AT&T Bets on Open-Source AI, Targeting 70%-80% Adoption
Nashnova编辑部
AT&T is shifting its AI workload from proprietary models to open-source, targeting 70%–80% of total usage — up from 25% today — after seeing 80%–90% cost cuts in some cases. This may be the clearest pressure signal yet for closed-model vendors' enterprise business.
How much AI does AT&T burn daily — and why switch?
AT&T processes roughly 45 billion AI tokens per day — a scale where model choice directly determines the bill.
This means → even a tiny per-token price gap, multiplied by 45 billion, becomes a massive expense. Picking cheaper models is not penny-pinching — it is survival-grade cost management.
Chief Data and AI Officer Andy Markus calls this heavy consumption "token doomsday," but says the company sees it as a manageable problem.
How much cheaper are open-source models, exactly?
Markus says switching from proprietary to open-source in specific use cases has cut costs by 80%–90%.
To auto-match each task to the most cost-effective model, AT&T built a "smart router" — it assigns different tasks to different models automatically.
In plain terms = not every task needs the priciest model. Simple jobs go to cheap models; complex ones get the expensive ones. The router makes that call for you.
Which core operations has open-source already taken over?
AT&T runs over 1,000 AI use cases internally, spanning legal, finance, field technician support, and network operations.
Customer-call summarization and analysis is now handled entirely by open-source models — not a fringe experiment, but a core business function.
For network management, AT&T customized an open-source model called OTel using proprietary telecom data. It powers AI agents — automated programs that diagnose root causes of network faults.
How does AT&T handle data security and vendor lock-in?
Open-source models can run in AT&T's own data centers, with no need to rent external cloud infrastructure — data stays in-house, and costs drop further.
Markus says the company may adopt a hybrid architecture: cloud-based AI models alongside open-source models on its own hardware.
AT&T is also testing multiple open-source models, including ones from China, specifically to avoid dependence on any single vendor.
What does this mean for the broader AI industry?
Gartner analyst Chirag Dekate notes that many enterprises hit a cost-unsustainability tipping point after adopting closed frontier models, then shift to open-source — driven also by vendor lock-in and IP concerns.
Gartner forecasts: within two years, open-source models will power over 50% of enterprise AI use cases — up from under 10% today.
This reflects a broader industry inflection, not an AT&T-only story. If this path becomes mainstream, closed-model vendors face real pressure on their enterprise market share.
Content is for reference only, not financial advice.