Why AT&T Is Betting Big on Open-Weight AI－By Belle Lin, by WSJ

Corporate America’s AI playbook is changing as more businesses substitute proprietary AI models in favor of open alternatives to save on costs. AT&T’s effort to remake its business for the AI era shows just how far that shift has come.

Such open models currently power about 25% of the telecommunications company’s overall AI usage, including helping manage its network, said AT&T Chief Data and AI Officer Andy Markus, “and we believe we can get way higher than that.”

Over time, he expects open models to power 70% to 80% of the company’s total AI usage.

That conviction has driven Markus to experiment with a number of different open models, including powerful new models from China , which help prevent AT&T from becoming “beholden to any one solution,” he said.

The company’s focus on using open models comes at a time when U.S. companies are starting to put tighter controls around their AI spending. Companies say one way to lower AI token costs is to use open models rather than state-of-the-art proprietary ones, because they generally offer more flexibility and lower cost.

Silicon Valley meanwhile remains locked in heated debate over whether AI should be “open” or “closed.” Nvidia is among the most prominent tech giants backing open models and Meta Platforms on Monday announced a new open-weight model , Muse Glimmer.

Proprietary models from companies like OpenAI, Anthropic or Google aren’t freely available for users to download and modify, while open-source and open-weight models are. True open-source models allow full access to training data and code, while open-weight models typically share only the numerical parameters, or “weights,” that underlie them.

AT&T uses both open-source and open-weight models.

Facing down the ‘token apocalypse’

Because AT&T uses an average of 45 billion AI tokens each day, directing user prompts toward cheaper models makes an impact, Markus said. The company built what it calls a “smart router” to automatically pick the most cost-effective model for a specific task.

Switching from closed, proprietary AI models to open models has already resulted in savings of 80% to 90% for AT&T in certain applications, he said.

“You hear the stories of the ‘token apocalypse,’ and we’re not scared of the token future,” he said. “It’s something that we feel like we can manage.”

AT&T has over a thousand internal uses for AI, from supporting back-office functions like legal and finance to assisting field technicians and running its core network operations. Summarizing and analyzing customer service call transcripts—what Markus describes as an intensive process—is supported entirely by open models, he said.

The work of managing AT&T’s sprawling network is also done by open models. OTel, an open model that AT&T customized using telecom-specific data, supports AI agents that can detect the root cause of network issues, according to Markus.

Another benefit of open models is that they can be run on AT&T’s own data centers rather than rented infrastructure from a cloud-computing provider—a setup that trims AI costs even further. In the future, Markus says the telecom will likely use a mix of cloud-hosted AI models and open models that run on its own hardware.

Although many enterprises say they need the premium capabilities offered by proprietary frontier models, they soon reach an inflection point where the costs of such models become unsustainable, said Chirag Dekate, an analyst at market research and IT consulting firm Gartner.

Plus, they don’t want to risk it all with a single AI vendor, and feel the need to guard their intellectual property from those AI labs, Dekate said.

Over the next two years, open models will underpin more than 50% of use cases for businesses, up from less than 10% today, according to Gartner.

AI sovereignty

Critics have argued that open models—many of which come from China—present potential risks to national security.

For AT&T, however, using open models is nonnegotiable because of what Markus described as data protection and intellectual property concerns. Open models are generally preferred for data protection because they offer more control over the flow and access of corporate data.

“We have to keep our information safe and protected, and we have to control the inputs and the outputs of the models,” he said. “The enterprise data is the gold mine, and the tools are just a way to mine the gold.”

Enterprises have long sought to keep their proprietary business data away from tech vendors, aiming to keep a strict hold over what they consider their most important assets. The same thinking applies to AI vendors, who train their models on vast amounts of data.

While labs like OpenAI and Anthropic have said enterprise customer data isn’t used for training their models, some companies fear that’s not enough.

“As our data flows through the model ecosystem, we just want to make sure that it is safe and secure from a security standpoint, but also an IP standpoint,” Markus said. “The concept of AI sovereignty has become truly paramount to us.”