AT&T’s 70%–80% Open-Model Target Needs a Switchboard

AT&T handles an average of 45 billion AI tokens a day and says open models power about 25% of its overall AI usage. It wants that share to reach 70%–80% while controlling token costs and protecting proprietary data. Which tasks can make the move?

In 2021, more than 500 researchers were building an open-source language model for research independent of any company. That project’s purpose was access to research, not a telecom’s daily operating budget. Five years later, enterprises are making model choices against bills, data policies and service requirements. “Open” describes access to the weights; it does not settle what running them costs or who answers when they fail.

AT&T’s workload mix is more revealing than a model ranking

In an August Wall Street Journal account, AT&T cited token-cost control and protection of proprietary data as reasons for shifting toward open models. Its prospective 70%–80% share is a target, not a measured result; its usage percentage cannot be converted into a token count without knowing how it measures usage.

The September coverage record put open models at 56% of Vercel’s tokens in August and attributed to AlphaSense a sixfold year-over-year rise in open-model mentions on US earnings calls and at conferences in August and September. Vercel’s tokens measure activity on one platform, while mentions measure discussion, not deployment. Neither establishes an enterprise-wide market share. Together with AT&T’s account, they show why buyers now have to evaluate open models as production options rather than as research artifacts.

Harvey, Abridge, Ramp and Rogo have likewise adopted open-weight models or trained their own to reduce costly reliance on frontier labs as their usage—and inference bills—grow. But a downloadable model is not necessarily cheap to serve. A model that consumes more tokens, needs expensive hardware or fails a task more often can lose its apparent price advantage. Buyers should compare the cost of successfully completed tasks across models.

The work around a model outlasts a model choice

AWS made a multi-supplier choice visible in 2023 when it launched Bedrock with models from Anthropic, Stability AI, AI21 Labs and AWS. That platform let customers select among suppliers without treating one provider’s model as the entire application. In 2025, AI21 Labs added an orchestration system aimed at reducing hallucinations and improving reliability alongside its own model-building work. Those are different businesses, but both address the work that remains after a buyer gains access to a model.

Databricks illustrates another part of that work. In August, the company said it had crossed a $7 billion revenue run rate, while its CEO argued that enterprise agents need organizational context unavailable to a general model. A buyer must still identify the source system for that context, control who can retrieve it and test whether a model uses it correctly. Databricks’ reported run rate does not measure revenue from model routing; its argument identifies the data dependency that a model swap cannot remove.

A production team therefore needs a record for each task: which models meet its quality threshold, what each costs to serve, what data each may receive and what happens when the preferred model is unavailable. The team also needs monitoring after deployment, not just a benchmark before it. That work can make an application more productive, but enterprise adoption may follow a J-curve: investment in data, infrastructure and organizational change can precede visible returns.

Private deployment changes who must answer for a failure

AT&T’s proprietary-data concern helps explain the appeal of hosting a model within a buyer’s own controls. Huawei-led vendors were already selling on-premises “AI-in-a-box” systems to Chinese companies in 2024. Hosting gives a buyer more say over access, retention and updates; it also gives that buyer more to operate and investigate.

Chip Huyen notes in AI Engineering that open-model users may carry data-lineage and copyright exposure that a commercial provider’s contract would otherwise help cover. Private hosting cannot establish where a model’s training data came from. Buyers also need evaluations, incident response and a defensible account of which version handled a sensitive request. Those obligations are deployment accountability.

Hosted suppliers face the same buyer demand from the other side. OpenAI is testing Private Safety Processing to detect misuse while preserving zero-data-retention protections for early customers. A buyer comparing that service with a self-hosted model is comparing two governance arrangements, not simply “closed” against “open.”

A router needs task evidence, not faith in interchangeability

Artificial Analysis, Collinear, IBM, Nvidia and Vercel formed the Cyber Index Alliance to evaluate how AI agents find and fix vulnerabilities. Its CWE-Bench-AA includes 120 held-out patching tasks across ten OWASP Top 10 categories and six programming-language families. Such tests give a buyer a way to ask whether a candidate works on a specified job, rather than whether it sits near the top of a general leaderboard.

The limit matters. An analysis of cyber capabilities found recent open-weight models lagging closed frontier systems by four to seven months, though the reported gap had narrowed from six to ten months through much of 2025. A buyer cannot assume that a cheaper open model can replace a frontier model on a sensitive cyber task. A useful router must be allowed to select the closed model—or reject the task—when an open candidate misses the required threshold.

Supply policy narrows the choice set too. OpenRouter reported that Chinese models drew more than 30% of token usage by US companies on its service each week since February 8, up from 11% over the preceding 12 months. That is OpenRouter usage, not a census of US businesses. It does show why a rule excluding Chinese weights would affect buyers already using that channel. The Qwen open-weight ecosystem makes the procurement question especially concrete: a model’s license, provenance and security approval can matter as much as its task score.

To pursue its proposed 70%–80% mix, AT&T must show which tasks an approved open model can handle within its quality, cost and data-policy requirements. Where no candidate does, access to the weights changes nothing; where several do, AT&T can change suppliers without rebuilding the work.