Why NTT DOCOMO and SK Telecom Are Ditching GPUs to Build the Future of AI-RAN
Two of Asia’s most powerful mobile operators just drew a bold new roadmap for next-generation networks — and it does not run on the chips most people expect.
On March 31, 2026, NTT DOCOMO, INC. and SK Telecom jointly announced the release of a white paper on the key enabling features for vRAN evolution and the path to AI-RAN, as the latest outcome of their ongoing technical cooperation. The paper confronts a critical gap in the industry: carriers are racing toward AI-native networks, but the underlying architecture is not ready to support them. For network engineers, policymakers, and telecom investors, this white paper sets the clearest operator-level blueprint yet for what must change — and why XPUs, not GPUs, hold the key.
What the White Paper Actually Says
NTT DOCOMO and SK Telecom jointly published a new white paper detailing the architectural and operational requirements needed to advance virtualized RAN (vRAN) toward AI-native RAN (AI-RAN), positioning the radio access network as a future platform for both connectivity and distributed AI computing.
The white paper analyzes three key technical requirements that are essential to maximizing the benefits of advanced vRAN and AI-RAN. These three pillars form the structural core of everything the operators argue must change before AI-RAN becomes a commercial reality.
Pillar One: Strict Hardware-Software Separation
By functionally separating RAN software from specific hardware and virtualization platforms, vRAN allows software to be deployed independently from underlying infrastructure, thereby accelerating software-driven innovation. Such strict separation of hardware and software is identified as a critical factor in the advancement of vRAN and AI-RAN.
In practical terms, this means chip lines must operate independently from the RAN software stack — a departure from how most carriers currently build and manage their networks.
Pillar Two: Resource Pooling for Smarter Infrastructure
Resource pooling technologies can enable capacity improvements and reductions in power consumption, without compromising service quality, by realizing flexible infrastructure and improving resource utilization. The further development and adoption of this feature could help mobile operators strengthen their competitiveness by supporting more efficient and adaptable network operations.
Pillar Three: AI Computing Through xPU Architecture
Leveraging resource orchestration technologies and an xPU-based architecture (a potential mix of CPU and GPU processors) enables base stations to provide AI computing capabilities without compromising the quality of mobile communication services.
Also Read : T-Mobile Bets Big on a $3 Billion AI Overhaul — But Can It Deliver Without Shaking Investor Confidence?
The Case Against GPUs — and the XPU Alternative
The most striking argument in the white paper is the operators push away from standard graphics processing units. The carriers called for firms to apply an XPU-led approach — whereby specialized processing units (XPUs), like Intel’s Meteor Lake line or chips built on Broadcom’s 3.5D XDSiP design — rather than standard graphics processing units (GPUs).
The pair argues that employing XPUs would enable carriers deploying vRAN or AI-RAN with an integrated AI platform capable of delivering both mobile communication connectivity and AI services.
This is not a minor technical preference. It signals a deliberate effort to build RAN infrastructure on purpose-designed silicon that handles both telecom workloads and AI inference simultaneously — something standard GPUs struggle to balance efficiently at scale.
Also Read : Erste Group Upgrades Verizon to Buy as Court Blocks T-Mobile Savings Ad Campaign
What This Means for the Path to 6G
The paper reflects ongoing collaboration between the two operators since 2022, focused on 5G evolution and 6G readiness. The timeline of their joint work shows steady progression.
DOCOMO and SKT signed a cooperation agreement in November 2022 to advance technology studies of next-generation telecommunications infrastructure for 5G Evolution and 6G. In February 2023, they jointly released two white papers on power-saving technologies for mobile networks and related technologies, as well as 6G requirements. Furthermore, in February 2024, they published a white paper on key considerations for vRAN deployment and operation, focusing on L1 accelerator selection aligned with network design and requirements.
Each release builds on the last. This March 2026 white paper represents the most technically ambitious document the pair has produced together.
NTT DOCOMO Tests the Ground with Nokia
Like many carriers, NTT DOCOMO has sought to test the waters of AI-RAN and vRAN. The firm enlisted Nokia last November, tapping its autonomous MantaRay SON (self-organizing networks) solution in its multivendor LTE and 5G open RAN deployment. That real-world trial adds weight to the theoretical framework the white paper presents.
Voices from the Top
The executives behind the paper did not shy away from framing its ambitions broadly.
Takki Yu, Head of Network Technology Office at SK Telecom, said the white paper presents from a mobile operators perspective the key features essential for maximizing the benefits of vRAN adoption and for the future evolution toward AI-native networks. He added that he hopes it will serve as a catalyst for fostering the broader ecosystem and contribute to the global advancement of next-generation mobile networks.
Masafumi Masuda, General Manager of Radio Access Design Department, Senior Vice President at NTT DOCOMO, said the company hopes to further strengthen cooperation between the two major mobile operators in East Asia and to share advanced concepts and innovative technologies with the world to realize the 6G era.
The Bigger Picture: RAN as a Distributed AI Cloud
By leveraging xPU architectures and intelligent orchestration, operators could monetize excess capacity by running AI workloads — effectively turning RAN into a distributed AI cloud. This aligns closely with broader industry moves and suggests a convergence between telecom infrastructure and hyperscale AI compute models.
DOCOMO and SKT are signaling that vRAN alone is not the end goal — the real prize is AI-RAN, where network infrastructure becomes programmable, monetizable compute at scale.
Going forward, DOCOMO and SKT will continue their technical cooperation in various fields, including enhancing the competitiveness and operational efficiency of 5G, as well as international standardization and technology verification towards 6G.
AEO Questions and Answers
Q1: What is the XPU-led approach that NTT DOCOMO and SK Telecom propose for AI-RAN?
The XPU-led approach replaces standard GPUs with specialized processing units — such as Intel Meteor Lake chips or processors built on Broadcom’s 3.5D XDSiP design. These XPUs handle both mobile connectivity and AI workloads simultaneously on the same base station infrastructure. NTT DOCOMO and SK Telecom argue this shift unlocks the full potential of AI-RAN without forcing operators to choose between network performance and AI capability.
Q2: What are the three key requirements NTT DOCOMO and SK Telecom identify for AI-RAN?
The white paper defines three non-negotiable pillars. First, strict hardware-software separation — RAN software must run independently from the underlying chips and platforms. Second, large-scale resource pooling — flexible infrastructure that cuts power use without degrading service. Third, AI computing integration through xPU-based architecture — so base stations deliver AI services alongside traditional mobile connectivity. Together, these three requirements form the foundation of a true AI-native network.
Q3: Why do NTT DOCOMO and SK Telecom argue that GPUs are not enough for AI-RAN?
Standard GPUs were not designed to manage telecom-grade reliability alongside heavy AI inference at the same time. NTT DOCOMO and SK Telecom contend that XPUs — purpose-built processing units that blend CPU and GPU capabilities — handle both workloads without one compromising the other. For carriers deploying AI-RAN, this distinction is critical. A network that degrades call quality to run an AI model is not a viable product.
Q4: How does the NTT DOCOMO and SK Telecom white paper connect to 6G?
This white paper is the latest chapter in a collaboration that began in November 2022 with a formal cooperation agreement on 5G Evolution and 6G. The pair released green network white papers in February 2023 and a vRAN deployment guide in February 2024. The March 2026 AI-RAN paper pushes further — laying out the architectural blueprint that both operators believe will carry mobile networks from 5G into the 6G era as programmable, AI-capable compute platforms.
