector, beyond Silicon Valley, where IT companies are concentrated.
According to research platform AlphaSense, executives mentioned open-weight and open-source models six times more often during corporate earnings announcements and investor conferences in August and September than they did during the same period last year.
Companies are seeking open-weight models because of their cost competitiveness.
Open-weight models are a type of open model whose publicly available AI-model weights can be downloaded by companies and customized to run on their own hardware.
They cost far less than paying to use premium versions of ChatGPT or Claude. Open-weight models can also be offered for free to lock users into specific hardware or to sell advertising.

With Chinese companies such as DeepSeek dominating the market, the influence of Chinese firms including DeepSeek and Z.ai is growing in the open-weight market. France's Mistral AI and U.S. companies such as NVIDIA, Reflection AI and Thinking Machines Lab have also released models based on a similar approach, but the market is currently being led by Chinese companies, FT reported.
The adoption of open-weight models is spreading beyond the technology industry to telecommunications, finance, logistics and manufacturing. AT&T currently handles about 40% of its AI workload with open models and plans to raise that share to 70% within the next year.
AT&T explained that the cost savings are significant because it processes 45 billion tokens a day. Fine-tuning open models with proprietary data can deliver performance comparable to or better than that of closed models for specific tasks, the company said.
Dating app Tinder is also shifting some tasks to open-weight models because of the cost burden. Tinder's annualized AI spending rose tenfold, from about $1 million in January to $10 million in July.
Security, in addition to cost, is cited as another reason for choosing open-weight models. Operating models on internal corporate infrastructure allows companies to process customer information and sensitive internal data without sending it to external AI providers.
This shift is expected to create new competitive pressure for OpenAI and Anthropic. Although a substantial portion of corporate AI spending still goes to the two companies' most advanced models, the rapidly improving performance of affordable open-weight models could reduce companies' need to use expensive models.
FT reported that companies are increasingly seeking to reduce AI costs by using high-performance models for complex tasks and less expensive open-weight models for repetitive, large-scale workloads.


As a result, OpenAI and Anthropic have urgently launched lower-priced models and entered the competition.
Open-weight models accounted for 7% of all tokens processed through the AI Gateway of AI development platform Vercel in December last year, but that share surged to 56% in August this year.
Vinay Kuruvila, Tinder's chief technology officer, said, "AI spending, which was $1 million annually in January, soared to $10 million in July, so we began shifting some inquiries from nontechnical users to open models," adding, "If the performance of open-weight models catches up with high-performance frontier models such as OpenAI's Astra or Anthropic's Claude Fable, there will be no reason to use expensive closed models."
Andy Markus, AT&T's chief data and AI officer, also said, "We currently process 40% of our AI workloads with open models and plan to raise that figure to 70% within a year."
FT noted that, beyond cost savings, securing data sovereignty and addressing security concerns are also important factors pushing companies toward open models.
Scott Wallace, a senior director at global data center operator Digital Realty, said, "We never enter customer data or sensitive, complex internal confidential information into external frontier models," adding, "We addressed security concerns by building an internal chat interface on private infrastructure based on an open-weight model."
FT assessed that "the rapid advancement of affordable AI systems could threaten the revenue of OpenAI and Anthropic, which have driven a data center boom with massive investments."

[email protected] Lee Seok-woo, international affairs correspondent Reporter