# Why AT&T Routes 40% of AI Requests Through Open Weights

> AT&T doubled open-model routing to 40% to hold vendor spend flat, but practitioners argue the percentage hides which tasks still need frontier models.

Published: 2026-09-06
URL: https://daniliants.com/insights/why-at-t-routes-40-percent-of-ai-requests-to-open-weights/
Tags: open-weights, enterprise-ai, cost-optimization, llm

---

## Summary

AT&T now routes 40% of its internal AI requests through open-weights models, up from 20% in May 2026, and expects 60% within months, explicitly to hold its Anthropic and OpenAI spend flat while employee usage keeps climbing. The New York Times frames this as a broad corporate shift away from frontier vendors. The Hacker News discussion around it is more useful than the article: practitioners converge on the point that cost is the stated reason but vendor risk is the real one, and that the win comes from tiering tasks, not from replacing frontier models wholesale.

## Key Insight

**The hard numbers from the article**

- AT&T runs AI for roughly 100,000 employees. Open models went from 20% in May 2026 to 40% now, with 60% projected in the coming months and 60-70% over a few years.
- Original workloads sent to Anthropic and OpenAI: customer service, call transcription, and coding.
- The stated goal is not to cut the AI bill but to keep it flat while usage grows, a ceiling strategy rather than a cut.
- AT&T researches Chinese models but does not deploy them, citing regulation and data-privacy concerns. It uses Google's Gemma and Meta's Llama instead.
- Reported quality position: open models are "on par with, and in some cases better than, older versions of the frontier closed models".

**What the practitioners add that the article misses**

- **The percentage is a composition trick, not a capability claim.** "40% of AI use" is unweighted by task. Call transcription is trivially open and is almost certainly the bulk of AT&T's request volume, while coding, the expensive and hard workload, likely stayed frontier. A high open-model percentage can coexist with frontier models doing all the work that matters.
- **Risk, not cost, is the buying driver.** A private-equity tech-diligence assessor reports that "how easily can they change models?" is now a standard diligence question, and that investors dislike dependence on vendors that are not trustworthy, predictable, financially safe, or capable of hitting three nines. The threshold offered: at 80-90% of frontier capability, almost as good with far less risk is the better deal.
- **The strongest counter-argument is to measure cost per task, not per token.** Open models add router complexity and security surface, sending company data through more and less trustworthy providers, and the Pareto curve is still owned by closed models. The interesting conclusion is that you capture most of the benefit of open models without adopting them, because their existence forces frontier prices down.
- **Trust and paternalism are a procurement factor.** Refusal behaviour on benign tasks is itself a business risk: if a vendor's policy layer can decline a translation today, it can decline your proprietary workflow tomorrow. A related theme is that labs competing with their own customers make enterprises reluctant to outsource core intelligence.
- **The credible bear case** calls this a narrative blip comparable to the RAG push, with legacy enterprises signalling cost-responsible innovation to Wall Street and the projects dying once operational overhead bites. Worth holding as the null hypothesis.
- **A terminology objection worth adopting:** these are open-weights, not open-source. There is no source to read, patch, or learn from. Using the precise term avoids importing open-source assumptions about auditability that do not hold.
- **Legal indemnity is a real blocker.** Corporations want someone to sue. An open model from a US company gives some indemnity cover, and that, more than benchmark scores, is why Gemma and Llama beat Chinese models here.