AINEWS
AT&T routes 45 billion daily A.I. tokens through Meta's Llama and Google's Gemma, open source models, because cost, legal, and political math made them the only practical choice for office work.
AT&T now runs about 45 billion tokens a day through its internal AI assistant, up from roughly 8 billion in early 2026 and 27 billion in February 2026. A growing share of that traffic goes to open-weight models, AI systems whose trained parameters are published and downloadable, so any company can run them on its own hardware rather than renting access. AT&T's picks are Meta's Llama, Google's Gemma, and Nvidia's Nemotron. By one re-report of The New York Times's reporting, roughly 40% of those tokens already come from open-weight systems, with a stated target of 60% to 70% within a few years. A separate re-report pegs the present share closer to 25% and the ceiling at 80%. The two baselines are likely different windows or different metrics, but both point the same direction: open-weight is gaining, not retreating.
Open-weight models still trail frontier closed models by 6 to 10 months on the hardest tasks, AT&T has told reporters. They get the job because the other constraints have hardened first. Per-seat spend against a 60 to 85 thousand dollar salary is the obvious gate, as Hacker News readers mapped the math: roughly 2 to 3 thousand dollars a year in open-model spend versus 40 thousand or more for the frontier alternative. On code generation, AT&T has reported a 56% cost cut with only a 2% quality drop, while a separate re-report cites 80 to 90% savings in narrow workloads, according to BigGo Finance and Value Add's re-reports of the NYT piece. The numbers differ; the order of magnitude does not.
Frontier providers bundle indemnity, compliance, and SLAs that open-weight deployments have to assemble themselves. At the volume AT&T runs, that wrapper costs more than the model. The cheap path also runs partly on the company's own Nvidia and AMD servers rather than rented cloud compute, per BigGo, which makes the per-request cost a function of hardware depreciation rather than API markup.
AT&T actively researches Chinese open-weight models, DeepSeek and Moonshot, and does not deploy them. The reason given is security, compliance, and indemnity. In a corporate setting, "deployable" requires a chain of warranties that Chinese open-weight providers cannot extend, regardless of how the models benchmark. US open-weight models from Meta, Google, and Nvidia clear that gate because the providers are reachable in US courts and the licensing is unambiguous.
AT&T's internal dispatcher, described as a Lite LLM-style router, scores each request and assigns it to the cheapest capable model rather than defaulting to the most expensive one. As Hacker News readers noted in the wake of the NYT piece, the structural test for Anthropic and OpenAI is whether they can slash prices fast enough to stay in the routing mix. "Unless they both dramatically slash prices... they're in big trouble," one comment summarized the framing.
Menlo Ventures' State of Generative AI in the Enterprise report from December 2025 shows open-source's share of enterprise AI workloads falling from 19% to 11% over 2025, according to Value Add's re-report of the same chain. Menlo attributes the decline to licensing and governance friction, not to model quality. That is the strongest counter-signal to the AT&T story, and the way to reconcile it: AT&T is a leader, not the median. Most enterprises are still stuck in the licensing and governance mud. AT&T has built the routing and indemnity plumbing that lets open-weight models actually fire in production, and most companies have not.
The same week AT&T's deployment landed, Anthropic was reportedly preparing to file IPO paperwork as soon as the end of September, on the back of 11.5 billion dollars in quarterly revenue, up 14 times year over year. The open-weight pivot is now a direct margin test for the frontier providers, not a research curiosity. If the cheap middle layer absorbs the volume, the frontier layer has to either price closer to it or prove it is worth the premium on the tasks where it still wins.
For now, the boring office work has a name and a price tag: Llama, Gemma, Nemotron, running underneath a stack the headlines have not yet caught up to.