Corporate America is getting addicted to open source AI
AT&T has turned to artificial intelligence models from Anthropic and OpenAI in recent years to help with customer service, call transcription and coding. The AI models charged a fee and were “closed”, which meant that AI companies did not share their underlying code.
This year, AT&T changed course.
As the cost of AI skyrocketed, Andy Markus, the company’s director of data and AI, looked for cheaper alternatives. He landed on “open” models that can be downloaded and modified without payment or approval.
By May, open models accounted for 20 percent of AT&T’s AI usage. That has since risen to 40 percent and may jump to 60 percent in the coming months, Mr. Markus said in an interview.
“We believe it could go much, much higher,” he said, adding that AT&T has saved up to 80 percent of AI costs compared to earlier this year.
Open-source artificial intelligence has become all the rage in corporate America. Like AT&T, companies such as Airbnb and Deloitte are also turning to open AI models that are easy to customize and cheaper to use. Last month, open models accounted for 58 percent of AI usage, up from 10 percent a year ago, according to US user data from OpenRouter, a platform that lets people choose different AI models to complete tasks.
On Thursday, chipmaker Nvidia announced it was buying Hugging Face, a library of open artificial intelligence models, for $12.9 billion, another sign of the technology’s importance.
“Together, we will make AI more open, capable and accessible to people and institutions around the world,” said Jensen Huang, chief executive of Nvidia.
Open models come in two variants. One is an open source model that publishes its core code. The second is an open weights model that publishes its “weights”—numerical calculations that determine how the model reasons.
Many of the most popular open AI models are produced by Chinese companies such as Moonshot AI, DeepSeek, and Alibaba. Rather than spending millions in fees to use a closed model, developers can download one of China’s systems and build their products on it, paying mainly for the servers. OpenAI and Anthropic charge subscriptions and other fees depending on the work.
Some Chinese open source models are now 80 to 90 percent more powerful than those from OpenAI and Anthropic while costing just 20 percent of the price, said Jerry Tang, chief executive of Atlas Cloud, a start-up that provides access to various artificial intelligence models. He compared leading models from OpenAI and Anthropic to expensive sports cars – powerful but not always necessary.
“Mazda does a good job of getting you where you need to go, and Maserati just isn’t worth the investment,” Mr Tang said. Open-source models, he added, “offer better bang for your buck.”
The popularity of open models could spell trouble for Anthropic and OpenAI, both of which are headed for successful initial public offerings. Companies have to charge more because they spend billions on R&D and computing power.
OpenAI and Anthropic did not immediately respond to requests for comment. (The New York Times has sued OpenAI and Microsoft, alleging copyright infringement of news content related to AI systems. The companies have denied the suit’s claims.)
The rise of open models plays into the broader debate in Silicon Valley and Washington about control over AI Dario Amodei, chief executive of Anthropic, argued that AI should be tightly regulated to avoid national security concerns. Other tech leaders, such as Mr. Huang of Nvidia and Mark Zuckerberg, chief executive of Meta, have said that open source technology is essential to a democratic and balanced artificial intelligence ecosystem.
After OpenAI released ChatGPT in 2022, many companies and users gravitated towards closed AI models. At the time, open-source models were popular mainly among developers.
The models started gaining traction last year after Chinese startup DeepSeek released one called the DeepSeek-V3. It was characterized by being almost as high-end as the leading closed models, while requiring less processing power. Since then, open models from Chinese companies have increasingly caught up with the performance of closed models from Anthropic, OpenAI, Google and others.
Many open models can now be powered by a phone or laptop, said David Stout, chief executive of artificial intelligence startup WebAI. This makes the technology more accessible and even cheaper. Front closed models are increasingly complicated and require more computing power.
Most American companies still use a combination of open and closed models. Closed models remain best for demanding tasks such as encoding and image and video generation, while open models often excel at simple specialized tasks, said Atlas Cloud’s Mr. Tang.
At AT&T, the telco is now using open scale models and adapting them to create new tools that can transcribe calls and facilitate customer service, said Mr. Markus, the company’s chief data and AI officer.
Some US firms are reluctant to use Chinese AI models due to regulatory and privacy concerns. AT&T is exploring Chinese models but not using them, Mr. Markus said. Instead, it works with popular alternatives made by American companies, such as Google’s Gemma AI model and Meta’s Llama AI models.
As Chinese models become more popular, some American companies are unveiling new open AI models of their own. Last month, Meta made its flagship AI system an open-weight model and said it would offer the model at a lower price than other high-end US models.
“I do not believe that restricting access to foreign open source models is an effective solution,” Mr. Zuckerberg said in a statement at the time. “Our goal should be to make American open-source models the best globally.”