AT&T has drawn a line in the sand, and OpenAI and Anthropic should probably take note. The US telecoms group wants to hold its spending on their models flat, even as its own use of artificial intelligence keeps climbing.
Its weapon is open-source software, the free-to-use models that anyone can download and run.
Leverage play
AT&T already runs about 100,000 employees on AI tools, for tasks from writing code to helping call-centre staff and handling HR paperwork.
That adds up to roughly 45 billion tokens a day, the units of text these systems chew through.
Bought entirely from frontier providers, that workload would cost hundreds of millions of dollars a year.
So the company now routes nearly 40% of it to open-source models, and wants to push that share towards 60% or 70%.
The trick is a model router, software that sends each request to a cheap open model or an expensive frontier one depending on how hard the task is.
Simple jobs, like summarising code, go to the budget option; the tricky work still goes to the best.
This is leverage dressed up as engineering.
Gap is closing
The old objection was that open models lagged too far behind to matter.
AT&T reckons that gap, once six to 10 months, is narrowing, and that for routine work the difference no longer shows.
It is customising models from Nvidia, Meta and Google, running them in its own data centres and testing them against internal benchmarks.
It is also eyeing Chinese models such as DeepSeek and Kimi, though not for real work yet, wary of political backlash from Washington.
If the performance gap keeps shrinking, the pricing power of the big labs shrinks with it.
A small puppy problem
Open source looks free, but it behaves like a puppy: cheap to acquire, costly to keep.
Running your own models means maintaining infrastructure, managing upgrades and doing the fiddly work of bending each model to your systems.
Those hidden costs can swallow the headline savings, and AT&T concedes its budget could yet rise if that burden grows.
The labs are not standing still either.
OpenAI and Anthropic have rushed out privacy protections and confidential computing to reassure nervous enterprises that their data will not linger.
A price war and a privacy war are being fought at once. The wider lesson is bigger than one phone company.
Enterprises are learning they hold more power than they thought, and the model makers now have to earn every dollar of the premium they charge.
That is a healthier market than one run by two suppliers. The direction of travel is what matters, and it points away from unquestioned dependence on a handful of labs.
Whether AT&T's sums add up is beside the point.
