Verizon Pushes Ericsson Toward Open vRAN for AI-Native 6G Readiness
As operators eye 6G timelines and the integration of AI agents directly into the RAN, hardware and interface flexibility have become non-negotiable. On June 29, 2026, Verizon CTO Yago Tenorio made this explicit in comments to Light Reading, underscoring that the carrier’s extensive Samsung-based vRAN footprint—nearly half its network—sets a benchmark Ericsson must meet to support future AI-driven autonomy and multi-vendor disaggregation.
The Scale of Verizon’s Existing vRAN Deployment
Verizon has deployed approximately 60,000 vRAN sites, encompassing macro cells, small cells in high-density venues like stadiums, and enterprise private networks. This architecture runs on general-purpose x86 servers (primarily HPE with Wind River software layers) paired with Samsung radios and baseband software. The operator defines “open RAN” not merely as O-RAN Alliance interface compliance but as the ability to execute RAN workloads on off-the-shelf CPUs rather than proprietary ASICs.
This deployment has grown steadily: in early 2025 Verizon reported 22,900 vRAN sites representing 40% of its footprint. The progression to 60,000 sites demonstrates production-scale validation of cloud-native RAN, providing a real-world testbed for the very capabilities—dynamic resource pooling, rapid software iteration, and third-party application insertion—that AI-RAN promises.
Why Open Interfaces and CPU-Based Compute Matter for AI
Tenorio emphasized that open standards and interoperability form “non-negotiable architectural foundations” for 6G. Critically, these interfaces enable embedding “network language models or agents” directly into the radio operating system. In an AI-native RAN, near-real-time RIC (RAN Intelligent Controller) loops—already standardized in O-RAN—can host xApps that perform closed-loop optimization of beamforming, link adaptation, and interference coordination using lightweight ML models.
Running these workloads on general-purpose x86 (with future Granite Rapids silicon cited as sufficiently powerful for early 6G) allows operators to co-locate telco and AI inference tasks without dedicated accelerators in every site. This contrasts with purpose-built baseband ASICs that, while power-efficient for PHY layer processing, limit the insertion of new AI functions post-deployment.
The same openness facilitates swapping radio units or baseband vendors without forklift upgrades, a prerequisite when AI models evolve faster than hardware refresh cycles. Verizon sees this flexibility as essential to avoid vendor lock-in while scaling AI-driven services such as intent-based slicing and predictive energy management.
Ericsson’s Current Cloud RAN Position and the Pressure Points
Ericsson has stated its cloud RAN software is commercially available on Intel Xeon processors and carries prototype support for AMD, Arm, and Nvidia platforms. However, the majority of its live RAN deployments remain tied to custom baseband hardware. Tenorio noted that Ericsson “is not ready” for vRAN parity with Samsung “as of today,” though the vendor now has “a good plan.”
This creates a clear timeline pressure. AT&T’s parallel $14 billion, five-year deal with Ericsson (announced 2023) has so far yielded cloud RAN at only 21 sites in two cities on older Sapphire Rapids silicon; the operator is now scaling with Granite Rapids to avoid deploying multiple servers per site. Both carriers are effectively signaling that multi-silicon support—including Nvidia for GPU-accelerated AI workloads—is required for AI-RAN roadmaps.
Analyst commentary from Omdia’s Gabriel Brown captures the underlying trade-off: purpose-built ASICs deliver superior performance density, while vRAN offers the flexibility needed for AI-native evolution. Ericsson’s own AI in RAN software subscription (launched earlier in 2026) brings telco-grade models for scheduler optimization and beam management into basebands and radios, but full realization of agentic, multi-vendor AI-RAN likely demands the disaggregated compute model Verizon advocates.
Architectural Implications for AI-RAN and Open RAN Convergence
The Verizon-Ericsson dynamic highlights how Open RAN’s disaggregation (RU/DU/CU split with open fronthaul and midhaul interfaces) directly enables AI-RAN. With standardized E2, A1, and O1 interfaces, operators can insert rApps and xApps that ingest real-time telemetry and push policy updates into the near-RT RIC. When the underlying compute is x86-based rather than ASIC-locked, these AI components can leverage containerized frameworks (Kubernetes, with acceleration via SR-IOV or DPDK) and even offload heavier inference to co-located GPUs when needed.
For 6G, this architecture supports the transition from today’s AI-for-RAN (optimization of existing functions) to true AI-native RAN, where the radio itself becomes a programmable AI compute node. Verizon’s emphasis on generic hardware aligns with broader industry moves—Nvidia’s Aerial platform, Nokia collaborations, and O-RAN Alliance work on AI/ML extensions—while underscoring that silicon diversity (CPU + optional GPU/accelerator) is the practical path.
Energy efficiency, always a RAN concern, also benefits: modern x86 platforms with improved power management, combined with AI-driven dynamic voltage/frequency scaling and sleep modes, can match or exceed ASIC performance per watt when workloads are bursty or AI-optimized.
Risks and Realistic Timelines
Transitioning legacy purpose-built sites to vRAN is non-trivial. Operators must validate interoperability across multi-vendor RUs, ensure fronthaul timing (eCPRI) stability under AI-induced traffic variations, and manage the increased software complexity of RIC orchestration. Memory price volatility driven by AI demand could also raise CPU costs, as noted in recent vendor warnings.
Nevertheless, Verizon’s public stance—coupled with AT&T’s measured rollout—establishes a clear operator requirement: by the time 6G products emerge around 2030, open vRAN must be mature enough to serve as the default foundation. Ericsson’s response will likely involve accelerated validation of its multi-silicon cloud RAN stack and tighter integration with O-RAN Alliance specifications for AI workloads.
Conclusion
Verizon’s latest comments crystallize a pivotal moment for the Open RAN and AI-RAN intersection. Open interfaces and CPU-based compute are no longer optional nice-to-haves; they are prerequisites for embedding intelligent agents, avoiding lock-in, and preparing infrastructure for 6G. As Ericsson responds to this pressure alongside Samsung’s proven deployments, the industry moves one step closer to a truly disaggregated, AI-native radio access network.