Red Hat is telling mobile operators to treat AI-RAN as a staged network upgrade, not a blanket GPU replacement for every radio site.

RCR Wireless News reported the comments from Shujaur Mufti, Red Hat's director of telco ecosystem solution architecture, during its Telco AI Forum.

The near-term focus is AI for RAN, where operators apply models to existing radio networks for operating cost, energy, spectral-efficiency and fault-detection gains.

Mufti framed early adoption around operational proof before operators move AI workloads and radio functions onto shared infrastructure.

Red Hat Sees AI For RAN Leading Through 2027

Mufti described three stages for AI-RAN.

The first uses AI to improve current radio access networks, the second combines AI and RAN workloads on common infrastructure, and the final stage makes the RAN a platform for AI-native services.

AI for RAN is expected to remain the main industry focus through roughly 2027, according to Mufti.

He said operators can use existing RAN infrastructure for use cases such as energy savings, network efficiency, spectral efficiency and fault detection, rather than beginning with major radio upgrades.

Traditional self-organising network systems are also moving towards AI-enhanced SON platforms.

Service management and orchestration systems for Open RAN are adding AI-powered xApps and rApps that can manage both Open RAN and conventional radio networks.

Shared AI And RAN Tests Run Towards 2030

Between 2027 and 2030, Mufti expects operators to expand proof-of-concepts for shared AI and RAN infrastructure as 6G research develops and early standards emerge.

He named SoftBank and T-Mobile as operators already exploring the shared-infrastructure model.

The later phase would put AI into the mobile network itself alongside commercial 6G deployments after 2030.

In that model, the radio access network becomes a platform for AI-native applications and potential new revenue services, rather than only a connectivity layer.

Red Hat's work with SoftBank, Fujitsu and Nvidia was cited as early evidence for GPU-accelerated RAN trials.

Mufti said those deployments showed Layer 1 and Layer 2 radio functions can run without requiring a real-time kernel, while Red Hat has been extending its AI Grid initiative as a RAN-ready edge platform.

GPU Rollout Depends On Operator Economics

Mufti cautioned that operators should not assume GPUs will appear everywhere in the RAN.

He pointed instead to selected sites where AI inferencing at the network edge has a clear business case before radio workloads are added.

The economic threshold is central to the deployment sequence.

Mufti said AI-RAN has to improve network quality and create measurable business value for mobile operators before widespread rollout makes sense.

That leaves the commercial case narrower than the technical roadmap.

Red Hat has not named operator deployment dates, GPU site counts, signed commercial RAN contracts or measured revenue from AI-native network services.