Large companies are shifting meaningful parts of their artificial intelligence work from closed, paid services to open-source models that can be downloaded and run without a vendor gatekeeper.
AT&T, which had been using models from Anthropic and OpenAI for customer service, transcription and coding, began switching after AI costs rose.
By May, open models accounted for 20% of its A.I. use; that share has since climbed to 40% and, the company told reporters, could reach 60% in the coming months.
The carrier said the move is producing big savings, as much as up to 80% versus earlier spending, and is training and routing models to improve telco-specific accuracy.
Practical, simple
The practical appeal is simple: open models can be downloaded and modified, making them cheaper to run at scale and easier to tailor to a company's internal data and processes.
That combination of lower cost and greater control is the reason other large buyers are following suit, the New York Times identified firms such as Airbnb and Deloitte among the early adopters, and an industry router service reports open models made up 58% of AI use last month, up from 10% a year earlier.
The shift is already attracting corporate capital. Chip-maker Nvidia announced it would buy the open-model repository Hugging Face for $12.9 billion, a bet that hosting and tooling around open weights will be central infrastructure as more firms self-host or fine-tune models.
Open-source models are not simply free plug-ins.
Fine tuned
Companies still need engineers to fine-tune, validate and secure deployments, and some workloads will keep using frontier closed models where specialised capability or vendor-managed safety is required.
But the balance of enterprise economics has moved: for tasks where accuracy and governance can be achieved in-house, open models are becoming the low-cost default.
“We believe it could go much, much higher,” Andy Markus, AT&T’s chief data and AI officer, said of open-model adoption.
This is not a fad; it is a shift in how corporate AI is being provisioned, cheaper, more controllable models are being turned into infrastructure rather than experiments.
