Microsoft's Phi-4 helped AT&T build telecom AI while slashing infrastructure costs

The US telecom giant, AT&T, has now revealed how it built its next-generation telecom AI using Microsoft’s Foundry platform, processing enormous amounts of data while significantly lowering development costs. The project also highlights Microsoft’s growing push into enterprise AI infrastructure.

Microsoft Foundry helped power OTel 2.0

AT&T developed Open Telco 2.0 (OTel 2.0) to better understand telecom networks, standards, and operations. Rather than relying on one AI model, the company combined several open models for different tasks.

Those included Microsoft Phi-4, OSS 120B, and Gemma 4. Phi-4 handled around 700 billion tokens every month, mainly for synthetic data generation and data preparation. Overall, AT&T processed roughly 1 trillion tokens and trained the model using around 400 billion tokens.

To keep everything running, Microsoft Foundry Managed Compute supplied access to about 530 GPUs, including 430 AMD Instinct MI300X GPUs.

Lower AI costs without sacrificing scale

AT&T says using open models through Microsoft Foundry dramatically reduced AI development expenses. According to the company, generating data with Phi-4 and other open models saved tens of millions of dollars compared to frontier AI models.

Speaking of scale, the company says OTel 1.0 has already surpassed 25 million downloads, showing growing interest in telecom-focused AI. Microsoft says Foundry lets organizations choose different AI models, mix GPU hardware, and deploy workloads much faster.