In recent years, AT&T has relied on AI models from Anthropic and OpenAI to help with customer service, call transcription and programming. These models are paid and "closed" — the AI companies don't share their underlying code. But this year, AT&T changed course.
As AI costs soared, the company's chief data and AI officer, Andy Markus, began looking for cheaper alternatives. He chose open models — ones that can be downloaded and modified for free, without payment or approval. By May, open models accounted for 20% of AT&T's AI usage; Markus said in an interview that the figure has since risen to 40% and could jump to 60% in the coming months. "We believe the number could go higher," he said, adding that AT&T is now saving up to 80% on its AI costs compared with earlier this year.
Open-source AI has become a hot trend across corporate America. Like AT&T, companies including Airbnb and Deloitte are shifting to open models that are easier to customize and cheaper to use. According to data from OpenRouter — a platform that lets users choose different AI models for tasks — open models accounted for 58% of US AI usage last month, up from just 10% a year ago.
On Thursday, chipmaker Nvidia announced a $12.9 billion acquisition of Hugging Face, the open AI model hub, underscoring the technology's importance. "We will work together to make AI more open, more powerful, and more accessible to individuals and institutions around the world," Nvidia CEO Jensen Huang said.
Open models come in two varieties: open-source models, which make their underlying code public, and open-weight models, which publish their "weights" — the numerical parameters that determine how a model reasons. Many of the most popular open models are built by Chinese companies such as Moonshot AI, DeepSeek and Alibaba. Developers don't need to spend millions on closed models; they can download one of these Chinese systems, build products on top of it, and mainly pay for server costs. OpenAI and Anthropic, by contrast, charge subscription and usage fees per task.
Jerry Tang, CEO of Atlas Cloud, a startup offering a variety of AI models, said some Chinese open models now reach 80% to 90% of the performance of OpenAI and Anthropic models at just 20% of the cost. He compared the leading models from OpenAI and Anthropic to expensive sports cars — powerful, but not always necessary. "A Mazda will get you to your destination just fine, and the Maserati isn't a great investment," he said. Open models, he added, "offer better value."
The rise of open models could threaten Anthropic and OpenAI, both of which are preparing massive initial public offerings. Having poured billions into R&D and computing, they must charge more. The trend has also ignited a broader debate in Silicon Valley and Washington over who controls AI: Anthropic CEO Dario Amodei argues AI should be strictly regulated to prevent national security risks, while Nvidia's Huang and Meta's Zuckerberg say open-source technology is essential to a democratic, balanced AI ecosystem. As companies vote with their wallets, the "open vs. closed" debate is turning from ideology into a real fight for market share.