AT&T says open models currently power ~25% of its overall AI usage, and expects that to rise to 70%-80% over time; AT&T uses an average of 45B AI tokens per day
Context & Ripple Effects
AT&T’s stated shift toward open models lands as enterprise model use becomes less concentrated: a Global 2000 survey found OpenAI was widely deployed in production, while OpenRouter reported substantial US-company token use from Chinese models. AT&T’s 45 billion daily tokens make its intended mix change a meaningful procurement signal rather than a limited pilot.
The report adds a concrete enterprise-scale example to the broader move toward multi-model sourcing, where buyers can direct workloads among model suppliers instead of standardizing on one provider.
First-order effects
- AT&T intends to move a majority of its AI workload toward open models over time, up from roughly one-quarter today, changing the model mix behind its reported daily token demand.
- Proprietary-model providers serving AT&T face a customer signaling that openness will increasingly matter in workload allocation, alongside model performance and operating cost.
Second-order effects
- AT&T’s planned diversification strengthens large enterprises’ negotiating leverage with model vendors, because high-volume workloads can be shifted among alternatives rather than committed to a single supplier.
- Providers of open models and the infrastructure used to run them gain a clearer path to compete for enterprise workloads that proprietary APIs currently serve.
Third-order effects
- If other large enterprises follow AT&T’s approach, enterprise AI purchasing may evolve into a portfolio market in which model vendors compete continuously for workload share rather than win durable single-provider standards.
- The pattern would shift value toward the tools and infrastructure that let buyers evaluate, route, and operate multiple models at scale.
The trend: Large enterprise AI buyers are moving toward multi-model sourcing, using open models to increase choice and bargaining power at production-scale token volumes.