Speaking during RCR Wireless News’ Telco AI Forum, Red Hat’s Shujaur Mufti said operators are initially focusing on AI for RAN because it delivers measurable operational benefits without requiring major upgrades to existing radio networks
In sum – what to know:
Operational gains – AI for RAN is leading adoption because it delivers immediate benefits—including lower operating costs, improved energy efficiency and better network performance—without requiring major RAN upgrades.
Shared infrastructure – Red Hat expects operators to gradually move toward common AI and RAN infrastructure, with proof-of-concepts accelerating through the remainder of the decade ahead of commercial 6G.
Economic value – AI-RAN deployments will expand only where they improve network quality and generate measurable business returns, making monetization and operational benefits the industry’s primary decision criteria.
AI-RAN will evolve through a phased transition spanning the rest of the decade, beginning with operational optimization before progressing toward shared AI and radio infrastructure and eventually enabling AI-native services across telecom networks, according to Shujaur Mufti, director of telco ecosystem solution architecture at Red Hat.
Speaking during RCR Wireless News’ Telco AI Forum, Mufti said operators are initially focusing on AI for RAN because it delivers measurable operational benefits without requiring major upgrades to existing radio networks. “I think AI for RAN is starting first because you can see, visualize the savings, for example, opex,” Mufti said.
He highlighted use cases including energy savings, network efficiency, spectral efficiency, and fault detection, noting that operators can deploy these capabilities on existing RAN infrastructure. “You don’t have to modernize your RAN network, and then you can get the benefits right there,” he said.
Mufti added that traditional self-organizing networks (SON) are evolving into AI-enhanced SON platforms, while service management and orchestration (SMO) systems for Open RAN are incorporating AI-powered xApps and rApps that can also manage conventional radio networks.
He described AI-RAN as a three-stage evolution. The first phase focuses on AI for RAN, followed by AI and RAN, where AI and radio workloads share common infrastructure. The final stage is AI on RAN, where the radio access network itself becomes a platform for AI-native applications and new revenue opportunities.
According to Mufti, AI for RAN will remain the industry’s primary focus through roughly 2027. Between 2027 and 2030, operators are expected to expand AI and RAN proof-of-concepts as 6G research matures and early standards emerge. He pointed to SoftBank and T-Mobile as operators already exploring this shared infrastructure model.
The final phase, expected after 2030 alongside commercial 6G deployments, would see AI becoming native across the entire mobile network.
Mufti also cautioned against assuming GPU acceleration will become universal across radio networks. Instead, operators are likely to begin with targeted deployments where the business case is strongest, particularly for AI inferencing at the network edge before introducing RAN workloads. “We should not think GPU everywhere in the RAN,” he said. “Maybe some selected sites as a starting point.”
Drawing on Red Hat’s work with SoftBank, Fujitsu and Nvidia, Mufti said early GPU-accelerated RAN deployments have already demonstrated technical advantages, including the ability to run Layer 1 and Layer 2 functions without requiring a real-time kernel.
He added that Red Hat has expanded its ecosystem collaborations around AI-RAN proof-of-concepts while extending its AI Grid initiative as a RAN-ready AI infrastructure platform at the edge.
While AI-RAN continues to gain momentum, Mufti said that widespread deployment will ultimately depend on demonstrating clear operational and financial value. “AI-RAN only makes sense if it has technology and economic benefits for the mobile operators,” he said.
He argued that operators will expand deployments only if AI-RAN improves network quality while creating new monetization opportunities. One possible approach is to begin with AI inferencing workloads at the edge, assess the revenue potential, and then determine how much GPU capacity should be allocated to radio functions.
Mufti concluded by encouraging operators to treat AI-RAN as part of a broader AI-native transformation rather than a standalone radio initiative.
Instead, operators should apply lessons learned from AI deployments across the core, OSS/BSS, and autonomous networks when designing future radio architectures. He argued that a common cloud-native platform and AI fabric spanning the data center, core, edge, and RAN will ultimately provide the operational consistency needed as telecom networks evolve toward AI-native infrastructure.
Facts Only
* Operators are initially focusing on AI for RAN because it delivers measurable operational benefits without requiring major radio network upgrades.
* AI for RAN yields immediate benefits including lower operating costs, improved energy efficiency, and better network performance.
* Use cases highlighted include energy savings, network efficiency, spectral efficiency, and fault detection deployed on existing RAN infrastructure.
* Traditional self-organizing networks (SON) are evolving into AI-enhanced SON platforms.
* Service management and orchestration (SMO) systems for Open RAN are incorporating AI-powered xApps and rApps to manage conventional radio networks.
* The evolution of AI-RAN is a three-stage process: AI for RAN, AI and RAN sharing common infrastructure, and AI on RAN where the radio access network acts as a platform for AI-native applications.
* AI for RAN is expected to remain the primary focus through approximately 2027.
* Between 2027 and 2030, operators are expected to expand AI and RAN proof-of-concepts as 6G research matures.
* Operators are exploring shared infrastructure models, with SoftBank and T-Mobile already investigating this model.
* Widespread deployment depends on demonstrating clear operational and financial value for mobile operators.
* Early GPU-accelerated RAN deployments have demonstrated technical advantages in running Layer 1 and Layer 2 functions without requiring a real-time kernel.
Executive Summary
Operators are currently prioritizing AI for RAN because it provides immediate operational benefits, such as reduced operating costs, improved energy efficiency, and better network performance, without necessitating major radio network upgrades. This focus stems from the ability to deploy these benefits directly on existing infrastructure. The transition is envisioned as a phased evolution over the next decade: first focusing on AI for RAN optimization, followed by shared AI and RAN infrastructure, and finally evolving toward AI-native services across the entire mobile network.
The economic viability of AI-RAN deployments hinges entirely on whether they improve network quality and generate measurable business returns, making these factors the primary drivers for adoption and monetization. Operators are expected to explore shared infrastructure models, with proof-of-concepts accelerating in anticipation of commercial 6G. Further deployment of GPU acceleration is anticipated to be targeted initially at edge AI inferencing before expanding to RAN workloads across the network.
Full Take
The narrative positions AI-RAN not as an end goal, but as an iterative journey rooted in incremental operational wins before achieving transformative architectural shifts. The framework of a three-stage evolution—AI for RAN leading to shared infrastructure and culminating in AI on RAN—suggests a necessary sequence where technological maturity must precede full deployment. This sequencing implies that ignoring the immediate operational gains (like OPEX reduction) might stall adoption, reinforcing the pragmatic focus on measurable value before pursuing high-level architectural concepts.
A significant tension exists between the technical possibility of integrating AI into radio networks and the practical barriers of infrastructure modernization and economic justification. The caution against universal GPU acceleration across the RAN suggests a necessary skepticism toward technological determinism; it implies that market readiness and demonstrable ROI must gate subsequent architectural leaps. This structure guards against a scenario where excitement over emerging technology leads to premature investment in capital-intensive upgrades without proven, value-generating use cases at the network edge.
The ultimate call for operators to view AI-RAN as part of a broader AI-native transformation, drawing lessons from core network functions (OSS/BSS), suggests that the radio layer is an implementation constraint rather than an independent initiative. This points toward a systemic pattern: true advancement in telecom infrastructure requires holistic, cloud-native fabric adoption across all layers, suggesting that focusing solely on RAN optimization risks creating siloed improvements rather than achieving end-to-end systemic coherence. The implied cost structure suggests that value extraction will naturally follow wherever operational quality and monetization opportunities are established first.
Sentinel — Human
This text appears to be a high-quality summary or transcription of an expert interview, characterized by structured reasoning derived from specific industry perspectives rather than pure generative prose.