Story
September 7, 2026
Soaring AI Bills Are Driving Corporate America Toward Open Models
Corporate buyers see open AI as a way to escape punishing bills and tailor tools to their own needs, while closed-model leaders still hold an edge on the hardest jobs. The bargain is complicated by unease over data, regulation and the growing influence of Chinese model makers.
AT&T spent recent years using paid, closed models from Anthropic and OpenAI for customer service, call transcription and coding. But as its AI spending rose, the telecom began seeking systems it could download and modify without paying for access or asking permission.1
By May, open models made up 20% of AT&T’s AI use. That share has since reached 40% and could hit 60% within months, chief data and AI officer Andy Marcus said. The company has cut some AI costs by as much as 80% compared with earlier this year. “We think it could be higher,” Marcus said of open models’ eventual share.2
AT&T is not alone. Airbnb and Deloitte are also moving toward cheaper, more customizable open systems, while OpenRouter’s U.S. data showed open models accounting for 58% of usage last month, up from 10% a year earlier. Nvidia’s $12.9 billion purchase of open-model library Hugging Face underlined how seriously the industry is taking the shift. Nvidia chief executive Jensen Huang said the deal would make AI “more open, more powerful, and more accessible.”3
The economics are especially stark for Chinese-developed models from companies including DeepSeek, Moonshot AI and Alibaba. Atlas Cloud chief executive Jerry Tang said some can deliver 80% to 90% of the capability of OpenAI and Anthropic models at 20% of the cost. “A Mazda can get you to your destination just fine, and a Maserati doesn’t have a lot of investment value,” he said.2
Yet the turn is not a wholesale rejection of closed AI. Companies still favor proprietary systems for demanding coding, image and video work, while open models suit simpler, specialized tasks. And AT&T, wary of regulatory and data-privacy concerns, says it is studying Chinese models rather than using them; it has opted for U.S.-developed alternatives such as Google’s Gemma and Meta’s Llama.2
That leaves a widening strategic divide. Anthropic’s Dario Amodei argues for strict AI controls on national-security grounds; Huang and Meta’s Mark Zuckerberg say openness is essential to a democratic, balanced ecosystem. Zuckerberg’s answer to foreign competition is not restriction but better domestic models: “Our goal should be to make U.S. open-source models the best in the world.”2