arXiv CSopen access

6G: From Connectivity Infrastructure to Guaranteed Digital Services

This manuscript was successfully submitted online on 27 July 2026 and is presently being given full consideration for publication in IEEE Access.

Manuscript ID Access-2026-37371

6G: From Connectivity Infrastructure to Guaranteed Digital Services David Soldani¹ (Senior Member, IEEE), Petrit Nahi¹, Awn Muhammad¹, Nikhil Dwivedi¹, Francesco Monaco¹, and Francis Jebamani¹ ¹Rakuten Mobile Inc., Tokyo 158-0094, Japan Corresponding author: D. Soldani (e-mail: [email protected]).

ABSTRACT Sixth-generation mobile networks are approaching a structural inflection point. Five generations of vendor-led architecture have left operators dependent on platforms they cannot fully modify and artificial-intelligence inference layers they cannot audit. This article argues that 6G should reverse that trajectory by reordering five priorities: control first; customer outcomes before peak rates; business guarantees before megabytes; software-driven operations with governed agentic artificial intelligence; and technology in service of those priorities. Four contributions operationalize the thesis. The Control Compact is an own-federate-consume taxonomy that allocates architectural sovereignty by strategic value. The Guarantee Economy is a six-tier outcome-priced model aligned with IMT-2030 usage scenarios and converts operator control into enforceable service-level objectives. An operator-grade Network MCP Platform shows how autonomous agents could enter the service-based architecture through a governed tool plane with auditable hooks for identity, charging, lawful intercept, enforcement, and digital-twin validation. A standardization section states Rakuten Mobile's public position on AI-agent scope, radio access, migration, non-terrestrial networks, physical layer, core, and spectrum. The framework distinguishes operational evidence from national-scale cloud-native Open RAN and core network deployments, standards-grounded extrapolation, and forward-looking architecture and commercial proposals. A three-phase roadmap separates standards milestones from operator deployment targets and identifies validation gates and stakeholder implications. INDEX TERMS 6G, Agentic AI, AI-RAN, IMT-2030, Model Context Protocol (MCP), eBPF, network sovereignty, Open RAN, O-RAN, standardization, zero-touch network management.

I. INTRODUCTION

Mobile communication has connected more than five billion unique subscribers [1] and reinvented its technical foundations roughly every decade, yet beneath this record lies a structural paradox: each generation delivered new capabilities while ceding more architectural control to a small set of equipment vendors. Operators have become experts at procuring, integrating, and operating technology they do not own, using interfaces they did not design, on platforms they cannot modify, progressively outsourcing their own strategic autonomy. As the International Telecommunication Union Radiocommunication Sector (ITU-R) IMT-2030 framework [2]–[4] sets the technical horizon for sixth-generation (6G) mobile networks, the central question is not whether 6G will be technically superior to fifth-generation (5G) networks – it will be – but whether operators will design and control the platform VOLUME XX, 2026

or merely consume it. The architectural decisions made during the 2026–2028 window will determine whether the industry’s default 6G model is operator-controlled or vendor-dependent. This article argues that 6G requires operators to reorder five fundamental priorities – not incrementally, but structurally, as illustrated in Fig. 1: Control First (owning the softwaredefined control plane); Customer First (delivering verifiable outcomes beyond peak data rates); Business First (outcomebased contracts, not connectivity pricing); Operations First (networks operated as software with agentic AI at the helm); and Technology Last (architecture and enabling technologies serve the four preceding priorities, not the antithesis). Technology Last is the most counterintuitive reordering, yet the most imperative: it does not mean that technology is unimportant; enabling technologies such as, e.g., sub-terahertz (sub-THz) spectrum, reconfigurable intelligent surfaces (RIS), and integrated sensing and communication (ISAC) remain 1

genuinely transformative. The argument is about sequencing: technology selection should follow from the four preceding priorities. Rakuten Mobile provides evidence that Control First is achievable at scale: its network runs on a fully disaggregated, software-defined architecture controlling the complete stack from the distributed unit (DU) to the core and orchestration layer. Public FY2025 full-year earnings before interest, taxes, depreciation, and amortization (EBITDA) disclosure [5] provides contextual commercial evidence, but it is not treated here as proof of the article’s broader architectural thesis, as presented in the following sections. The article makes four principal contributions, each graded by its evidentiary status: • An operator-grade agentic network platform and its proposed realization as the Network MCP Platform – the core technical contribution: six cooperating subsystems that admit autonomous artificial-intelligence (AI) agents into the 3GPP service-based architecture (SBA) while providing auditable design hooks for operator control, identity, charging, lawful intercept, sub-second enforcement, and digital-twin-validated safety, together with a tool catalog, migration path, and productionviability analysis. This proposal addresses five control problems that the NGMN AI-agent reference framework leaves open [6]: N×M routing and trust explosion, attested agent workload identity, charging and lawful intercept on agentic traffic, sub-second enforcement of nondeterministic agent behavior, and adversarial runtime defense inside the service-based architecture. • The Guarantee Economy: a six-tier, outcome-priced commercial catalog – one tier per IMT-2030 usage scenario, with Sovereign Trust as a cross-cutting assurance premium – extending ETSI Zero-touch Network and Service Management (ZSM) and TM Forum IG1252 [7] with specific service-level objectives (SLOs), breach consequences, billing mechanisms, and the three enforcement prerequisites that condition commercial deployment. • A standards-grounded strategic framework: The 6G Control Compact ownership mapping (Own | Federate | Consume), a six-family mapping of the IMT-2030 usage scenarios onto commercial tiers with quantified key performance indicators (KPIs) anchored in 3GPP TR 22.870 [8] and the ITU-R Technical Performance Requirements (TPR) [4], an integrated trust-sovereigntysustainability compliance argument, and a three-phase deployment roadmap. • Rakuten Mobile’s public position on seven principal standards development organization (SDO) architecture decisions. The evidence model is deliberately tiered to avoid overclaiming. First, production-validated evidence: national-scale cloud-native Open RAN infrastructure, TM Forum GB1059H-certified Level 4 autonomous energy optimization, large-language-model (LLM)-based network infrastructure 2

optimization, and extended Berkeley Packet Filter (eBPF) observability are publicly documented [5], [9], [10], [11]; of these, [9] and [11] are peer-reviewed, while [5] is corporate financial disclosure and [10] is an industry-foundation report. Second, standards-grounded extrapolation: the service tiers, control architecture, and deployment roadmap are derived from ITU-R, 3GPP, ETSI, O-RAN, NGMN, and TM Forum baselines cited throughout. Third, forward-looking proposal: the Network MCP Platform and the Guarantee Economy are architectural and commercial proposals built on that substrate; this article analyzes their design rationale and production viability rather than claiming field deployment results for them. The remainder of this article is organized as follows. Section II positions the contributions against related work. Sections III-VI develop the four positive “reorderings”: the Control Compact, the six outcome families, the Guarantee Economy, and the agentic AI operating model. Section VII presents the operator-grade agentic network platform and its realization as the Network MCP Platform. Section VIII introduces Rakuten Mobile’s public position on the 6G standardization landscape. Section IX describes the end-toend 6G architecture (Technology Last). Section X addresses zero-trust security, data sovereignty, and sustainability. Section XI illustrates the phased roadmap and stakeholder implications; Section XII identifies open research questions; and conclusions are drawn in Section XIII. II. RELATED WORK

The operator-sovereignty argument advanced here is positioned against six bodies of prior work. Each contribution addresses one or more of the five reorderings in isolation; the integrated framework synthesized here is original to this work. The distinguishing characteristic relative to the cited prior art is not that every element is already deployed, but that the paper explicitly separates production-validated operational substrate, standards-grounded extrapolation, and forwardlooking architectural proposal, rather than presenting them as one undifferentiated claim set. 6G vision, requirements, and AI integration. The ITU-R IMT-2030 framework [2]–[4], the AI-RAN Alliance architecture and working-group reports [12]–[14], and 3GPP TR 22.870 [8] establish the standards baseline and usage scenarios; NTT DOCOMO’s 6G vision [15] provides the 01 CONTROL FIRST Own the SW-defined control plane OSS/BSS · RAN · Core · AI · data 03 BUSINESS FIRST The Guarantee Economy closed-loop SLA · ETSI ZSM

02 CUSTOMER FIRST Verifiable outcomes — not peak Gbit/s ITU-R TPR · six outcome families

6G Operator-Controlled Network

04 OPERATIONS FIRST Networks as software NGMN L4/L5 · agentic AI

05 TECHNOLOGY LAST Architecture serves priorities 01–04 sub-THz · RIS · ISAC · AI-native PHY

FIGURE 1. Five-axis priority reordering for the 6G era. Technology Last is a sequencing principle, not a downgrade. Each wedge represents an operator investment and architectural decision set. VOLUME XX, 2026

cyber-physical fusion context. Chatzieleftheriou and Liotou [16] survey AI techniques across the 6G physical layer, RAN, and network management, synthesizing over 200 publications; their survey documents the AI/ML toolbox layer by layer but does not address operator control, ownership allocation, or outcome-priced commercial frameworks – the present article specifies which of those tools must be owned, federated, or consumed, and ties each choice to a revenue engine. Zheng et al. [17] demonstrate an AI-native physical layer with crossmodule optimization and cooperative control agents, grounding the Stage 2–3 air-interface evolution of Section IX. The AI-RAN Alliance roadmap [12]–[14] is the closest to Sections V and IX: this article extends it by integrating the AIRAN platform into a complete commercial and operational framework grounded in national-scale deployment evidence. Open RAN and RAN-Core convergence. Salmi et al. [18] address real-time adaptation and conflict resolution in AInative Open RAN: their Conflict Management Engine exemplifies the mechanism any operator-owned RIC fleet must implement to honor Guarantee-Economy SLOs across overlapping xApp scopes. The O-RAN nGRG reports on the Near-RT RIC [19], dApps [20], and scalable RAN [21] provide a research and prototype baseline for the three-tier intelligence architecture of Section IX (nGRG reports are contributed research reports, not O-RAN normative specifications). Harkous et al. [22] show that RAN-Core convergence – a distributed UPF co-located with the CU-UP – is a latency-critical requirement for the sub-10 ms edgeinference targets, not an architectural elegance; this article frames that choice within the Control Compact. Autonomous network management and agentic AI. ETSI ZSM [23]–[25] defines the closed-loop automation framework; TM Forum IG1252 [7] the Autonomous Networks Levels; the NGMN Agentic AI framework [26] the Level 3–5 taxonomy; and the broader closed-loop servicemanagement literature is collected in the IEEE Communications Magazine special issue on advanced AI for zero-touch service management [27]. Sun et al. [28] survey explainability requirements, reinforcing the case for an operator-auditable AI substrate. Provvedi et al. [29] instantiate the Intent → Observe → Decide → Act loop through LLMgenerated network configuration code. Yohannes et al. [9] report production deployment of an LLM-enhanced, architecture-aware Kubernetes resource-optimization framework at Rakuten Mobile – an average potential cost reduction of 61% across 68 production microservices in five namespaces (range 21–88% per namespace) with a human-inthe-loop approval workflow – the operating point (Level 4 autonomy under Level 5 governance) this article advocates throughout. Ferrag et al. [30] argue the same architectural direction from the research side: LLM agents as bounded, policy-governed reasoning entities in a semantic control plane layered above – and explicitly complementing – the deterministic 3GPP control plane, evaluated through a device– edge–core feasibility study showing that no single model 3

simultaneously satisfies latency, throughput, and reasoningaccuracy constraints. This provides independent support for two positions taken here: that agent autonomy must be bounded by policy above, not inside, the deterministic control plane, and that heterogeneous model tiering across the device– edge–core continuum is a deployment necessity rather than an optimization. Security for autonomous and AI-native networks. NIST SP 800-207 [31] establishes zero-trust principles and ETSI GR ZSM 017 [32] applies them to closed-loop automation. Choudhary et al. [33] frame the 5G-to-6G post-quantum transition as a multi-year sequencing problem, combining NIST-finalized ML-KEM [34] and ML-DSA [35] with classical primitives. Altintaş et al. [36] categorize adversarial threats specific to AI-native 6G – physical-world, inputmanipulation, model-level, and cross-domain – specifying the attack surface the ZSM 017 mechanisms must defend. Karahan et al. [37] demonstrate engineering-grade, multimethod explainable AI (permutation importance, SHAP, LIME, partial dependence) of the maturity that EU AI Act Article 14 oversight requires. 6G standardization. 3GPP TR 38.914 [38], the 3GPP 6G workshop contributions from Nokia [39] and Qualcomm [40], and Japan MIC spectrum planning [41] ground Section VIII. Rakuten Mobile’s operational publications [5], [9], [10], [11] provide the deployment evidence prior literature has lacked. The aggregate gap can be stated precisely, and at two levels. The sharp, checkable gap is this: to the best of our knowledge, no prior work – including [18], [26], [29], [42], [6], and the agentic AI-native architecture of [30] – specifies how autonomous software agents acquire attested workload identity, charging, lawful intercept, and sub-second enforcement inside the 3GPP service-based architecture; [30] identifies workload identity and agent-to-agent trust as open requirements of its distributed multi-agent fabric but does not bind them to 3GPP procedures, and leaves charging, lawful intercept, and enforcement latency unaddressed; Section VII proposes an architecture that addresses that gap. The broader gap is integrative: prior surveys [16], [36] catalog AI paradigms and threats; prior architectures [17], [18], [22] specify component-level innovations; earlier automation work [7], [26], [29], [42] specifies operating models; previous security work [32], [33] specifies trust mechanisms; no single treatment (i) classifies every architectural element by operator ownership, (ii) ties that classification to an outcome-priced commercial model, (iii) sequences the migration over three deployment phases aligned with 3GPP Releases 18-21, and (iv) makes trust, sovereignty, and sustainability first-class architectural constraints rather than retrospective compliance items, while grounding the argument in a live national-scale cloud-native deployment context. Three adjacent literatures are deliberately outside the scope of this positioning: the network-slicing service-levelagreement (SLA) economics and pricing corpus, the rapidly expanding body of work on LLM-based agents for network VOLUME XX, 2026

operations beyond [29], [9], [42], and [30], and the intentbased networking lineage; the Guarantee Economy and the intent translator of Section VII should be read as operator-side syntheses that those literatures will refine. III. CONTROL FIRST – THE 6G CONTROL COMPACT A. Control Illusion

The history of mobile generations is conventionally told as technical progress; what it omits is the corresponding loss of operator agency. The 2G GSM era established the pattern: a standards-defined air interface combined with proprietary vendor implementations of the core and the operations and business support systems (OSS/BSS), creating de-facto lockin inside a formally open standard. The 4G all-IP shift was transformative, but the Evolved Packet Core (EPC) was delivered almost universally as a monolithic appliance with proprietary northbound interfaces; an operator running a fourvendor network would typically operate four separate OSS instances with no common data model and no ability to implement cross-domain AI optimization without costly bespoke systems integration. 5G’s Service-Based Architecture (SBA) and network functions virtualization (NFV) offered genuine promise, but virtual network functions (VNFs) were frequently ports of appliance software, management and orchestration (MANO) became another integration battleground, and the AI/ML capabilities embedded in 5G stacks remained deliberately opaque – fragmentation NGMN Framework for Network Simplification [43] identifies as a primary driver of operational cost. Open RAN is a necessary step, but open interfaces at the RAN layer alone do not give operators control of the full stack. This is the Control Illusion: operators believe they are operating their networks, when the networks are operating themselves according to logic the operator cannot see, modify, or own. An operator that cannot access its own network data cannot train its own AI models, audit vendor AI claims, or build differentiated automation. Hyperscale cloud providers are actively building toward the same control-plane value – each positioning itself as the AI and orchestration layer inside operator infrastructure. The operator’s defensible advantage – licensed spectrum, physical presence at the radio edge, carriergrade SLA accountability, and subscriber relationships – is durable only if operators own the software layer that sits above the spectrum and below the application.

transition has been partial: the network-slice promise has been realized in proofs-of-concept, rarely at commercial scale, because orchestration and assurance remained immature. Earlier work framing mobile networks as a software operating system [44] anticipated this transition; the Control Compact operationalizes it. • Phase 3 – Intelligent Growth (6G) – casts the operator as a digital service platform owner: a network that senses application requirements, guarantees outcomes, exposes capabilities through APIs, and prices on value delivered [6]. Phase 3 requires three capabilities no previous generation delivered at scale: real-time AI-driven SLA assurance, an operator-owned data layer, and a business model mapping network outcomes to customer outcomes. Rakuten Mobile’s evidence that the Phase 3 foundation is in place is architectural: TM Forum GB1059H-certified Level 4 autonomous operations, a nationwide RIC deployment, and LLM-based infrastructure optimization in production [5], [9]. The FY2025 commercial results – revenue of JPY 374.7 billion (+32.0% year-on-year), first full-year EBITDA profitability, 10+ million subscribers, and a 48.8% higher annual Rakuten Ichiba grossmerchandise-sales (GMS) for Rakuten Mobile subscribers than non-subscribers, with net ARPU of JPY 2,467 (+JPY 59 year on year) [5] – provide commercial context for that foundation rather than proof of the architectural thesis. The outcome-priced revenue model that fully defines Phase 3, however, remains in early commercialization (Sections V, IX.B, and XII). The structural insight is that competition shifts discontinuously between phases: Phase 1 was a coverage race; Phase 2 was expected to be a slicing/API race but largely reverted to rollout speed because differentiation capabilities were not operationally ready; Phase 3 will be the AI-substrate and data-sovereignty race. Unlike the 5G transition, Phase 3 competitiveness requires the Control Compact investments to be made now – the 2026–2027 procurement, open-source, and talent decisions determine which operators arrive at Phase 3 with an owned AI substrate and which arrive dependent on vendor-managed intelligence.

B. Three Phases of Telecom Evolution

As illustrated in Fig. 2, the industry’s trajectory comprises three phases. • Phase 1 – Connectivity (1G–4G) – defines the operator as a bit-transporter; ubiquitous connectivity collapsed differentiation and forced price competition under high capital intensity. • Phase 2 – Programmability (5G/5G-Advanced) – replaced appliances with software on commodity compute, but the

PHASE 1 · Connectivity 1G → early 4G

PHASE 2 Programmability 5G / 5G-Advanced

PHASE 3 Intelligent Growth 5G-Advanced → 6G

KEY TECH voice · data pipes · bit transport

KEY TECH SDN · NFV · cloud core · APIs

KEY TECH Open RAN · agentic AI · CAMARA

REVENUE MODEL flat-rate bandwidth

REVENUE MODEL capacity and latency tiers

REVENUE MODEL outcome contracts · APIs · NaaS

ARPU SIGNAL stagnant / declining

ARPU SIGNAL flat in real terms

ARPU SIGNAL Rakuten: +JPY 59 net ARPU YoY; GMS +48.8% (FY2024)

OPERATOR VALUE voice · SMS · data caps

OPERATOR VALUE private-5G pilots · slice PoCs

OPERATOR VALUE Guarantee Economy · AI-on-RAN

OWNERSHIP vendor stacks · integrator-led

OWNERSHIP hybrid — vendor core, closed OSS

OWNERSHIP operator-owned service platform

Operator revenue potential ↑ · Vendor architectural control ↓

FIGURE 2. Three Phases of Telecom Evolution: From Pipe to Platform. Operator revenue potential rises as vendor control falls; Phase 3 reverses two decades of disaggregation drift. .

4

VOLUME XX, 2026

C. Own, Federate, Consume

The central design question for 6G operator strategy is not “how much should we build versus buy?” but “what must we own to remain sovereign?” Fig. 3 distinguishes what operators MUST OWN (ceding it creates structural dependency), what they can FEDERATE (shared ownership creates mutual value without dependency), and what they should CONSUME AS COMMODITY (ownership creates cost without strategic benefit); Table 1 maps the taxonomy across the full 6G stack. Operators must own: (1) the control plane – the software layer expressing policy, enforcing service guarantees, managing spectrum, and coordinating cross-domain behavior; (2) the AI substrate – data pipelines, fine-tuning and adaptation, retrieval corpora, inference infrastructure, and feedback loops – AI the operator cannot audit, modify, or redirect cannot serve operator-specific objectives [26], [6]; frontier foundation-model weights are the deliberate exception: they are federated, licensed under the contractual protections defined below, while everything that grounds them in the network remains owned; (3) the data layer – telemetry, analytics, customer behavior, and performance records, the raw material for AI training and SLA verification; (4) the OSS/BSS stack; and (5) spectrum policy and licensing. Operators can federate roaming and interconnect, API marketplaces (GSMA Open Gateway [45]), open-source platform stacks, and shared rural infrastructure; they should consume cloud infrastructure-as-a-service (IaaS), conformance-tested multi-vendor cloud-native network functions (CNFs), and hardware behind open interfaces at commercial off-the-shelf (COTS) economics. Rakuten Mobile’s deployment exemplifies the Compact in production: major core, RAN-software, and management functions are virtualized or containerized and managed by Kubernetes with open APIs at every layer, while RF, real-time PHY, firmware, and acceleration functions continue to rely on specialized hardware and deterministic execution environments. The commercial outcomes are quantified [5]: more than 10 million subscribers, FY2025 full-year EBITDA profitability, TM Forum GB1059H-validated Level 4 autonomous operations achieving approximately 20% RAN OWN

FEDERATE

CONSUME

operator sovereignty

open ecosystems

COTS economics

Control plane & intent policy · slicing · SLA

Cloud-native stack Sylva · Nephio · CNCF K8s

Silicon & accelerators x86 · ARM · GPU · DPU

AI substrate & MLOps models · pipelines · loops

RAN reference impls O-RAN SC · OAI · OCUDU

COTS hardware cell-site servers · switches

Data layer & identity telemetry · auth · audit

Foundation AI models domain-tuned LLMs

IaaS · PaaS services generic cloud services

OSS/BSS & customer billing · ecosystem identity

API marketplaces CAMARA · GSMA Open Gateway

Multi-vendor CNFs conformance-tested

Spectrum policy licensing — the base asset

Spectrum sharing WRC-27 · regional accords

DevOps / CI-CD GitHub · Argo · Helm

energy savings, and nationwide RAN Intelligent Controller (RIC) deployment demonstrating 15–20% power reduction (the same nationwide program reported under two measurement bases). Production maturity is further quantified by the LLM-enhanced Kubernetes optimization framework [9] (61% average potential cost reduction, human-in-the-loop) and independently corroborated eBPF adoption for anomaly detection and cloud-native security [10]. For brownfield incumbents, migration is a multi-year sequencing program: new spectrum deployments adopt Control Compact architecture from the outset while existing infrastructure migrates as vendor contracts expire. No brownfield incumbent has yet completed a full ControlCompact migration; the sequencing below is a proposed path, not a validated precedent. The minimum viable first step is threefold: require Cloud Native Telco initiative (CNTi) CNF conformance [46] and open API documentation as contractual conditions for all new NF procurement from 2026; deploy Cloud Native Computing Foundation (CNCF)-conformant machine-learning-operations (MLOps) tooling on a parallel data plane; and insert data-access rights as a non-negotiable clause in all vendor renewals. TABLE 1. THE 6G CONTROL COMPACT: OPERATOR-OWNED, FEDERATED, AND COMMODITY LAYERS Operator MUST OWN Operator should FEDERATE

Operator may CONSUME as Commodity

End-to-end operating model and system architecture

Foundation AI models Silicon (SoCs, GPUs, (LLMs, domain-adapted) FPGAs)

AI strategy: model selection, training data, MLOps governance

Shared spectrum capacity (neutral host, national roaming)

Optical transport components

Customer relationship, identity, and trust layer

Network API exposure marketplaces (CAMARA, Open Gateway [45])

Generic cloud-native services (undifferentiated compute)

Data layer: telemetry Sustainability accounting Standard pipelines, model training and carbon reporting undifferentiated network data, inference context platforms functions Security governance: zero-trust policies, PQC key management

Partner settlement and onboarding infrastructure

Off-the-shelf COTS hardware (servers, switches)

Network observability: eBPF telemetry, realtime analytics

Spectrum-sharing and dynamic allocation frameworks (WRC-27)

Standardized conformance testing tools

OSS/BSS stack; spectrum policy and licensing

Multi-operator Public cloud regions for federation for roaming non-latency-sensitive and API interoperability workloads

Agentic AI operations: closed-loop automation, digital twin

Cross-operator research programs (Hexa-X-II, SNS JU, Next G)

Generic DevOps tooling (CI/CD pipelines)

FIGURE 3. The 6G Control Compact. Operator sovereignty is allocated by layer: the control plane is owned, ecosystems are federated, hardware and IaaS are commoditized.

5

VOLUME XX, 2026

Where the structurally limited supply of combined telecom/cloud-native/AI talent [43] forces tactical reliance on vendor-managed AI platforms, four contractual protections are non-negotiable: full access to training data and model architecture documentation; standard open inference APIs; the right to migrate models without vendor approval; and real-time API access to all telemetry generated by vendor NFs. D. What 6G Must Not Repeat

The most important design inputs for 6G are the failure modes of previous generations. Six are consequential. First, AI and data locked in vendor black boxes: TR 38.914 [38] makes AI/ML fundamental to 6G RAN design; delivered as opaque vendor systems, operators will run networks whose behavior they cannot explain, audit, or modify. Second, standards written by vendors for vendors – the NGMN Simplification framework [43] is the direct response. Third, RAN integration lock-in re-emerging through systems-integration dependency. Fourth, open-source platform capture by dominant contributors: Nephio [47], Sylva [48], Cilium [49], O-RAN SC [50], and kagent [51], among other projects, prevent vendor lock-in only if operators maintain the engineering presence that makes strategic redirection detectable; CNCF neutrality requirements and the O-RAN contributor licensing framework [52] are the safeguards. Fifth, standardization drift between open-source implementations and specification baselines. Sixth, misalignment between standardization velocity and commercial readiness: 5G NR was standardized in 2018, but the automation, exposure, and AI-driven operations that would have monetized it matured years later. The service-layer specifications – CAMARA [53], ETSI ZSM [23]–[25], TM Forum IG1252 [7] – must mature through Release 20 study work and be ready for Release 21 normative 6G specifications; operators should treat a 12-18-month slip in the Rel-21 package as an explicit planning scenario, given standardization and ecosystem pressure. IV. CUSTOMER FIRST – THE SIX OUTCOME FAMILIES A. The Specification Trap

The mobile industry celebrates the metrics it can measure rather than the outcomes customers buy. The specification trap has three manifestations that 6G must counter. The first is the peak-rate fallacy: ITU-R TPR [4] peak data rates of 36 Gbit/s downlink and 18 Gbit/s uplink are achieved with a single user, maximum antenna gain, and perfect channel conditions. The commercially relevant metrics are the 5th-percentile userexperienced data rates – 300 Mbit/s downlink and 50 Mbit/s uplink in the Dense Urban test environment – and the userplane interaction latency of ≤4 ms [4]; the 6G commercial narrative must lead with these, not the laboratory maxima. The second is the coverage gap: a guaranteed service proposition fails if it cannot be delivered where customers experience the use case – hospital corridors, factory floors, basements, 6

Immersive Experience ITU-R Immersive Communication DL 300 / UL 50 Mbit/s (5%-ile) · ≤4 ms Use cases: XR · holographic · spatial media

Mission-Critical Determinism ITU-R HRLLC ≤1 ms · 1−10⁻⁵ · 32-byte PDU at edge Use cases: Industry 4.0 · robotics · V2X

Massive IoT Fabric ITU-R Massive Comm · Ambient IoT 10⁶ devices/km² · QoS-bound delivery Use cases: smart city · utilities · logistics

AI-Native Communication ITU-R AI and Communication ~26% UL AI share · <10 ms edge inference Use cases: agentic AI · edge LLMs

Sensing & Positioning ITU-R ISAC 0.75 m / 6 m accuracy · 95% detection Use cases: asset tracking · navigation

Sustainable Extended Connectivity ITU-R Ubiquitous Connectivity low-load energy · ≥10× energy-per-bit Use cases: rural · disaster · net-zero

SOVEREIGN TRUST — cross-cutting assurance premium PQC · zero-trust · data residency · trusted AI lifecycle · auditability Assurance foundation — full ITU-R TPR dimension set: rate · efficiency · capacity · bandwidth ≥400 MHz · density · mobility · latency · reliability · positioning · sensing · AI · energy · resilience

FIGURE 4. Six Outcome Families – The 6G Service Design Space. Six tiles aligned with the IMT-2030 usage scenarios – Immersive Experience, Mission-Critical Determinism, Massive IoT Fabric, AI-Native Communication, Sensing and Positioning, and Sustainable Extended Connectivity – with Sovereign Trust as the cross-cutting assurance foundation.

moving vehicles; 6G design must begin from the coverage requirement per service tier and work backward to spectrum, densification, and non-terrestrial network (NTN) integration. The third is the contract gap: even where operators can deliver a guaranteed service technically, no standardized, operatorindependent SLO specification, breach-consequence, and billing framework yet exists – the gap the Guarantee Economy (Section V) addresses. 3GPP TR 22.870 [8] acknowledges the commercial reality by framing requirements around use-case families; the translation into deployable, monetizable service tiers remains an unsolved industry problem. B. The Six Outcome Families

Aligned with the six ITU-R IMT-2030 usage scenarios and the ITU-R WP 5D draft new Report M. [IMT-2030.TECH PERF REQ] (Feb. 2026) [4], this article defines six outcome families – the 6G service design space illustrated in Fig. 4 and mapped in Table 2 – with Sovereign Trust as a cross-cutting assurance foundation spanning all six: post-quantum cryptography (PQC) readiness, zero-trust operation, data residency, trusted AI lifecycle, and auditability. The assurance foundation rests on the full TPR dimension set: peak rate, spectral efficiency, area capacity, bandwidth (≥400 MHz), connection density, mobility, latency, reliability, positioning, sensing, AI, energy efficiency, and resilience. Immersive Experience (ITU-R Immersive Communication) requires low-variance, consistently maintained throughput: TPR [4] establishes a 5th-percentile user-experienced data rate of 300 Mbit/s downlink / 50 Mbit/s uplink (Dense Urban) and ≤4 ms user-plane latency for the interactive extended-reality (XR) and spatial-media service class. Mission-Critical Determinism (ITU-R hyper-reliable and low-latency communication, HRLLC) industrializes deterministic connectivity for autonomous systems, remote machinery, and emergency control: ≤1 ms user-plane latency with 1−10⁻⁵ reliability for a 32-byte layer-2 protocol data unit (PDU) at the coverage edge [4]; safety-critical contracts can specify stricter targets commercially (Section V). VOLUME XX, 2026

Massive IoT Fabric (ITU-R Massive Communication, including Ambient Internet of Things, IoT) provides 10⁶ devices/km² connection density with QoS-bound message delivery and low-overhead, sensor/actuator-centric access for smart-city, utility, logistics, and industrial telemetry. AI-Native Communication (ITU-R AI and Communication) marks the point at which AI becomes a radio-interface capability – network for AI and AI for network – spanning data collection, distributed learning, compute, and model execution; Ericsson [54] documents that AI-native sessions generate an uplink-to-total-traffic ratio of approximately 26%, versus around 10% for conventional broadband, driven by continuous upload of context, sensor data, and reasoning artifacts. Sensing and Positioning (ITU-R ISAC) elevates the air interface to a shared sensing fabric: 95% object-detection probability at 5% false-alarm rate, and horizontal positioning accuracy of 0.75 m (indoor factory) and 6 m (urban macro) [4]. TABLE 2. MAPPING OF THE SIX OUTCOME FAMILIES TO IMT2030 USAGE SCENARIOS, ITU-R TPR KPIS, AND GUARANTEE ECONOMY TIERS Outcome Scenario Targets / status Tier Ph. source Immersive

IMT-2030 immersive comms

5%-ile DL 300 / UL 50 Consumer Mbit/s (Dense Urban); user-plane latency ≤4 ms; peak DL 36 / UL 18 Gbit/s

1→2

Critical

HRLLC / critical comms

User-plane latency ≤1 Enterprise ms; reliability 1−10⁻⁵ (32-byte L2 PDU, coverage edge); jitter <1 μs (author target)

2

Massive IoT Massive Connection density 10⁶ IoT comms / devices/km²; QoS-bound ambient IoT message delivery; lowoverhead control

1→2

AI-native

AI and comms / distributed intelligence

Edge inference latency <10 ms (author target); AI-native UL share ~26% [54]; data/training/inference support

2

Sensing

ISAC / positioning

Positioning ≤0.75 m Sensing indoor / ≤6 m urban macro; detection 95% at 5% false alarm; update <1 s (author/vertical target)

Sustainable

Ubiquitous Energy-efficient lowSustainable 1→2 connectivity load operation (unloaded, ≤30% load) [4]; energy-per-bit improvement an operator target (not a TPR figure); extended reach (HIBS, relays)

7

AI service

2

Sustainable Extended Connectivity (ITU-R Ubiquitous Connectivity plus sustainability) combines energy-efficient low-load operation – evaluated for unloaded and ≤30% load conditions against a full-load reference [4] – with resilience and extended reach through high-altitude IMT base stations (HIBS), relays, repeaters, and shared infrastructure for rural coverage, disaster recovery, and net-zero networks. Each family is anchored in quantified targets with explicit evidence status: draft ITU-R TPR evaluation minima where [4] applies, 3GPP TR 22.870 use-case requirements where [8] applies, and author commercial SLOs where the target deliberately exceeds the standards baseline. This enables the Guarantee Economy tiers of Section V to translate outcome requirements into enforceable SLOs without treating every figure as a final deployment guarantee, and positions operators to demonstrate standards alignment and, where requirements are finalized, compliance through live network operations with independently auditable data; the IOWN GF MetricEES review [55] surveys the fragmented energy-efficiency metrics (including ATIS TEER, ITU-T NCIe, and ETSI ES 203 228) and proposes their harmonization toward auditable, comparable energy performance across operators and vendors. C. Customer Segments and the Outcome Matrix

The families form a configurable matrix across segments, as illustrated in Fig. 5. Consumers in dense urban markets prioritize Immersive Experience, AI-Native Communication, and trust assurances; enterprises in manufacturing, logistics, energy, and healthcare weight Mission-Critical Determinism, AI-Native Communication, Massive IoT Fabric, and Sensing most heavily; public-sector organizations place the Sovereign Trust foundation, resilience, and Sustainable Extended Connectivity at the top, reflecting accountability to citizens and geopolitical risk exposure. Two scoping dimensions condition the matrix: • The first is market economics: the Guarantee Economy applies most immediately in high-ARPU markets with enterprise density; for operators serving subscribers at two-to-four-dollar monthly ARPU across South Asia, SubSaharan Africa, and Southeast Asia, the near-term priority is connectivity-first, with three Control Compact elements applying directly – spectrum-policy ownership, national data-layer control, and open-source adoption to reduce capital cost – which the Bharat 6G Vision [56] frames as a national strategic priority. • The second is jurisdiction. In the EU, the EU AI Act [57], GDPR [58], EHDS [59], NIS2 [60], and the Cyber Resilience Act [61] bind service design – EU AI Act Article 14 human oversight must be satisfied at service activation, not at audit. The UK operates a separate regime – UK GDPR, the Data (Use and Access) Act, and a sectorled ‘pro-innovation’ AI approach rather than the EU AI Act. Japan and the Republic of Korea are governed by APPI/PIPA and telecommunications business law; the United States by the Secure and Trusted Communications VOLUME XX, 2026

Networks Act [62] and CHIPS Act [63] supply-chain provisions, with state privacy laws layered on top. India operates under the DPDP Act and the trusted-supplier framework [56], Australia under the SOCI Act and Privacy Act amendments, and Singapore under IMDA codes. These differences shape which outcome families can be offered, at what tier, under what audit obligation: a SovereignTrust-assured tier for a Japanese government customer cannot be operated identically to one for an Australian criticalinfrastructure customer. The commercially scalable approach is a regionally parameterized template – one product with jurisdiction-specific configuration enforced by the operatorowned policy engine – not dozens of distinct catalogs. V. BUSINESS FIRST – THE GUARANTEE ECONOMY A. The Connectivity Trap

Operators were forecast to invest approximately USD 1.1 trillion in mobile network capital expenditure over 2020– 2025, roughly 80% of it 5G-related [1] (a 2021 GSMA projection), yet industry revenue growth remained in low single digits: in leading European and North American markets, ARPU has broadly stagnated since 4G launch even as traffic grew by orders of magnitude. The cause is structural: connectivity was commoditized while the economic value of digital services migrated to over-the-top platforms. The 5G enterprise case – private networks, slicing, edge – was constrained by the absence of programmable interfaces, verifiable SLAs, and cloud-competitive provisioning velocity; enterprises that deployed private 5G frequently bypassed operators entirely. The lesson for 6G is unambiguous: connectivity alone is not a sustainable business. Rakuten Mobile’s trajectory illustrates the path beyond it: its subscribers generate approximately 48.8% higher annual average gross merchandise sales (GMS) on the Rakuten Ichiba marketplace than non-subscribers [5] – evidence of the ecosystem flywheel that converts connectivity into crossservice value through Rakuten Points, e-commerce referral, and affiliated financial services across the group’s 100+ million registered users; FY2025 net ARPU reached JPY 2,467, up JPY 59 year on year [5]. This demonstrates in commercial deployment the Business First principle: revenue Outcome family

Consumer urban dense

Enterprise B2B verticals

Immersive

Public sector gov · civic —

MC Determinism

Illustrative geography (Example – not exhaustive)

EU / Japan EU / Japan / US

Massive IoT

Global

AI-Native Comm

Global

Sensing & Pos.

Japan / US

Sustainable Ext. Primary

Secondary

EU No priority in this example

Illustrative mapping only; geography is not a market-addressability or regulatory conclusion. Sovereign Trust applies across segments.

FIGURE 5. Customer Segment × Outcome Matrix. Three segments × six outcome families with Primary/Secondary/– designations; regional dimensions add a third axis reflecting divergent national 6G priorities.

8

anchored in service value, not data volume. The caveat: Rakuten’s uplift depends on a loyalty flywheel built over decades; operators without a comparable ecosystem realize Business First through API monetization, enterprise Guarantee Economy contracts, or partnership-based service aggregation. B. Three Revenue Engines

Three mutually reinforcing revenue engines are available to 6G operators; cumulative ARPU uplift accrues only when all three are operationally live. The Guarantee Economy is the most immediately actionable: AI-RAN Alliance WG3 research [14] indicates nearly half of global 5G users experience connectivity challenges during peak hours and approximately half indicate willingness to pay for guaranteed reliability. It requires three architectural components – granular SLO definition, real-time closed-loop enforcement, and transparent auditable verification; without all three, a guarantee is a marketing claim. (This willingness-to-pay evidence derives from alliance member research rather than independent demand studies; survey-stated willingness does not imply willingness to sign enforceable SLAs – the private-5G bypass precedent is the caution.) Programmable Network APIs expose network capabilities as developer-consumable services: CAMARA [53] and GSMA Open Gateway [45] established the 5G vocabulary (Quality on Demand (QoD), Location Verification, Location Retrieval, Geofencing Subscriptions, Number Verification, SIM Swap, and Device Reachability APIs); 6G extends to AIinference offload, Sensing-as-a-Service, and network digitaltwin APIs, governed by ETSI ZSM [23]–[25]. A representative CAMARA/Open Gateway candidate set is cataloged in Table 8 (Appendix A). Network-as-a-Service (NaaS) delivers end-to-end network functionality as an ondemand, intent-driven service – slices provisioned in minutes rather than months – an operator design target for Nephio [47]based orchestration. First-generation NaaS is a velocity improvement, not yet an outcome-contracted product; the transition requires adding real-time SLA monitoring, automated breach detection, and consequence execution, with ETSI ZSM 019 [23] as the vendor-agnostic abstraction. The cloud precedent calibrates the prize: Amazon Web Services grew to over USD 100 billion in annual revenue in under two decades [64], with provisioning velocity central to its adoption; mobile infrastructure adds licensed spectrum, physical edge proximity, and carrier-grade accountability that public cloud cannot replicate. C. The Guarantee Economy in Practice

Table 3 deliberately defines six commercial tiers – one per IMT-2030 outcome family – with guaranteed SLOs, breach consequences, and billing models; this is the canonical taxonomy used in this article, with Massive IoT Fabric treated as a distinct tier rather than folded into sustainable VOLUME XX, 2026

connectivity. SLO values draw from draft ITU-R TPR [4], 3GPP TR 22.870 [8], and explicitly marked author commercial targets. Sovereign Trust is not a seventh tier but a cross-cutting assurance premium attachable to any tier: PQC readiness, zero-trust operation, data residency, trusted AI lifecycle, and auditability, contracted as an uplift on the underlying SLO package. Commercial SLOs may exceed draft TPR minima where the customer pays for stricter targets: Enterprise Determinism contracts specify reliability up to 99.9999% and jitter below 1 μs against the TPR baseline of 1−10⁻⁵ [4]. SLA enforcement spans five dimensions: latency (mean and variance), reliability, throughput (peak, sustained, and fifthpercentile), availability, and security posture. At the service layer, TMF641 [66] provides service ordering and TMF657 [65] represents and exposes service-quality information, including Service Level Specifications (SLSs) and associated objectives and thresholds. Assurance, breach detection, and consequence execution are separate operator functions subject to applicable charging, lawful-intercept, and legal controls. Operational use requires a standardized measurement method accepted by both parties, jurisdiction-specific legal enforceability of breach consequences, and unambiguous attribution across operator, device, and application domains. TM Forum IG1252 [7] is an autonomous-networks evaluation methodology, not a master SLA template; bilateral contracts must define the service specification, measurement authority, breach attribution, remedies, and liability. A Guarantee Economy claim is commercially testable only when a named buyer accepts a disclosed premium for a specified SLO and remedy. Each pilot should pre-register its measurement authority, breach-attribution procedure, liability cap, renewal decision, and exit criterion; until such results are published, tier-level willingness-to-pay claims remain hypotheses. The first is partially met; the second and third remain open (Section XII). Operators should use bilaterally negotiated master service-agreement templates that define the service specification, measurement authority, breach attribution, remedies, and liability. TM Forum IG1252 [7] may support maturity evaluation but is not a contracting template. Price anchoring is external: the Smart Networks and Services Joint Undertaking (SNS JU) eHealth study [67] projects 40-60% CO2 reduction per clinical episode (a study-portfolio projection, not audited production data), and the AI-RAN WG3 [14] edge-versus-central inference differential anchors AI-tier pricing; without external anchors the Guarantee Economy reduces to cost-plus pricing, commercially insufficient to recover 6G capital expenditure. Two boundaries apply. SLA enforcement covers domestic services under single-operator control; international roaming makes end-to-end guarantees unenforceable under current inter-operator architecture, positioning international coverage as a Phase 3 objective built on GSMA Open Gateway QoD extensions. 9

TABLE 3. THE GUARANTEE ECONOMY: SIX-TIER SERVICE CATALOG WITH SLOS, BREACH CONSEQUENCES, AND BILLING MODELS (SOVEREIGN TRUST AS CROSS-CUTTING PREMIUM) Tier

Target Segment

Guaranteed SLOs

Breach Consequence

Billing Model

Premium Consumer Immersive

Consumer: gaming, XR, live events

DL ≥500 Mbit/s; E2E latency ≤20 ms; availability 99.99%

Bill credit; Monthly automatic QoS premium re-routing subscription

Enterprise Determinism

Manufacturing, logistics, robotics, remote surgery (subject to separate clinical/safety/re gulatory validation)

Latency ≤1 ms SLA penalty; Per-SLO (device-to-edge); incident report; outcome reliability to 99.9999% automated contract; (author commercial remediation framework + target ≈1−10⁻⁶, vs TPR usage baseline 1−10⁻⁵); jitter <1 μs (author target; packet-delay-variation, vertical-sourced)

Massive IoT Fabric

Smart city, utilities, logistics, industrial telemetry

Density to 10⁶ devices/km²; QoSbound message delivery; battery-life and deep-coverage SLOs

AI Inference Edge

Sensing-as-aService

Messagedelivery credits; fallback access (incl. NTN IoT)

Per-device / per-message; volume commitments

Enterprise AI, Inference latency <10 real-time ms (device-to-edge); CV/NLP at edge model availability 99.99%; throughput per agreed tokens/s

Fallback to cloud; bill credit

Perinference; committed capacity

Smart city, autonomous logistics, monitoring

Positioning ≤0.75 m indoor / ≤6 m urban; update <1 s; coverage SLA

API SLA Per-API-call credit; fallback or per-area sensing source subscription

Slice provisioned <5 min (operator product target); ≥99.9% slice integrity; guaranteed BW/latency isolation; per-slice energy and carbon-intensity reporting (research objective; not an operational standard)

Provisioning credit; automated reinstantiation; carbon-report credit

Sustainability- Hyperscaler, certified NaaS enterprise private networks, netzero-committed enterprises

Usage + committed capacity; outcome contract

The catalog assumes a facilities-based operator: mobile virtual network operator (MVNO), neutral-host, and spectrum-sharing arrangements introduce multi-party SLA pass-through chains for which standardized mechanisms do not yet exist. VI. OPERATIONS FIRST – AGENTIC AI AND INTELLIGENT GROWTH A. The Operations Gap

Despite a decade of NFV, software-defined networking (SDN), and cloud-native investment, most live 5G networks are still operated through command-line configuration, ticketbased change workflows, and reactive incident response. On the NGMN five-level agentic-AI adoption scale [26], Level 3 – conditional autonomy with human escalation – is the industry ceiling; Levels 4 and 5 remain aspirational for most operators. Two level frameworks appear in this article and are kept distinct: the NGMN agentic-AI levels [26] state the target operating model, while the TM Forum IG1252 Autonomous Networks levels [7] – the scale against which GB1059H certification is issued – provide the auditable evidence; the two scales are aligned but not identical.

VOLUME XX, 2026

The gap is a structural revenue constraint: the guaranteed tiers and NaaS of Section V are operationally undeliverable without autonomous management. A 6G network at Level 4 on the order of 10^5-10^6 cells generating O(10^6) configuration decisions per day (an engineering assumption for a dense national deployment) - cannot be operated manually under any plausible staffing assumption. The NGMN scale describes a maturity trajectory, not a uniform economic conversion factor; no general operating-cost reduction should be inferred for a move between levels [26]. TM Forum GB1059H validation of Rakuten Mobile's Level 4 autonomous RAN management is retained only as a scoped result of approximately 20% RAN energy conservation in live commercial operation [5], not as a generic autonomy return. Automation itself is not new: self-organizing networks (SON), software-defined networking (SDN), and rule-based closed loops already absorb this decision volume. What they cannot express is the residual for which no rule can be prewritten – decomposing open-ended intent and diagnosing novel faults that span RAN, core, and transport. That residual is what caps operators at Level 3, and where large-languagemodel (LLM) reasoning is warranted. Closing it requires three capability investments layered on top of existing automation – cross-domain telemetry ingestion, an LLM agent with retrieval-augmented generation (RAG) and verified toolcalling authority scoped to the management plane, and a risktiered pre-execution validation layer. High-impact, crossdomain changes justify digital-twin validation, while routine or time-sensitive decisions can be screened by a latent worldmodel approach inspired by JEPA, maintaining an abstract representation of network state and predicting likely outcomes without replaying every low-level interaction [68], [69]. Without some combination of these capabilities, the operator remains capped below Level 4. B. Networks as Software

The architectural foundation is a cloud-native operating model in which the virtualized and containerized portions of the network stack – from DU/CU and core functions to OSS – execute as software on shared, horizontally scalable compute orchestrated by Kubernetes, while RF, hard real-time PHY, firmware, and acceleration paths remain specialized deterministic functions, as illustrated in Fig. 6. Kubernetes is the network operating system; the networking and security layer is eBPF-based – Cilium/Tetragon [49] or Rakuten Mobile’s Sauron platform [10], [11] – providing kernel-level policy, observability, and security without proprietary hardware; Nephio [47] provides intent-driven lifecycle management through Kubernetes-native reconciliation. Production evidence quantifies the model’s efficiency: the LLM-enhanced, architecture-aware Kubernetes optimization framework [9] – integrating High-Level Design documents, 30-day telemetry, and deterministic right-sizing rules (request at 1.2× peak 24-hour utilization, limit at 1.3× request) – demonstrated a 61% average potential cost reduction across 10

68 production microservices in five namespaces (range 21– 88% per namespace). The saving derives from the deterministic right-sizing rules, with the LLM performing design-document and telemetry ingestion and recommendation rather than autonomous control – LLMassisted optimization under a human-in-the-loop approval workflow that instantiates Level 4 execution under Level 5 governance. The wide savings range reflects architectureawareness: latency-sensitive services receive conservative adjustments; flexible workloads receive aggressive rightsizing. C. Agentic AI: The Operating Model

LLM-based agents shift automation from rule-based scripts to intent-driven operation. As illustrated in Fig. 7, the autonomous agent proceeds through four stages on four timescales: • Intent – an operator goal in natural language (minutes). • Observe – telemetry, counters, probes, eBPF events, and customer KPIs (100 ms-1 s). • Decide – LLM reasoning, multi-agent planning, and validation (seconds). • Act – execution via APIs for operations, steering, energy management, and remediation (10 ms to minutes). The validation layer is not required to exhaustively simulate every action. For high-impact management-plane and intent-level changes, a synchronized digital twin remains the strongest safety gate. For repetitive or time-bounded decisions, a latent world-model approach inspired by JEPA can evaluate actions in a compressed state space, suppress redundant state, and escalate only ambiguous or high-risk cases to full twin analysis [68], [69]. Real-time control loops below the twin’s validation latency, such as Near-RT RIC and dApp timescales, execute within policy envelopes that this tiered validator has pre-certified, rather than being gated on a per-action basis. The operational target is NGMN Level 4 within a Level 5 governance framework [26]. Subject to pending European AI Office implementation guidance, this combination is

FIGURE 6. Cloud-Native Telecom Stack – Networks as Software. Six layers (Layer 0 Agentic AI at top, Layer 5 Infrastructure); Sylva and CNTi brackets indicate open-source governance scope. VOLUME XX, 2026

structured to support alignment with EU AI Act Article 14 [57] at the policy-and-exception level rather than per-decision; whether that ultimately satisfies “effective oversight” of O(10⁶) daily decisions remains an open regulatory question and should be treated as a design target requiring validation, not as settled compliance. The nGRG generative-AI report [42] documents agents autonomously creating, configuring, optimizing, and repairing network slices – the bottleneck currently blocking NaaS delivery. RAG grounds agents in live network state; multi-agent coordination is defined in ETSI ZSM 020 [24]; the Model Context Protocol (MCP) router provides the agent-to-tool interface: an open protocol binding (MCP is an open project under the Linux Foundation’s Agentic AI Foundation, not a 3GPP/ETSI telecom standard) combined with operator-defined authorization, policy, rate limiting, audit, and lifecycle controls — auditability is a property of those operator controls, not of the protocol. The production stack comprises four layers, requiring no proprietary components beyond CNCF [70], Linux Foundation, and O-RAN SC [50] open source: • An agent orchestration layer (kagent [51]) coordinating specialized agents named consistently with Section VII – the Optimizer (RAN and resource optimization), the SLA Enforcer (Guarantee-Economy SLO assurance), the Fault Remediator (detection, diagnosis, remediation), the Energy Saver, plus infrastructure-lifecycle (Nephio) and security (ZSM 017 closed-loop) agents – each within an authorization scope encoded in its trust certificate. • A tool-calling layer exposed through the MCP router whose audit trails can support EU AI Act Article 12 record-keeping evidence, where the use case falls within scope, and ZSM 017-aligned access-control evidence simultaneously. • A model-serving layer (vLLM or equivalent) on operatorowned hardware. • A governance layer – MLflow versioning, Digital Twin validation, and a policy engine encoding regulatory obligations.

• Federated learning where training data cannot leave jurisdictions, coordinated at the RAN layer via the NearRT RIC [19]. • Continuous retraining pipelines with drift-triggered retraining and shadow deployment. • Explainability tooling – grounded in auditable decision provenance: the natural-language intent, retrieved context, tool calls, and pre-execution validation verdict recorded per decision, producing justification artifacts that survive regulator audit, complemented by intent-to-code translation validated with human feedback [29]; where individual component models are feature-based, established XAI methods (permutation importance, SHAP, partial dependence) of the maturity Karahan et al. [37] demonstrate supplement this provenance, subject to their known sensitivity to configuration. Observability is eBPF-native: Sauron [10], [11] provides kernel-level telemetry without instrumentation overhead, extended by LLM-based anomaly detection identifying degradation before customer-visible failure, as presented at Cloud Native Telco Day EU 2026 [72]. VII. THE OPERATOR-GRADE AGENTIC NETWORK PLATFORM AND ITS MCP REALIZATION

The four preceding re-orderings converge on a single engineering question: how can autonomous, LLM-driven agents be admitted into a live mobile network without surrendering the operator control that Sections III-VI establish as non-negotiable? An agent that can reconfigure a policy function, steer traffic, or quarantine a cell can also, if unconstrained, breach an SLA, exfiltrate subscriber data, or propagate an adversarial action across domains faster than any human can intervene. This section presents our core technical contribution in two parts: the platform architecture (six cooperating subsystems) and its proposed realization as the Network MCP Platform, including the tool catalog, migration roadmap, and production-viability constraints. The platform Closed loop

D. AI/ML Infrastructure and Observability

Production agents require full-lifecycle MLOps: unified crossdomain telemetry; MLflow experiment tracking, versioning, and deployment [71]; and Network Digital Twins for simulate-before-action safety. Governance aligned with the EU AI Act [57] requires every model to be version-controlled, auditable, and documented for training data, evaluation, and performance boundaries; alignment risk – technically correct but commercially or regulatorily unacceptable optimization under distribution shift – is addressed at two levels: multiobjective reward design encoding regulatory SLA floors and fairness constraints is a first line of defense; because reward specification is itself susceptible to reward-hacking, the binding control is the runtime pre-execution validation and enforcement of Section VII.E, not the reward function alone. Three capabilities distinguish scale-grade MLOps: 11

INTENT

OBSERVE

DECIDE

ACT

Goal in natural language

eBPF · probes · customer KPIs

LLM reasoning & planning

Execute via API slice · energy

minutes

100 ms – 1 s

seconds

10 ms – minutes

AGENTIC AI ENGINE kagent · MLflow · MCP router · tiered LLMs across device–edge–core

TIERED VALIDATION — gates Decide → Act Network Digital Twin simulate before act high-impact changes PQC-signed records

Latent world-model screening (JEPA) routine decisions escalates hard cases

Pre-validated policy envelopes real-time loops re-checked on change

HUMAN OVERSIGHT policy · governance · audit · EU AI Act operators set objectives and thresholds — not per-action approvals

FIGURE 7. Agentic AI – Closed-Loop Operations Model. Intent → Observe → Decide → Act loop with human oversight (policy, governance, audit, EU AI Act) and a tiered validation layer gating Decide → Act: Network Digital Twin for high-impact changes, latent world-model screening for routine decisions, and pre-validated policy envelopes for real-time loops.

VOLUME XX, 2026

extends, rather than replaces, the 3GPP Service-Based Architecture (SBA), consistent with the networksimplification trajectory of NGMN [43], [6]. A. Five Unresolved Control Problems

The baseline emerging in the industry – overlaying agent-toagent (A2A) and MCP interfaces directly onto the 5G SBA, as sketched in the NGMN AI-agent reference framework [6] – is directionally correct but leaves five control problems unsolved. Table 4 maps each problem to the resolving subsystem. First, N×M routing and trust explosion: if every agent is a direct MCP client of every network function, the operator inherits a quadratic mesh of trust relationships, and the MCP specification [73] defines neither normative rate limiting nor multi-server routing. Second, absent cross-domain workload identity: 5G authentication and key agreement (AKA) [74] authenticate the subscriber, not the agent runtime; an agent on user equipment (UE), in a tenant, or across multi-vendor infrastructure has no attested identity verifiable across UE, RAN, and core. Third, no charging, observability, or lawful intercept on agentic traffic: an MCP invocation is observed by neither the 3GPP Charging Function [75] nor the lawful-intercept architecture [76], breaking revenue assurance and regulatory compliance simultaneously. TABLE 4. FIVE UNRESOLVED CONTROL PROBLEMS AND THE SUBSYSTEMS THAT RESOLVE THEM Unresolved control problem

Resolving subsystem(s) Operator-grade effect

N×M routing and trust explosion

(A) MCP Router with SBI/MCP arbitration

Single arbitrated, ratelimited entry point; O(N) integration

Absent cross-domain workload identity

(B) Cross-domain SPIFFE identity

Attested agent identity across UE/RAN/Core; attach/detach revocation

No charging / observability / lawful intercept

(A) Router + (C) eBPF plane

Kernel-sourced metering events and LI-design hooks on routermediated invocations; eBPF-aggregate telemetry for real-time bypass tiers; compliant CHF/LI mediation remains open

No closed-loop enforcement on nondeterministic behavior

(E) validation + (C) eBPF + (D) lifecycle manager

Pre-execution validation blocks unsafe decisions (E); sub-second throttle / quarantine on execution anomalies (C/D)

No adversarial runtime defense

(C) eBPF (sub-ms) + (E) Sub-millisecond local digital twin event capture or detection; sub-second containment target, subject to classification, policy-decision, and distributed-enforcement latency; simulate-beforeact safety gate

12

Fourth, no closed-loop enforcement on non-deterministic behavior: LLM agents are stochastic, and existing policy functions react at application-layer timescales with no path from kernel-level anomaly to agent-level rate limiting in nearreal time. Fifth, no adversarial runtime defense: input manipulation, model poisoning, RIC supply-chain compromise, and cross-layer propagation move faster than application-layer telemetry can detect [36]. B. Six Cooperating Subsystems

The platform, shown in Fig. 8, resolves the five problems through six subsystems designed to be complementary: each addresses a control problem the others do not, and the operator guarantees emerge from their combination. (A) MCP Router with SBI/MCP arbitration. A proposed logical mediation function — not a new standardized 3GPP network function — receives every agent invocation and resolves it against two registries – the network repository function (NRF) view of standardized network function (NF) service operations and a tool/schema registry of MCP-exposed capabilities (OpenAPI, YANG). A deterministic arbitration rule selects the route: a Service-Based Interface (SBI) request over HTTP/2 where a standardized API exists; MCP to a governed MCP endpoint where none exists or multi-vendor mediation is required; the route decision also weighs actor class and trust domain, operation risk and reversibility, management- versus control-plane target, interface ownership and transaction semantics, authorization and data sensitivity, and latency/availability class. The router enforces rate limits and per-agent-class quotas, acts on the enforcement signals of subsystem (C), and critically, on every router-mediated invocation across SBI and MCP routes, emits the audit and mediation hooks from which jurisdiction-specific charging and lawful-intercept support can be implemented where legally applicable — kernel-observed cycles and bytes from the eBPF plane (C) provide the usage measurement, but a compliant 3GPP charging record still requires chargeableevent definition, rating group, duplicate control, correlation, and Charging Function (CHF) mediation, and lawful intercept still requires the standardized information model, target correlation, and delivery functions, which remain open standardization and legal questions; router-bypassing realtime tiers are metered in aggregate by the eBPF plane; for RAN-core traffic requiring near-real-time delivery it supports MCP over Quick UDP Internet Connections (QUIC). (B) Cross-domain workload identity. An operator-issued Secure Production Identity Framework for Everyone (SPIFFE) trust domain spans UE, RAN, and Core through a hierarchical root of trust, reducing O(N²) bilateral federation to O(N) [77]. Agent runtimes – including UE-resident agents – receive short-lived identity documents only after hardwareanchored remote attestation (trusted platform module or trusted execution environment, TPM/TEE) and present them for mutual TLS at the router, NF MCP servers, and the network exposure function (NEF); authorization and VOLUME XX, 2026

protected-resource discovery follow RFC 9728 [78], while workload identity and attestation rely on the SPIFFE/SPIRE trust domain [77] and TPM/TEE attestation. Revocation is bound to network attach/detach, so a compromised agent loses credentials at the network edge, not at application timeout. (C) eBPF telemetry and enforcement plane. eBPF probes on NF containers, MCP servers, and optionally UE runtimes yield per-flow, per-syscall, per-resource telemetry: chargingrelevant metering inputs from kernel-observed cycles, accelerator milliseconds, and bytes – replacing user-space service-mesh sidecars for east-west agentic traffic; the 30-50% overhead reduction is an engineering estimate based on [11] and requires operator-specific validation. A runtime enforcement layer encodes anomalous syscall signatures as tracing policies; on a match, the router throttles, quarantines, or re-routes within a sub-second loop. The eBPF plane detects infrastructure-level anomalies; a semantically wrong but wellformed configuration write is instead caught upstream, by the validation tier of subsystem (E) before execution and by the behavioral baselines of subsystem (D) after it – each layer’s detection scope is deliberately delimited. (D) Closed-loop AI lifecycle manager. Consumes eBPF telemetry, maintains per-agent behavioral baselines, and performs drift detection, lifecycle management, and conflict resolution, emitting real-time policy updates to the router and the policy control function (PCF) – operating at NGMN Level 4 under Level 5 human governance [26], and supporting highrisk-AI oversight evidence [57] at the policy level where the deployed use case falls within the EU AI Act scope. A priorityweight scheduler arbitrates competing policies before commitment. (E) Network Digital Twin – simulate before act where warranted. Every proposed management-plane action is checked against a tiered validation layer: a continuously synchronized twin for high-impact or cross-domain changes, and a latent world-model approach inspired by JEPA for routine decisions where exhaustive simulation would be too slow or costly [68], [69]. Only passing actions are released, making the loop Decide-Validate-Act. Sub-10 ms control paths are not twin-gated per action: they operate inside prevalidated policy envelopes, and the twin re-validates the envelope whenever policy changes. Each validation is signed with post-quantum cryptography [34], [35], [79], providing tamper-evident records and pre-empting configuration conflicts and adversarial policy injection. (F) LLM intent translator – transactional intent binding. Natural-language intent (e.g., “improve uplink throughput for enterprise slice 7 while reducing energy by 15 per cent”) is translated by a domain-fine-tuned LLM, grounded by retrieval over NF schemas and live telemetry, into an executable sequence of MCP tool calls, PCF primitives, and Non-RT RIC A1 policies – committed across all targets with saga-like orchestration — precondition checks, bounded leases, idempotency, and compensating rollback for targets that lack native prepare/commit/abort semantics — designed to prevent 13

a partially applied intent from persisting; any residual inconsistency window is surfaced to the lifecycle manager as an exception rather than silently absorbed; coordinator failure is handled conservatively through bounded lock leases with timeout-abort semantics, so a crashed translator cannot strand locks on live network functions – ambiguous intents abort rather than complete. Each translation logs an explainability record for audit. Subsystem (F) additionally addresses a failure mode the five control problems do not capture: partial application of a multi-domain intent, which without transactional binding would leave the network in a state no single-domain rollback can repair. C. The Network MCP Platform: From RAN-Only to Whole-Network Control

Today, AI models – forecasting, pattern learning, anomaly detection – drive the closed loop but remain focused largely on the RAN, mediated through CMaaS (Configuration Management as a Service) and element management system (EMS) workflows designed for human-operated configuration management, not the high-frequency, low-latency tool calls agents require. Core NFs – the access and mobility management function (AMF), session management function (SMF), Policy Control Function (PCF), and user plane function (UPF) – remain largely unreachable by agents; probe, deep packet inspection (DPI), and eBPF data are not first-class AI inputs; every new application requires custom integration; and agents cannot discover, compose, or validate network capabilities at runtime. The Network MCP Platform is the critical evolutionary step: it enables the transition from rigid, hard-coded workflows to dynamic, intent-driven operations spanning RAN, core, and transport in one governed loop, with existing AI/ML models migrating as MCP-native agents. The architectural contrast is illustrated in Fig. 9. Rakuten Mobile’s position is that this is a production-engineering program, not a research prototype, designed around three principles:

FIGURE 8. Operator-Grade Agentic Network Platform – Six Cooperating Subsystems. (A) MCP Router with SBI/MCP arbitration, (B) cross-domain SPIFFE workload identity, (C) eBPF telemetry and enforcement plane, (D) closed-loop AI lifecycle manager, (E) Network Digital Twin (simulatebefore-act), and (F) LLM intent translator, positioned between the agentic-AI layer and the 5G/6G SBA across UE, RAN, and Core.

VOLUME XX, 2026

TODAY — RAN ONLY AI models forecast · anomaly

CMaaS / EMS bespoke adapters

RAN only EMS · vCU · K8s

Core NFs unreachable by agents · brittle, no tool discovery probes · DPI · eBPF not first-class AI inputs critical evolutionary path WITH NETWORK MCP PLATFORM — RAN + CORE + TRANSPORT AI AGENTS — Optimizer · SLA Enforcer · Fault Remediator

Read

Write

Network performance data: PM counters, active/passive probes, subscriber experience, DPI flow inspection

DPI traffic shaping: rate- Probe streams: RTT, limit profiles, QoS reloss, jitter per marking, flow link/segment; DPI event blocking/throttling flows and QoS-violation triggers

Stream / event

FMaaS, service desk, inventory; RAN Commander as lightweight digital twin

Kubernetes resource operations: scale CNF replicas, drain nodes, ConfigMaps, rolling restarts

eBPF kernel events (Sauron): AF_XDP RAN L2 telemetry, flow events, runtime security syscall traces

NETWORK MCP PLATFORM discovery · schema registry · auth & policy · rate limiting · audit RAN tools NETCONF/YANG · EMS

Core NF tools AMF · SMF · PCF (SBI)

Observability probes · DPI · eBPF

Step 0 · 2026 wrappers over EMS

Step 1 · 2027 native Core + TN tools

Step 2 · 2028+ tool auto-registration

FIGURE 9. Network MCP Platform: before (RAN-only, CMaaS/EMSmediated) and after (RAN + Core + Transport, intent-driven). Today’s AI models reach only the RAN via brittle CMaaS/EMS adapters; with the Network MCP Platform, the Optimizer, SLA Enforcer, and Fault Remediator gain full-network access through a unified tool plane with discovery, schema registry, authentication and policy, rate limiting, and audit logging.

standardize enablers – interfaces, policy, observability, audit, and rollback – not fixed AI models or toolchains; own a governed data layer across RAN, core, cloud, and OAM; evolve existing integrations first, redesign only where lifecycle value is clear and measurable. The tool catalog – Table 5 – defines what agents can READ, WRITE, and STREAM across access, core, observability, and cloud-control domains. Its role is architectural rather than vendor-specific: it identifies the minimum governed tool surface required for whole-network agency, regardless of whether a capability is reached through a legacy wrapper, a standards-based API, or a native agent interface. The migration therefore follows three abstract stages: expose existing operational systems through governed tool wrappers; replace wrappers with native tool surfaces as network functions and platforms mature; and finally allow 6Gready functions to advertise self-describing capabilities directly into the governed catalog. These stages align with the phased roadmap in Section XI. TABLE 5. NETWORK MCP TOOL CATALOG: READ, WRITE, AND STREAM ACROSS RAN, CORE, AND TRANSPORT Read

Write

vCU/NETCONF GET: vCU config push running config, cell (NETCONF editparameters, neighbor config): PSM, CIO lists, bearer state (interim offset, antenna CMaaS/EMS API) parameters, PCI

Stream / event NETCONF notifications: config-change, fault, SW-upgrade lifecycle events

Authorized standardized Authorized PCF/SMF NWDAF NF service operations procedures and governed analytics/events and and events, including OAM adapters; UPF authorized NF/OAM NWDAF analytics, plus forwarding and QoS notifications; operational telemetry rules are applied by the availability, event scope, exposed through SMF over N4/PFCP. and retention are governed OAM adapters; Any MCP wrapper is an deployment-specific no assumption of agent operator-provided, nonaccess to NF-internal standard adapter state

14

D. Production Viability: LLM Cost, Scalability, Conflict Resolution

Three systemic concerns determine production viability at operator scale; each is addressed by the architecture, not deferred. LLM cost: the LLM is engaged for intent decomposition and novel-fault resolution only – deterministic micro-agents handle routine loops with zero token spend; model tiering assigns a small fast model to KPI triage and a capable model to planning only; prompt caching on tool schemas and steadystate context cuts consumption. This tiering strategy is consistent with 6G-Bench feasibility measurements, which show that no single model simultaneously satisfies latency, throughput, and reasoning-accuracy constraints, and that quantization effects are model-specific rather than uniform, making heterogeneous device–edge–core model placement a systems-level requirement rather than an optimization [30]. A practical operator design target – stated as an unvalidated assumption to be confirmed against production event-rate data – is LLM engagement in approximately 1-5% of closed-loop cycles, with cost scaling with novel-event rate, not network size. Scalability: the Network MCP Platform is designed around stateless, horizontally scalable routing and policy services; tool calls are independent and parallel; rate limits and peragent-class quotas at the router protect underlying NFs; the Digital Twin sits in the decision path only for gated management-plane actions, while a latent world-model approach inspired by JEPA maintains compressed predictive representations of network state for fast screening, conflict pre-checks, and policy-envelope maintenance [68], [69]; this validation compute runs on parallel infrastructure that throttles neither live traffic forwarding nor the real-time control tiers; RAN, core, and transport operate in isolated toolset partitions; new closed-loop domains are added by registering intent types, without code changes. Conflict resolution: the router holds optimistic resource locks per network element, detecting conflicting write intents before the ACT step; twin pre-validation catches residual conflicts before network impact; a fixed priority hierarchy is defined once – SLA Enforcer > Fault Remediator > Coverage Optimizer > Energy Saver; unresolvable conflicts escalate to

VOLUME XX, 2026

human oversight with a full audit trail, aligned with EU AI Act Article 14 [57]; post-action monitoring detects emergent interference and applies mutual-exclusion zones automatically. Router engineering budget: the router is a stateless, horizontally scaled service tier – not a single network element – with N+K redundancy per region and SBI-only bypass as the degraded mode; the design target is single-digitmillisecond added latency on management-plane invocations (real-time control tiers bypass the router by construction, operating within pre-validated envelopes), with availability engineered to at least the level of the SBA functions it fronts. Because the router emits charging-relevant events and LI mediation hooks, it is also one of the platform’s highest-value attack targets: router compromise is treated as a first-class threat scenario, bounded by attested workload identity for router instances themselves, PQC-signed audit chains, and the SBI-only failsafe. E. Compliance by Design, Containment, and Standards Alignment

Because the router can emit charging-relevant events and LIdesign hooks on SBI and MCP paths, and the eBPF plane can source consumption telemetry from the kernel for routerbypassing real-time tiers, the audit and metering visibility that conventional signalling exposes can be extended across both router-mediated and bypass paths. However, compliant charging records, lawful-intercept handover interfaces, retention rules, warrants, and mediation functions remain open standardization and implementation items. The twin’s PQCsigned validation records and the translator’s explainability logs map directly onto high-risk-AI record-keeping and oversight obligations [57]. Adversarial containment is architecturally layered: among standardized, deployable mechanisms, kernel-level eBPF syscall tracing supports sub-millisecond detection [11], enabling router-level containment within a sub-second loop; in the RAN, an AF_XDP fast path carries L2 telemetry to Near-RT RIC inference at sub-millisecond latency, and dApps on the proposed E3 interface [20] are designed to support control within the sub-10 ms budget Mission-Critical Determinism requires. A deliberately conservative failsafe completes the design: forcing SBI-only mode and revoking the agent fleet’s identities returns the control path to standard 3GPP procedures, preserving a standards-based degraded operating mode; restoration to a known-good state is not automatic, additionally requiring explicit state snapshots, reconciliation, rollback ownership, dependency ordering, and defined recovery objectives, because disabling agents does not by itself restore prior distributed network state. The same discipline applies to the agent substrate itself: the six subsystems are protocol-substitutable by design – MCP and A2A are the current concrete bindings, not architectural commitments – so a license change, deprecation, or vertical15

Business

TM Forum · AI-Native Blueprint · TMF939/785 autonomous-network business enablers

Exposure

CAMARA open APIs · 3GPP SA6 (AIMLE · CAPIF) QoD · location · verification · AI-inference offload

Service & network mgmt

NGMN levels · 3GPP SA5 (intents · data · O&M) ETSI ZSM/ENI · O-RAN SMO (R1 · O1 · O2)

Network domains

3GPP SA2 (CN AI · UE agents) · RAN1/2/3 (Uu) O-RAN RIC/xApps · AI-RAN Alliance (non-SDO)

Underlying agent tech

IETF · Linux Foundation · MCP / A2A protocols OSS toolchains: MLflow · kagent · Argo

IN SCOPE data collection & exposure multi-vendor interfaces auth · policy · audit · rollback

AVOID specific ML models training pipelines one frozen agent architecture

FIGURE 10. Standardization scope for agentic AI in 6G: which SDO defines what – 3GPP SA2/SA5/SA6, O-RAN Alliance, ETSI ZSM/ENI, NGMN, TM Forum, CAMARA, IETF, and open-source toolchains – with the boundary principle that standards expose, secure, and govern capabilities for agents without freezing one AI-agent architecture.

integration move by any single protocol or model vendor is absorbed at the tool-plane binding, not in the architecture. Where the platform forms part of a safety-critical chain, functional-safety regimes (IEC 61508, ISO 26262, IEC 62443) apply beyond telecom SLA frameworks. The platform is designed to be standardized through, not around, existing bodies: 3GPP SBI [80], [81], security [74], charging [75], and lawful intercept [76]; O-RAN RIC and E3 [20], [19]; SPIFFE/SPIRE [77] and RFC 9728 [78]; MCP [73] and A2A [82]; and the OCUDU Ecosystem Foundation [83] for multi-vendor interoperability. The 3GPP SA2 6G study’s intent-handling key issue (KI#18) [84] independently confirms the problem space while leaving agent identity, charging, lawful intercept, enforcement latency, and transactional fulfillment open, highlighting a live standardization opportunity. VIII. 6G STANDARDIZATION: SDO LANDSCAPE AND RAKUTEN’S PUBLIC POSITION

Standardization of 6G is entering its most consequential phase: the IMT-2030 framework (Recommendation M.2160) is approved while the Technical Performance Requirements remain a WP 5D draft pending Study Group 5 approval (expected Dec. 2026) [2]-[4]; 3GPP Release 20 6G study items are active and Release 21 will carry the first normative 6G specifications, and first commercial 6G deployments are expected around the 2030 window. The decisions being fixed now determine whether 6G advances an operator-controlled platform vision or reproduces vendor dependency. This section presents Rakuten Mobile’s public position across seven dimensions, summarized in Table 6. A. Agentic AI: Standardize Enablers, Not Architectures

The AI-agent standardization landscape, illustrated in Fig. 10, is distributed across SDOs without a unified coordination VOLUME XX, 2026

framework. 3GPP SA2 studies AI-embedded core NFs and UE AI-agent registration and discovery (KI#18 [84]), while SA5 covers 6G operations and maintenance, intents, data management, and autonomous-management enablers. The O-RAN Alliance specifies SMO/R1/rApps, RIC/xApps, and O-Cloud automation, with agentic-AI evolution progressing through nGRG [19], [42]. ETSI ZSM/ENI covers agents in autonomous networks and closedloop automation [24], [32]; NGMN defines the autonomy levels [26]; TM Forum contributes the AI-native blueprint (TMF939/TMF785); CAMARA and the exposure layer sit above. The underlying agent technology – MCP/A2A-like protocols and open-source AI toolchains – evolves in the IETF, the Linux Foundation, and the Agentic AI Foundation. Rakuten’s position: standards should define the enablers and boundaries – data collection and exposure, businesscritical multi-vendor interfaces, authorization, policy, audit and rollback, management and exposure enablers, and UE/RAN/Core performance control – while avoiding overstandardization of specific ML models, training pipelines, or one frozen agent architecture. Existing 3GPP/O-RAN interfaces remain the execution path; a single frozen agent architecture standardized in 3GPP would recreate the lock-in Open RAN was designed to eliminate. B. RAN Architecture, Migration, NTN, PHY, Core, and Spectrum

RAN-core and the high-level split (HLS). The centralizedunit/distributed-unit (CU-DU) split is not only a signalingefficiency question: it determines openness, automation, and troubleshooting capability. The HLS must be standardized to preserve multi-vendor interoperability, with F1 retained and PREFERRED BASELINE 6G SA

MRSS

NR/6GR mobility

standalone 6G first simple day-1 anchor

multi-RAT spectrum sharing for reuse

coverage continuity no NSA complexity

OPTIONS UNDER STUDY Option A — 6G anchor + NR booster · may help day-1 performance Option B — 5G anchor + 6G cells · risks repeating NSA complexity Option C — dual stack outside RAN · continuity vs complexity Dual/split bearer — PDCP anchor · latency · charging still open

CORE / INTERWORKING VARIANTS AMF / SMF combined vs separate

4G/EPS interworking only for proven value cases

NTN continuity

Rel-17 NTN baseline delay/Doppler UE assistance

Rel-18/19 NR-NTN enhancements mobility · voice handheld path

NB-IoT/eMTC low-rate IoT deep coverage long-lived IoT

6G NTN target TN/NTN commonality handset-first · MRSS GNSS resilience PNT · priority mode

NR-NTN to 6G NTN

RAKUTEN DEFAULT — 6G SA + MRSS + NR/6GR mobility no mandatory feature without TCO proof · avoid migration explosion

FIGURE 11. 5G-to-6G migration and interworking: preferred baseline (6G SA + MRSS + NR/6GR mobility), additional options under study (6Ganchor booster, 5G-anchor, dual stack, dual bearer), and core/interworking variants (AMF/SMF combined vs. separate, 4G/EPS interworking, NTN continuity).

16

enhanced; a vendor-specific HLS would undermine the tool catalog of Section VII – if the vCU NETCONF/YANG surface is proprietary, the READ/WRITE tools of Table 5 cannot deploy multi-vendor, recreating the CMaaS/EMS fragmentation the Network MCP Platform is intended to remove. Rakuten can accept simple point-to-point RAN-core connectivity provided the standards supply observability and extension points (SBI) for non-connectivity services (ISAC, data/AI). Migration and interworking. Rakuten’s default is 6G Standalone plus Multi-RAT Spectrum Sharing (MRSS) for spectrum reuse and NR/6GR mobility for coverage continuity, as illustrated in Fig. 11. Of the options under study, a 6Ganchor booster may help day-1 performance (conditionally acceptable); a 5G-anchor risks repeating the non-standalone (NSA) complexity of the 5G transition; dual stack outside the RAN helps continuity but raises UE/core complexity; and dual/split bearer raises unresolved Packet Data Convergence Protocol (PDCP) anchor, latency, policy, and charging questions – none should become mandatory without totalcost-of-ownership (TCO) proof. On core variants: single registration is the baseline; combined AMF/SMF may simplify context transfer but risks complexity; separate functions preserve clean architecture but need robust mobility. The filter throughout: avoid migration explosion; support only proven value cases. Non-Terrestrial Network (NTN). Evolution runs from the Rel-17 NR-NTN baseline through Rel-18/19 enhancements and the NB-IoT/eMTC NTN low-rate path to a 6G NTN target of maximum TN/NTN commonality, as illustrated in Fig. 12. The 6G target requires day-1 NTN-ready basic support, handset-first design before very-small-aperture-terminal (VSAT)-heavy optimization, and MRSS with NR NTN alongside global-navigation-satellite-system (GNSS) resilience, positioning-navigation-and-timing (PNT) enhancement, and priority mode. NTN is for resilience and coverage extension – disaster recovery, maritime, IoT deep coverage – not satellite-specific overdesign that inflates device cost; 6G NTN is an evolution, not a start from zero. PHY and devices. The direction captured in the 3GPP 6G workshop discussions reuses the NR baseline – DL CPOFDM, UL CP-OFDM/DFT-s-OFDM, NR LDPC/Polar where technically justified – with FR1/FR2 focus and site-grid

already standardized: NR-NTN & NB-IoT/eMTC NTN

expected in 6G

Use cases — coverage TN cannot economically serve disaster resilience · maritime · deep-coverage IoT · emergency comms

FIGURE 12. NTN evolution: NR-NTN Rel-17 baseline → Rel-18/19 enhancements → NB-IoT/eMTC NTN low-rate IoT path → 6G NTN target with maximum TN/NTN common design, handset-first evolution, and MRSS/GNSS-resilience/PNT/priority-mode requirements.

VOLUME XX, 2026

reuse as the deployment-economics constraint. Nokia’s workshop analysis [39] projects 6–10 dB uplink gain through 6G migration on existing spectrum – a vendor projection pending independent validation, but one that would deliver coverage without new sites; Qualcomm’s chipset-level priorities align [40]. Open items – low peak-to-averagepower-ratio (PAPR) uplink (frequency-domain spectral shaping, FDSS, and DFT-s-OFDM enhancements), higherorder quadrature amplitude modulation (QAM) and shaping, new low-density parity-check (LDPC) code extensions, synchronization-signal-block (SSB) periodicity and raster around new bands, and low-tier UE bandwidth/RF-chain/HDFDD trade-offs – must each justify complexity with quantified gains. The operator priority: coverage first (uplink, cell edge, control channels), efficiency with evidence (average-user and loaded-cell gains over peak-rate optics), and no hidden cost in DU/UE compute, RF complexity, or energy. PHY choices should be deployment-gated by real coverage, capacity, and cost outcomes, not feature count. Integrated Sensing and Communication (ISAC). Sensing becomes useful only if measurements, fusion, exposure, and governance are standardized consistently across the Sense → Measure → Fuse → Expose → Assure chain. Open questions span where fusion is done (RAN node, sensing function, core, SMO, or distributed), which measurement level is standardized (object-level results versus lower-level radio observables for AI fusion), how UEs are involved, and how data is governed – authorization, consent, privacy, and charging (3GPP Rel-20 studies, TR 23.801-01/TR 38.914 discussions). ISAC decides who owns sensing data, fusion logic, and exposure control; fragmented vertical solutions would prevent Sensing-as-a-Service from reaching commercial scale. Standalone (SA)/Core. SA2 is deciding who controls 6G intelligence, data, exposure, and continuity. The safe baseline: start from 5G SBA and non-access-stratum (NAS) experience, retain the Packet Forwarding Control Protocol (PFCP) for SMF-UPF stability, and add exposure, policy, and authorization for AI and compute as incremental extensions. Open items include new versus evolved NAS, UPF SBI versus PFCP, AI-embedded core versus a separate AI subsystem, and UE AI-agent registration/discovery. Rakuten’s position: evolve first, redesign only where lifecycle value is measurable; standardize enablers, not one AI-agent architecture; keep compute tied to QoS, edge availability, and policy; own a governed data layer across RAN, Core, Cloud, and OAM; prefer open SBA/SBI evolution and avoid proprietary 6G control islands. Spectrum and timeline. Japan is investigating 7125–8400 MHz and 14.8–15.35 GHz for 6G [41]. Since sufficiently wide new spectrum may not be available for day-1 deployment around 2030, combined use of existing and new frequencies will be necessary – and 5G (in some scenarios 4G) will still be operating in both coverage and capacity bands in 2030, making MRSS an operational necessity rather than an option. 17

TABLE 6. RAKUTEN MOBILE’S PUBLIC POSITION ON SEVEN PRINCIPAL 6G SDO ARCHITECTURE DECISIONS Dimension

Established Open items predecessor baseline or strong study direction

Rakuten position

Agentic AI

Enablers: data, exposure, authorization, audit, rollback

AI-in-NF vs. subsystem; UE agent registration; cross-SDO data framework

Standardize enablers and governance, never one frozen agent architecture

RAN-Core / HLS

O-RAN-aligned disaggregation; open F1

Non-connectivity service interfaces (ISAC, data/AI)

Standardized HLS; P2P acceptable with observable, documented SBI extensions

Migration

6G SA first

6G-anchor 6G SA + MRSS + booster; 5GNR/6GR mobility; anchor; dual stack; extras only with TCO dual bearer proof

NTN

Rel-17→19 NRNTN evolution to 6G NTN

Feeder/ISL assurance; TNNTN convergence detail

Day-1 basic support; maximum TN/NTN commonality; handset-first; no overdesign

PHY / devices

NR waveform + coding reuse; FR1/FR2 focus

Low-PAPR UL, QAM/shaping, sync/raster, lowtier UE

Coverage first; evidence-gated efficiency; no hidden compute/power/site cost

SA/Core

5G SBA/NAS as starting point; PFCP retained

NAS evolution; UPF SBI; AI subsystem placement; Compute-aaS

Evolve first; open SBA/SBI; governed data layer; no proprietary control islands

Spectrum (Japan)

7125–8400 MHz and 14.8–15.35 GHz under study [41]

WRC-27 outcomes; day-1 bandwidth

Combined legacy + new spectrum; MRSS as operational necessity

The high-level timeline – M.2160 framework (2023), draft TPR (Feb. 2026, pending SG5 approval Dec. 2026), Release 21 6G Stage 1 freeze (Mar. 2027), Stage 3 functional freeze (Dec. 2028), ASN.1/OpenAPI freeze (Mar. 2029), commercial deployments around 2030 – aligns with the roadmap of Section XI; WRC-27 is the critical external dependency. IX. TECHNOLOGY LAST – THE END-TO-END 6G ARCHITECTURE A. The End-to-End Architecture

Technology Last argues that every technology choice must be derived from the operator, customer, business, and operational requirements of Sections III–VI, not from vendor research roadmaps. The derivation is explicit: Immersive Experience’s end-to-end application budget of ≤20 ms [85] – of which the network contribution must respect the TPR ≤4 ms user-plane latency [4] – mandates dense mid-band/mmWave and dApp co-deployment; Mission-Critical Determinism’s 1−10⁻⁵ TPR reliability floor (with stricter contractual sub-cases) mandates ISAC-based redundancy and predictive maintenance; AINative Communication’s <10 ms inference latency requires AI-on-RAN edge compute (the strict sub-10 ms bound applies VOLUME XX, 2026

B. AI-RAN: The Intelligent Radio Platform

AI SUBSTRATE kagent · MLflow · MCP · licensed foundation models · multi-agent manager

SERVICE EXPOSURE CAMARA · Open Gateway · ISAC-aaS · inference offload · digital twin

UE devices · XR IoT · UAV

RAN

INTELLIGENCE

CORE

O-RU · O-DU O-CU · dUPF AI-RAN node

Non-RT RIC · SMO Near-RT RIC · xApps dApps · E3 <10 ms

5GC to 6G SBA AMF · SMF · UPF NEF · NWDAF

EDGE

CENTRAL CLOUD

MGMT & ORCH

MEC · local breakout AI-on-RAN inference

hyperscaler IaaS AI training · ops

ZSM · Nephio NGMN L4/L5 ops

TRANSPORT NETWORK optical · SRv6 · fronthaul · backhaul · x-haul

FIGURE 13. End-to-End 6G Architecture. AI Substrate (top) → Service Exposure → UE → RAN → Intelligence (Non-RT/Near-RT RIC/dApps) → Core → Edge → Central Cloud → Management, with the transport network at the bottom.

to RAN-local, dApp-tier inference; edge-cloud paths serve the remainder of the latency budget); and Sensing’s sub-meter positioning mandates wideband signal processing. The architecture in Fig. 13 extends cloud-native disaggregation to the full stack – AI substrate on top (kagent, MLflow, MCP, foundation models, multi-agent manager), service exposure (CAMARA, Open Gateway, ISAC-aaS, inference offload, digital twin), UE, Open RAN (O-RU/ODU/O-CU, distributed UPF, AI-RAN node), three-tier intelligence (Non-RT RIC/SMO, Near-RT RIC, dApps), service-based core (5GC SBA → 6G SBA), edge, central cloud, management and orchestration (ETSI ZSM, Nephio, NGMN L4/L5), and transport. Three characteristics distinguish it from 5G. First, RAN-AI compute convergence: RAN functions co-hosted with AI inference on shared accelerated hardware [12]; a dUPF colocated with the RAN node enables local breakout to AI-onRAN workloads within the sub-10 ms budget; Harkous et al.’s flat-UP architecture [22] goes further, collapsing the GPRS Tunneling Protocol (GTP) tunnel between CU-UP and UPF with measured throughput, latency, and compute gains. Second, hierarchical multi-timescale intelligence: Non-RT RIC (seconds-minutes), Near-RT RIC [19] (10 ms-1 s), and dApps [20] (<10 ms). Third, unified standards-governed service exposure extending Open Gateway [45] and CAMARA [53] with 6G interfaces. A candid maturity note: published studies document Open RAN throughput gaps versus integrated RAN in some dense multi-vendor configurations; the argument is not present equivalence but that AI-for-RAN optimization, integration tooling, and interface stability are converging on a trajectory that can make Open RAN a credible long-term choice for operators prioritizing control, with Salmi-style conflict management [18] inside the RIC as a precondition for honoring GuaranteeEconomy SLOs across overlapping xApp scopes.

18

As shown in Fig. 14, AI-RAN [12] formalizes two complementary function classes on one accelerated cloudnative platform, creating the possibility that the cell site contributes revenue as well as reduces cost. AI-for-RAN optimizes the radio: neural channel estimators outperform minimum-mean-square-error (MMSE) estimation in complex massive-MIMO channels; learned channel-state-information (CSI) feedback compresses at higher fidelity and lower uplink overhead; transformer beam predictors anticipate mobility; deep-RL schedulers co-optimize throughput, fairness, and energy; and AI cell-sleep and antenna muting target deployment-specific energy-per-bit reductions, consistent with the 6G direction set by TPR [4], but no universal 10x improvement should be inferred; the mechanism class is production-validated by Rakuten Mobile’s nationwide RIC deployment (15-20% network power reduction; approximately 20% RAN energy savings under TM Forum GB1059H Level 4 validation; see Section III.C) [5]. DMIMO, RIS, and XL-MIMO are feasible only with ML codesign. AI-on-RAN monetizes the co-located accelerated compute for edge AI workloads [12], [14], [42], [71], [86] across five deployment profiles [14]: interactive GenAI (lowlatency uplink voice/multimodal), always-on agentic AI (sustained uplink, context persistence), physical AI (sub-10 ms safety-critical actuation), remote inspection (high-quality video uplink with real-time analysis), and batch broadcast inference (high-throughput downlink). Each maps to a Guarantee Economy tier, suggesting a commercial architecture from AI workload to network product. That commercial architecture remains a projection from demonstrated capability and published demand research rather than a reported large-scale market outcome. AI-FOR-RAN

AI-ON-RAN

optimize the radio

monetize edge compute

Interactive GenAI

PHY neural CSI · beam prediction

low-latency voice · multimodal

MAC / RRM

always-on · sustained uplink

Agentic AI deep-RL schedulers · energy-aware

Physical AI sub-10 ms closed-loop control

Energy AI cell sleep and muting benefit is study-specific

Remote inspection video uplink · AI analysis

Antennas D-MIMO · RIS · XL-MIMO co-design

Batch inference high-throughput broadcast

UNIFIED AI-RAN PLATFORM GPU·DPU·ASIC · real-time OS · CaaS · AI runtime · Orchestrator Rel 18 · AI-assisted

Rel-19 context AI integration and extensions

Future research AI-substituted or AI-native PHY; no release commitment

FIGURE 14. AI-RAN: AI-for-RAN and AI-on-RAN on a Shared Platform. Dual value streams on one accelerated cloud-native platform (GPU/DPU/ASIC, real-time host OS, CaaS, AI runtime) and AI-RAN Orchestrator; release references are maturity context only, not commitments for specific AI-substituted PHY functions.

VOLUME XX, 2026

The economic case rests on three compounding advantages over central-cloud serving: physical proximity (central roundtrips add an estimated 30–80 ms against the 10 ms budget of the AI-RAN WG3 remote-inspection profile [14]); a shared cost basis (AI-for-RAN energy and capacity gains subsidize the accelerated hardware AI-on-RAN monetizes); and data sovereignty (medical images, personnel video, and transaction records frequently cannot leave the facility, making edge inference compulsory rather than preferable). Realizing the moat requires operators to own the AI substrate; purchasing inference from a vendor-managed service converts the sovereign edge advantage into a managed service at vendorset prices. The standards context frames a possible path, not a commitment: Release 18/19 include AI-assisted CSI/beam work within 5G-Advanced, Release 20 studies 6G AI topics, and any Release 21 normative treatment of AI-substituted blocks or AI-native network management would depend on future 3GPP agreement and demonstrated gain. C. The AI-Native Air Interface: Three Stages

As illustrated in Fig. 15, the path to an AI-native physical layer is presented as a three-stage author research maturity model informed by 3GPP 6G study work [38], not as an approved release commitment. Stage 1 – AI-Assisted Enhancement (3GPP Release 18/19 | 2024–2026): AI augments blocks within a conventional specification – neural channel estimation, learned CSI compression, beam prediction – deployable via software on existing hardware, funding the later stages. Stage 2 – Candidate AI-Substituted Functions (3GPP Release 20/21 | 2025-29): learned equalizers, neural decoders, and deep-reinforcement-learning schedulers are candidates to replace MMSE, LDPC/Turbo, and rule-based mediumaccess-control (MAC) scheduling, each gated on measurable gain, fallback, and lifecycle value; standardize inputs/outputs, not algorithms; requires accelerated compute. Release 20 is the 6G study phase; any normative treatment would depend on future 3GPP agreement. Stage 3 – AI-Native End-to-End (3GPP Rel 21+ | 202932+): transmitter and receiver jointly optimized as a learned communication system, continuously adapting to channel, interference, and service – the enabling technology for cyberphysical fusion [15]. Zheng et al. [17] provide the concrete Stage 3 realization: cross-module optimization of uplink learned source-channel coding with modulation, downlink modulation with precoding and CSI feedback, and control agents performing model switching by channel state – a natural dApp target, since model switching at channel-state granularity needs the sub-10 ms E3 loop. Stage 3’s critical challenge is multi-vendor interoperability of jointly learned systems across UE chipset and infrastructure vendors – an open research problem (Section XII). Device timing constrains commercialization. Applying the 5G SA precedent (standardized in 2018; meaningful handset penetration in 2022–2023) and the 18–24-month silicon cycle, consumer AI19

STAGE 1

STAGE 2

STAGE 3

AI-assisted

AI-substituted

AI-native E2E

Rel 18/19 · 2024-26

Rel 20/21 · 2025-29

Rel 21+ · 2029-32+

AI augments blocks within conventional spec · deploy today

AI replaces DSP blocks · standardize I/O, not algorithms

TX/RX jointly learned · adapts to channel & service

Examples Neural channel est. learned CSI feedback AI beam prediction

Examples Learned equalizers neural decoders deep-RL schedulers

Examples Learned waveforms joint comms·sensing online adaptation

Hardware impact: software upgrade

Hardware impact: accelerated compute

Hardware impact: AI-native PHY platform

FIGURE 15. AI-Native Air Interface – A Three-Stage Evolution. Stage 1 (Rel 18/19 AI-assisted enhancements) → Stage 2 (6G study topic; AIsubstituted functions requiring accelerated compute; no approved release commitment) → Stage 3 (longer-term research toward end-toend learned transmit-receive chains; no approved release commitment).

native services at scale are a Phase 3 objective (around 2030 and beyond). Enterprise and IoT devices, for which custom silicon is procurable, accelerate ahead – which is why Phase 2 Guarantee Economy revenue depends predominantly on enterprise, industrial IoT, and fixed wireless (Section XI). D. dApps and the E3 Interface

The Near-RT RIC’s 10 ms–1 s loop is insufficient for realtime spectrum sharing, ISAC feedback, and deterministic control. The dApps framework (nGRG RR-2025-05 [20]) introduces a control tier executing directly on DU/CU infrastructure, as depicted in Fig. 16, with three E3 procedures: Registration (capability and resource declaration to the host CU/DU, enabling lifecycle management without bespoke integration), Subscription (user-plane data and telemetry streams at sub-millisecond cadence over Unix-domain-socket exchange), and Control Message (spectrum allocation, beam steering, and slice admission inside the sub-10 ms cycle). The nGRG prototype on OpenAirInterface validated real-time spectrum sensing and sub-meter positioning over the air [20]. The mapping to Section IV is precise but separates two latency budgets: Mission-Critical Determinism retains the radio/userplane latency target in Table 2, while dApps are proposed to provide a sub-10 ms E3 control-loop budget for control actions; Sensing requires direct user-plane access; AI-Native NON-RT RIC

NEAR-RT RIC

s – min

10 ms – 1 s

dApps < 10 ms

NEW dApps

Workload

rApps

Workload

xApps

Workload

Interface

A1 / O1

Interface

E2

Interface

E3

Loop

s – min

Loop

10 ms – 1 s

Loop

< 10 ms

Policy · AI model lifecycle · crossdomain optimization

Per-cell / per-UE inference · slice admission · steering

Real-time spectrum sharing · ISAC · beam steering

E3 PROCEDURES

Registration · Subscription · Control Message

FIGURE 16. [INDUSTRY RESEARCH / PROPOSED] dApps and the E3 Interface – proposed sub-10 ms edge control. Three intelligence tiers – Non-RT RIC/rApps (A1/O1, s–min), Near-RT RIC/xApps (E2, 10 ms–1 s), dApps (E3, target <10 ms) – plus the E3 procedures: Registration, Subscription, and Control Message. VOLUME XX, 2026

Communication benefits from in-node inference. Deployment readiness requires an E3-exposing CU/DU implementation, a dApp certification framework comparable to xApp onboarding, and a conflict-resolution mechanism arbitrating dApp actions against xApp policies – the Salmi-style engine [18] lifted into the dApp domain. Highest-tier Determinism SLAs cannot be priced credibly without a control plane that intervenes inside the radio-frame cadence.

SUB-THz · >100 GHz mmWave · 24–100 GHz urban · indoor · FWA

FR3 · 7–24 GHz primary 6G capacity

FR1 · sub-6 GHz coverage + ISAC

E. Spectrum Strategy: Coverage, Capacity, and Sensing

The spectrum strategy, illustrated in Fig. 17, is layered against the outcome families. FR1 (up to 7.125 GHz) is the coverage and ISAC foundation: proven 3.5/4.9 GHz mid-bands plus legacy sub-3 GHz deliver indoor penetration, reach, and the 10⁶ devices/km² density target [4]; the TPR positioning targets (0.75 m indoor factory, 6 m urban macro, 95% detection) [4] are achievable through wideband sub-6 GHz processing. The upper mid-band (7–24 GHz — an informal industry grouping, not an agreed 3GPP range) is the 6G primary capacity layer; WRC-27 Agenda Item 1.7 studies specific candidate bands — 4.4–4.8 GHz, 7.125–8.4 GHz, and 14.8–15.35 GHz — not the entire 7–24 GHz range. mmWave (24–100 GHz) serves dense urban, indoor, and fixed-wireless-access (FWA) hotspots, where the draft TPR peak-rate and low-latency figures (DL 36 / UL 18 Gbit/s; 1 ms HRLLC) are approached only under favorable hotspot conditions, not as intrinsic mmWave guarantees. Sub-THz (>100 GHz) remains extreme-capacity research – carrier bandwidths of up to 10 GHz for holographic and extreme XR. NTN constellations integrate with terrestrial assets – coordinated through the RIC framework – on a deployment-dependent basis for ubiquity and resilience, not as blanket coverage across all bands. WRC-27 contingency planning is essential: upper mid-band identification faces structurally organized regional opposition, so operators should model delayed or partial outcomes as the base case, sizing Phase 2 on existing allocations plus AI-forRAN spectral-efficiency gains, with WRC-27 success treated as material upside rather than prerequisite. Operators and administrations must engage the European, Asia-Pacific, and African regional preparatory processes (CEPT, APT, and ATU) now – while planning as though outcomes will be partial or late.

10 GHz carriers · extreme XR

extreme capacity · R&D

2024

36 / 18 Gbit/s peak · 1 ms HRLLC

NTN

ISAC + eMBB+ · WRC-27 · TPR

spans all bands

10⁶ dev/km² · 0.75 m positioning WRC-27

IMT-2030

2027

2030

2033

FIGURE 17. [STANDARDS-GROUNDED / DRAFT TPR] Spectrum Strategy – Coverage, Capacity, and Sensing. FR1 up to 7.125 GHz (coverage + ISAC foundation), upper mid-band 7–24 GHz as an informal industry grouping with WRC-27 studying only specified candidate bands, mmWave 24–100 GHz (deployment-specific dense urban/indoor/FWA), sub-THz >100 GHz (extreme capacity research), with NTN integration deployment-dependent.

F. The Near-RT RIC Toward 6G

The Near-RT RIC evolves for 6G [19] with enhanced AI/ML support (federated learning, distributed edge training), expanded service exposure, unified data management, multiRIC interworking, and Communication and Computing Integrated Network (CCIN) support enabling joint radio-andcompute reservation – the architectural precondition for endto-end Guarantee-Economy SLA enforcement. Cross-RIC federation lets an enterprise-campus xApp coordinate with the surrounding macro RIC so mobility preserves the guaranteed SLO without unfiltered access to public-network telemetry – cross-domain AI coordination without cross-domain raw-data sharing, a privacy-preserving pattern that can support GDPR [58] and EHDS [59] compliance-by-design when lawful basis, purpose limitation, retention, transfer, and sector-specific obligations are satisfied. Three extensions beyond current specification are needed: federated learning at the RIC tier itself; formal certification of xApp interaction graphs to make conflict resolution auditable at third-party scale; and first-class observability of the RIC itself via eBPF instrumentation [10], [11] – the engineering path to sub-millisecond drift detection, supporting the trust-management and tamper-evidence objectives of ETSI ZSM 017 [32]. X. TRUST, SOVEREIGNTY, AND SUSTAINABILITY A. Zero-Trust for Autonomous 6G

In 4G/5G, security was a perimeter concept; agentic AI demolishes it. At Level 4 autonomy, compromising the telemetry feeding Observe, poisoning the model behind Decide, or injecting commands into Act requires only access to the closed-loop automation infrastructure – not a perimeter breach. ETSI ZSM 017 [32] maps risks across each loop stage and proposes three interlocking mechanisms, illustrated in Fig. 18: Closed Loop Trust Management (per-domain cryptographic trust certificates, immutable audit trails, interloop identity verification), Closed Loop Access Control 20

VOLUME XX, 2026

(attribute-based least-privilege policies over telemetry reads and configuration writes), and Closed Loop Security Exposure (a real-time security-posture interface enabling automated anomaly detection and human audit under EU AI Act Article 14 [57]). The convergence is useful but bounded: ZSM 017 is an ETSI Group Report, not an AI Act conformity standard, so ZSM 017-aligned controls may contribute technical evidence relevant to an AI Act conformity assessment but do not by themselves establish legal conformity – a partial, not a dual, return. ZSM 017 does not cover LLM-specific threats: prompt injection, tool poisoning, and goal misgeneralization [36]. These require adversarial testing pipelines, output guardrails validating configuration commands against policy whitelists before execution, sandboxed tool execution, and immutable logs for forensic reconstruction – deployment prerequisites, not post-deployment items, though their efficacy against production telecom attack patterns remains to be empirically validated (Section XII). Post-quantum cryptography – NIST-finalized ML-KEM [34], ML-DSA [35], and SLH-DSA [79] – underpins the trust hierarchy and is identified in 3GPP TR 22.870 [8] as a security expectation for IMT-2030 systems; with cryptanalytically relevant quantum computers expected within 6G’s operational lifetime, migration is an operational priority today. Choudhary et al. [33] show the migration problem is operational rather than algorithmic: hybrid post-quantum AKA variants have measurable overheads compatible with commercial platforms, but cut-over sequencing across RAN, subscription permanent/concealed identifier (SUPI/SUCI) concealment, roaming, and lawful-intercept conformance must be coordinated – absent a 3GPP SA3 study item, operators will sequence divergently at predictable interoperability cost. And cryptographic resilience does not imply AI resilience [36]: an input-manipulable scheduling agent compromises service delivery even under post-quantum-strong cryptography.

Monitor

Analyze

Decide

Execute

telemetry

AI anomaly

LLM planning

config push

ETSI ZSM closed-loop automation · mechanisms span every stage

TRUST MANAGEMENT

ACCESS CONTROL

SECURITY EXPOSURE

per-domain certificates · immutable audit trail

attribute-based policies · least privilege per agent

security-posture API · anomaly alerts · human audit

Beyond ZSM 017 — LLM-specific threats: prompt injection · tool poisoning · goal misgeneralization → guardrails · sandboxed tools · adversarial testing

PQC FOUNDATION LAYER ML-KEM · ML-DSA · SLH-DSA · NIST 2024 · 3GPP TR 22.870 never trust · always verify

FIGURE 18. Zero-Trust Architecture for Autonomous 6G. Three ETSI ZSM 017 mechanisms – Closed Loop Trust Management, Access Control, and Security Exposure – spanning every Monitor → Analyze → Decide → Execute stage; the PQC foundation layer (NIST 2024; 3GPP TR 22.870) underpins the trust hierarchy.

21

SERVICE CO₂/episode · Scope 1/2/3

SYSTEM TEER · NCIe · ERF · network-wide

COMPONENT

SNS JU eHealth 40–60% CO₂ per clinical episode modeled projection

IOWN GF MetricEES unified EE metrics

energy/bit · sleep states · muting

FIGURE 19. Sustainability KPI Framework – Component, System, Facility, and Service levels. Pyramid integrating component metrics (energy/bit, sleep states), system/facility metrics (TEER, NCIe, ERF, PUE; MetricEESinformed harmonization), and service-level attribution research (carbon per clinical episode, Scope 1/2/3 accounting), with the SNS JU eHealth 40–60% CO2-per-episode evidence callout.

B. Data Sovereignty and Regulatory Compliance

Data sovereignty is now a primary procurement criterion for the highest-value Guarantee Economy segments. In the EU, GDPR [58], EHDS [59], and the EU AI Act [57] converge: critical-infrastructure AI may be high-risk depending on intended purpose and Article 6/Annex III classification, triggering human oversight (Article 14), robustness (Article 15), and pre-deployment conformity assessment where applicable. National-security policy is a structural accelerant – the EU 5G Toolbox [87], the US Secure and Trusted Communications Networks Act [62] and CHIPS Act [63], and analogous frameworks in the UK, Australia, Japan, Korea, and India [56] all increase scrutiny of supplier provenance and may impose high-risk-vendor restrictions depending on jurisdiction and asset class; operator-owned software materially reduces exposure to high-risk-supplier exclusion – though its open-source supply chain remains within the scope of the Cyber Resilience Act [61] and provenance regimes – broadly aligning regulation with the Control Compact. Three dimensions deserve operational treatment. First, lawful intercept for AI-mediated services: when an agent negotiates a slice or executes an SLA, the preservation obligation extends to the agent’s decision rationale, not only user traffic – the explainability stack (auditable per-decision provenance, intent-to-code records [29], and componentmodel XAI [37]) is proposed here as the natural evidence artifact for such obligations – an extension current lawfulintercept frameworks have not yet codified, and one that requires engagement with ETSI TC LI rather than assertion. Second, cross-border data movement for federated learning: high-risk training-data auditability, data-minimization duties, and member-state localization constraints make federated architectures a strong privacy-preserving option, subject to purpose, legal basis, transfer, retention, and sector-specific assessment. Third, the EU AI Act’s staged entry into application (prohibited practices from Feb. 2025; GPAI obligations from Aug. 2025; high-risk requirements deferred by the 2026 Digital Omnibus to 2 Dec. 2027 for stand-alone Annex III systems and 2 Aug. 2028 for high-risk AI integrated into regulated products, with transitions continuing thereafter) VOLUME XX, 2026

creates a moving compliance target that the Control Compact and MLOps stack absorb through configuration rather than infrastructure change. Jurisdiction-specific guidance evolves on a 12–24-month cycle; multi-jurisdiction operators must design the catalog as regionally parameterized templates (Section IV.C) and treat any static regulatory summary as a screening tool, not a determination. A fourth dimension is specific to sensing: ISAC observes non-subscribers – pedestrians, vehicles, premises – who have no contractual relationship with the operator and cannot meaningfully consent. Commercializing Sensing-as-a-Service therefore requires a lawful basis beyond contract – in GDPR terms, a legitimate-interest or public-task analysis with a dataprotection impact assessment – purpose limitation enforced in the exposure layer, and governance that pre-empts surveillance creep; absent that framework, the Sensing tier’s addressable market narrows to use cases where bystander data can be avoided or anonymized at the point of measurement. TABLE 7. 6G SUSTAINABILITY KPI CATALOG WITH PRIMARY STANDARD OWNERS KPI

Primary source

Measurement boundary

TEER

ATIS0600015

Equipment under Throughput/W; Operatora prescribed equipment defined vs. weighted load benchmark, not a own 5G GHG-scope baseline metric

Network energy efficiency

ITU-R TPR Operational §4.19; ETSI network; low/no ES 203 228 load vs. a fullyloaded reference case

NCIe

ITU-T L.1333

ERF

Facility/dat Facility heata-center reuse boundary standards

PUE

ISO/IEC 30134-2; facility standards

Cell-sleep utilization

Operator / Cell/carrier/RU O-RAN time by sleep measureme state nt

GHG Scope 1/2/3

GHG Protocol Corporate Standard; CSRD [88]

22

Unit / status

Referencealigned target

Relative energy consumption, % (evaluation metric)

No universal % target — a draft-TPR evaluation method

Explicit network kgCO2e/GB; boundary + emissions emission factor boundary must be declared

Track vs. ITU-T L.1470 Parisaligned −45% (2020→203 0) trajectory

Dimensionless fraction; facility efficiency metric, not a GHG-scope metric

Applicable only where useful heat recovery exists

DataDimensionless center/facility ratio; facility energy boundary efficiency metric, not a GHG-scope metric

Minimize where facility owned/contr olled; not a radionetwork KPI

% time; operational utilization metric, not a GHG-scope metric

Operatorspecific; not a universal standard target

Corporate tCO2e by scope; Report inventory: direct, corporate against the purchasedaccounting, not a applicable energy, and network or inventory service KPI boundary

KPI

Service carbon intensity

Primary source

Measurement boundary

Unit / status

Referencealigned target

and applicable ESRS

value-chain emissions

and transition plan; do not infer a service-level allocation or universal operator target

Research gap

Session, slice, gCO₂e per API, or delivered service unit outcome

Proposed standardizati on objective (no current standard target)

C. Sustainability as Competitive Architecture

Sustainability has moved from responsibility commitment to compliance criterion: the EU CSRD [88] mandates auditable Scope 1/2/3 disclosure, and the ITU-R TPR [4] treats energy efficiency as a formal IMT-2030 performance dimension and draft evaluation requirement, rather than a universal deployment guarantee. As illustrated in Fig. 19 and defined in Table 7, the KPI framework spans three levels. Component level: energy per bit, idle-versus-active power, cell sleep, and antenna muting (3GPP and O-RAN reporting). System level: the telecom energy efficiency ratio (TEER, ATIS-0600015), network carbon intensity (NCIe, ITU-T L.1333), energy reuse factor (ERF, facility standards), and power usage effectiveness (PUE, facility/data-center boundary), which IOWN GF MetricEES [55] reviews and seeks to harmonize. Service level: carbon per delivered outcome and greenhouse gas (GHG) Scope 1/2/3 accounting for Corporate Sustainability Reporting Directive (CSRD) reporting. Production evidence anchors the network layer: Rakuten Mobile’s nationwide RIC deployment demonstrated 15–20% power reduction, with TM Forum-validated Level 4 operations achieving approximately 20% RAN energy conservation (Section III.C) [5]. Beyond the network itself, the SNS JU eHealth program [67] projects 40–60% CO₂ reduction per clinical episode and 15–25% healthcare cost reduction at scale (modeled projections from the same study portfolio), showing that the 6G sustainability case extends beyond the network’s own consumption to vertical-service enablement. The three concerns of this section are architecturally convergent: zero-trust requires the same operator-owned data layer that sovereignty compliance demands, which is the same telemetry pipeline that auditable sustainability reporting requires, which is the same observability infrastructure agentic operations need. Operators that run trust, sovereignty, and sustainability as separate compliance workstreams will build redundant, inconsistent systems satisfying no one credibly. The weakest link is service-level energy attribution – watts and CO₂e per delivered service outcome (a streaming session, a control-loop iteration, an API invocation). It requires persession/per-slice/per-API energy telemetry, allocation models VOLUME XX, 2026

for shared infrastructure, and an audit-grade verification framework; none is operational at scale today, and without it the Guarantee Economy cannot price sustainability as a firstclass commercial dimension (Section XII). XI. FROM VISION TO DEPLOYMENT – ROADMAP AND STAKEHOLDER IMPLICATIONS A. Phased Deployment Roadmap

The roadmap, illustrated in Fig. 20, separates two timelines the industry often conflates: a July 2026 planning snapshot of standards and spectrum work, and the operator's own deployment targets. TR 38.914 v1.0.0 is cited as a 6G RAN study reference, not as a Rel-20 milestone. WRC-27, the illustrative Rel-21 planning gates shown for Mar. 2027, Jun. 2028, Dec. 2028, and Mar. 2029, and the ITU-R candidateRIT submission and evaluation period over 2027-2029 are planning assumptions subject to 3GPP, ITU-R, and conference decisions, not fixed commitments. The deployment phases mirror the three evolution phases of Section III.B; the profile shown is a deployment-grounded projection for an operator whose cloud-native infrastructure is already in production - an example, not an industry average [5]. Phase 1 (2025-2027): 5G-Advanced deployment and 6G studies. Releases 18-19 [89] deliver reduced-capability (RedCap) devices, ISAC study items, and energy-saving primitives, while CAMARA/Open Gateway exposure work proceeds in parallel outside 3GPP. 6G study planning proceeds in parallel; TR 38.914 v1.0.0 is used here as a 6G RAN study reference, not assigned to a Rel-20 milestone. Priorities include national-scale cloud-native Open RAN with Nephio-based lifecycle management; AI-for-RAN energy optimization with MetricEES-informed TEER/NCIe reporting [55]; first CAMARA Quality on Demand, Location Verification, and Location Retrieval APIs; an agenticoperations stack with Digital Twin validation and Level 4 preparation; and PQC pilots. Contribution to Nephio [47],

Sylva [48], Cilium/Tetragon [49], O-RAN SC [50], and kagent [51] is strategic because operators that shape reference implementations retain architectural influence. Procurement should require evidence appropriate to the purchased component: CNTi CNF conformance where applicable; Kubernetes conformance, security, and workload tests defined in operator requirements rather than a purported CNCF 'Kubernetes AI conformance' program; and, for eligible ORAN products and systems, O-RAN Alliance conformance, interoperability, or end-to-end certificates or badges issued by an OTIC. O-RAN SC results may support engineering qualification but are not the O-RAN Alliance certification route. Phase 1 investment appears in CAPEX/OPEX before revenue; boards should model it as foundational infrastructure with returns tested against explicit deployment gates. Phase 2 (illustrative Mar. 2027-Mar. 2029): future specification work and product development. This author planning window uses provisional Rel-21 gates (Stage 1 Mar. 2027, Stage 2 Jun. 2028, Stage 3 Dec. 2028, and interface freeze Mar. 2029) that remain subject to 3GPP finalization. It also covers chipset, device, and infrastructure development, pre-commercial trials, WRC-27 follow-up, and the ITU-R candidate-RIT evaluation period; 6G remains pre-commercial here. In parallel - as an operator company target, not a standards milestone - the Guarantee Economy can commercialize on mature 5G-Advanced, with revenue depending predominantly on enterprise, industrial IoT, and fixed wireless. Business cases should retain base-case and delayed-standards scenarios; in the slip case, operator commercialization proceeds on 5G-Advanced alone and 6G defers further into Phase 3. Phase 3 (around 2030 onward): initial commercial 6G and scaling. Commercial 6G introduction begins in this window; its timing depends on spectrum availability, device-ecosystem readiness, and the business case, not on standards alone. Scaling brings nationwide coverage, mature AI-RAN, NGMN Level 5 governed autonomy, and the Guarantee Economy as the default enterprise commercial model; Near-RT RIC

STANDARDS & SPECTRUM MILESTONES Rel-20 6G studies

WRC-27

Rel-21 package + Stage 1 freeze

Stage 2 freeze

Stage 3 freeze

ASN.1 / OpenAPI freeze

2027

Mar 2027

Jun 2028

Dec 2028

Mar 2029

TR 38.914 v1.0.0 Jun 2026

PHASE 2 | Rel-21 normative work + product development

PHASE 1 | 5G-Advanced + 6G studies 2025–2027

Mar 2027–Mar 2029

Rel-18/19 deployment and Rel-20 studies CAMARA/Open Gateway scaling and AI-for-RAN PQC pilots and operator architecture preparation

Rel-21 specification work and freezes Chipset, device and infrastructure development

2025

Commercial introduction depends on spectrum, devices ecosystem readiness and business case Later-release AI-native and above-100-GHz research

Pre-commercial trials and WRC-27 follow-up

2027

2028

PHASE 3 | Initial commercial 6G + scaling Around 2030 onward

2029

2030

2032+

ITU-R candidate RIT submissions and evaluation process: 2027-2029

FIGURE 20. 6G standards, product, and commercial roadmap. Fixed 3GPP/ITU standards and spectrum milestones (Rel-20 6G study with TR 38.914 v1.0.0, 2026; WRC-27; Rel-21 Stage 1 freeze Mar. 2027, Stage 2 Jun. 2028, Stage 3 Dec. 2028, ASN.1/OpenAPI Mar. 2029; ITU-R candidate-RIT evaluation 2027–2029) shown separately from the operator’s deployment phases, which are company targets, not standards commitments

23

VOLUME XX, 2026

extensions [19], dApp programmability [20], and scalable RAN [21] approach full potential as research and implementation mature, with third-party sub-10 ms control remaining a development objective rather than a deployed standards commitment; consumer AI-native air-interface services, Stage 3 learned waveforms, the mature CAMARA catalog plus author-proposed 6G extensions such as Sovereign Attestation and AI-RAN Exposure, and above-100-GHz operation follow as later-release research and deployment mature. Progression between deployment phases is gated, not assumed: it requires spectrum and regulatory availability; interoperable chipset and UE evidence within stated power and thermal budgets; measured network-energy and modelassurance performance; and signed enterprise demand with an independently verifiable SLO. Failure of any gate triggers a 5G-Advanced-only commercialization path and defers the 6G-dependent offer. B. Implications for Stakeholders

The operator-controlled 6G model has differentiated implications for nine stakeholder groups. The discussion below frames them as analytical implications of the architecture, economics, and standardization trajectory developed in Sections III-X. For operators, the immediate implication is that AIsubstrate and data-layer ownership become strategic architecture decisions rather than implementation details deferred to vendors. Procurement criteria should be component-specific: CNTi CNF conformance, Kubernetes conformance where relevant, and O-RAN Alliance certification or badging where applicable. Brownfield adoption should be evaluated as a staged option rather than a binary replacement: begin with a bounded, high-value domain and disclose integration, migration, assurance, datagovernance, training, and multi-vendor-accountability costs. For low-ARPU operators, broader cloud-native or agenticautonomy commitments should follow only after that domain has a positive unit-economic case. For vendors, the framework shifts competition toward performance, interoperability, and auditable openness rather than lock-in. O-RAN integration maturity, complete xApp/dApp documentation, NETCONF/YANG contributions, plugfest participation, and third-party AIsubstrate audits become more informative indicators of longterm value than proprietary interface breadth [50], [90], [91]. For hyperscalers, the analysis points to complementarity rather than simple substitution. Licensed spectrum and radioedge presence remain operator advantages, but managed-edge offerings can still displace value unless open serving stacks, CAMARA-aligned consumption models, and revenue-sharing structures keep enterprise data and control within operatorgoverned environments. The dependency also runs inward: the agentic operating model consumes ecosystem-supplied components – agent protocols, frameworks, and frontier 24

models – which the Control Compact deliberately treats as federated or consumed elements under the contractual protections of Section III.C, never as owned substrate. For regulators, the main implication is that Level 4 closedloop management should be assessed as an operational necessity rather than as an exceptional automation case. Proportionate oversight at the policy and exception layer, together with timely spectrum decisions such as upper midband availability at WRC-27, materially affects whether the proposed architecture can be deployed at scale. For standards bodies, the paper implies that service-layer and management specifications must advance in parallel with the radio layer. Priority areas include O-RAN dApp/E3 evolution, Near-RT RIC extensions, and governance of 6G exposure families such as AI inference offload, ISAC sensing, and sovereign attestation before proprietary implementations harden into de facto interfaces [19], [20]. For enterprise and public-sector buyers, the commercial test is independent verifiability: buyers need a named measurement authority, transparent breach attribution, bounded liability, and remedies that survive multi-provider service chains. Without those controls, an outcome guarantee is a marketing label rather than a contractible product. For device, module, and chipset suppliers, the roadmap creates a hard feasibility constraint. AI-native and sensing functions must fit declared power, thermal, memory, update, and certification budgets, and multi-vendor interoperability must be demonstrated before a learned air-interface stage can be treated as deployable. For academia, the remaining high-value contribution is not another abstract 6G vision but methods that close the gap between IMT-2030 targets and commercial-scale engineering. Multi-vendor training compatibility, formal verification of LLM agents in safety-critical loops, and field-grounded trials with ecosystems such as NICT, Hexa-X-II, and SNS JU are especially relevant. For investors, the framework reframes operator AI-platform expenditure as infrastructure for durable differentiation rather than discretionary experimentation. Under this view, sustainable 6G ARPU growth is most plausible where operators retain control of the AI substrate, the API exposure layer, and the data pipeline required for guaranteed-service delivery. XII. OPEN RESEARCH QUESTIONS

Eight open questions will determine the pace of the transition. (1) Brownfield transformation economics: Whether cloudnative virtualized architectures can deliver considerable, fulllifecycle cost and energy savings for brownfield operators transitioning from legacy infrastructure. Japan's NEDO [92] post-5G R&D initiative is directly addressing this question, with Rakuten Mobile and KDDI jointly selected in May 2026 to investigate the simultaneous optimization of virtualized base stations (vRAN) and data center computing, targeting a roughly 40% reduction in network energy consumption by VOLUME XX, 2026

fiscal year 2030. The initiative is well-positioned to produce the credible, independently overseen evidence the industry needs. Rakuten Mobile brings proven cloud-native Open RAN expertise from a greenfield deployment, while KDDI contributes the operational reality of transitioning large-scale legacy infrastructure – together covering both ends of the brownfield challenge. The expectation is that their findings will yield standardized frameworks applicable well beyond Japan. As results emerge, the NEDO initiative stands to move Open RAN transformation economics from an open question to an evidence-based foundation, giving brownfield operators worldwide a credible, replicable roadmap for adopting automated, energy-efficient cloud-native architectures with confidence. (2) Formal verification for LLM agents in safety-critical control loops: no accepted methodology exists to reconcile Level 4 execution with EU AI Act Article 14 conformity assessment at production scale; domain-specific reasoning benchmarks such as 6G-Bench [30] are a necessary first step, but no benchmark yet couples reasoning accuracy to closedloop safety and SLA impact at operator scale; a first evaluation of the Section VII platform – router added-latency and availability budgets, enforcement-loop timing, and agent reasoning quality – can follow the deployment-feasibility methodology of [30] on a testbed before production trials. (3) Guarantee Economy scale conditions: whether jurisdiction-level legal infrastructure for SLA enforcement, standardized SLA term frameworks, and independent performance verification can be established within the Phase 2 window. (4) Multi-vendor training compatibility for Stage 3 air interfaces: jointly learned transmitter-receiver systems require shared training or standardized learned-component representations across competing chipset vendors – no 3GPP precedent exists, and the answer determines whether Stage 3 is a platform technology or a vertically integrated proprietary capability; the UE side additionally lacks committed neuralaccelerator, model-update, and certification roadmaps. (5) Multi-agent governance at network scale: emergent behavior grows non-linearly with coordinating agents (federated RICs, cross-vendor dApps, enterprise-operator agent interaction); formal methods for specifying and verifying coordination protocols under the ZSM 017 [32] threat model and the adversarial taxonomy [36] remain insufficient. (6) Post-quantum migration sequencing: hybrid AKA variants are quantified [33], but no consensus cut-over criteria exist – joint 3GPP SA3/NIST/operator work is required. (7) Adversarial robustness benchmarks: the field lacks a MITRE ATT&CK-equivalent for AI-native RAN - a versioned catalog of production-observed attack techniques with mapped defenses - without which procurement cannot demand evidence-based robustness; relatedly, federated explainability that preserves audit-grade rationale [37] across operator and national boundaries without leaking training data 25

remains unsolved. (8) Service-level energy attribution: MetricEES [55], ITU-T L.1333, ATIS TEER, ISO/IEC 30134-2, and ETSI ES 203 228 cover equipment, network, and facility levels, but no standard ties energy to delivered business outcomes. Without such a standard, sustainability cannot be priced as a first-class commercial dimension and the SNS JU eHealth result [67] cannot be replicated systematically. XIII. CONCLUSION

This article has argued that 6G is unlikely to deliver structural revenue growth unless it is designed from the start as an operator-controlled platform – one in which the operator owns the operating model, the AI substrate, the data layer, and the customer relationship, and uses that ownership to deliver verified outcomes rather than undifferentiated connectivity. The Control Compact gives architectural and procurement decisions a strategic rationale; the six-tier Guarantee Economy translates control into contractible service products; and the agentic operations model, together with the proposed Network MCP Platform, provides a path from manual operations toward Level 4 autonomy under Level 5 governance. The operational substrate is supported by publicly documented national-scale commercial evidence including TM Forum GB1059H-certified Level 4 autonomous RAN energy optimization, a nationwide RIC deployment, and LLM-based infrastructure optimization [5], [9]. By contrast, the Network MCP Platform and the six-tier catalog are proposals grounded in that substrate and in current standards trajectories; this article analyzes their rationale, architecture, and production constraints rather than reporting deployment results for them. The software industry’s platform transformation offers a useful analogue: vendors that opened interfaces and competed on service quality adapted more successfully than those that defended proprietary integration. For telecom, the near-term implication is practical rather than rhetorical. Operator boards can assess current architecture against the Control Compact, instrument one high-value vertical against the outcome-family and Guarantee Economy framework and set a time-bounded Level 4 autonomy target under Level 5 governance. None of these steps depends on waiting for 6G standards, but each generates evidence about whether operator control can be translated into durable commercial advantage.

VOLUME XX, 2026

APPENDIX A. CAMARA / OPEN GATEWAY 6G API CANDIDATE LIST TABLE 8. [CAMARA MATURITY / AUTHOR 6G CANDIDATES] CAMARA / OPEN GATEWAY 6G API CANDIDATES API family (exact CAMARA name where applicable)

Primary use case

Exposed objective TR 22.870 / SLO family

Quality on Demand (QoD)

Guaranteed Guaranteed bitrate; URLLC+; bandwidth/l latency profile; eMBB+ atency for priority class enterprise apps

Mature (CAMARA GA)

Location Verification / Location Retrieval / Geofencing Subscriptions

Asset tracking, geofencing, logistics

Mixed: Location Verificatio n stable; Location Retrieval initial; check each API at submission

Number Verification / SIM Swap

Authenticati Real-time Sovereign on; fraud verification; Trust prevention response SLA to be contractually defined

Verify/retrieve device location; area subscription

Positioning; location

Mature (CAMARA GA)

Massive IoT Mature (CAMARA GA)

Discover Nearest MEC Edge / MEC Mature nearest edge endpoint resolution (CAMARA for app GA) deployment

Model As A Service / Optimal Edge Discovery

Edge AIDeploymentservice specific latency, management throughput, and and edge availability selection objectives

Network Slice Booking / Network Slice Assignment

Enterprise slice reservation and device assignment

AI/ML as a service; AI NaaS

Initial (CAMARA ; sandbox; status checked Jul. 2026)

Reservation and NaaS; assignment determinism requests; performance targets are deploymentspecific

Initial (CAMARA ; sandbox; status checked Jul. 2026)

ISAC Sensing Environmen Sensing objective, exposure tal sensing, authorization, and occupancy, assurance (not velocity internal radio timing)

ISAC

AI-RAN Exposure

Programmab Authorle RAN proposed; not a current CAMARA API

26

Objective, authorization, policy, status, assurance (not internal E3/xApp/dApp timing)

Exposed objective TR 22.870 / SLO family

Maturity / status

Verifiable data residency and jurisdiction

Data-residency Sovereign assertion and Trust jurisdictional evidence; cryptographic proof

Authorproposed; not a current CAMARA API

DISCLOSURE

D. Soldani is Senior Vice President, Advanced Research and Innovation, at Rakuten Mobile Inc., Tokyo, Japan; P. Nahi, A. Muhammad, N. Dwivedi, F. Monaco, and F. Jebamani are with the same group at Rakuten Mobile Inc. The assumptions and views reported herein are solely those of the authors and do not necessarily represent those of Rakuten or its affiliates. Generative-AI use disclosure: the authors used Claude (Anthropic) and Codex (OpenAI) for editing and proofreading assistance, literature-search assistance, figure and table caption drafting, and document review; no section of this article contains AI-generated scientific content, and all content, claims, analysis, and conclusions are the sole responsibility of the named authors. ACKNOWLEDGMENT

Simple Edge Discovery

Third-party real-time RAN control

Primary use case

application service

Maturity / status Sovereign Attestation

Device IoT Status update Reachability reachability latency; Status / Device and roaming reachability Roaming state accuracy Status

Energy Footprint Notification

API family (exact CAMARA name where applicable)

Authorproposed; not a current CAMARA API

Report E2E Declared-boundary Sustainabilit Initial energy energy and carbon y (CAMARA consumption information; no ; sandbox; and carbon universal perstatus footprint for session attribution checked an standard Jul. 2026)

This article draws on the lead author’s over 25 years of work experience and contributions to the information and communication technology sector, together with the published work of the ITU-R, 3GPP, NGMN Alliance, O-RAN Alliance, ETSI, AI-RAN Alliance, IOWN Global Forum, and the Linux Foundation, cited throughout. The authors thank the Rakuten Mobile Inc. engineering and operations teams, whose work operating a large-scale, fully cloud-native Open RAN mobile network provides much of the deployment context discussed in this article. The authors also acknowledge Sharad Sriwastawa of Rakuten Mobile and Rakuten Symphony for the opportunity to work on this subject and for his invaluable support and contribution. REFERENCES [1] [2] [3]

[4]

GSMA, “The Mobile Economy 2021,” GSMA, London, 2021. [Online]. Available: https://www.gsma.com/mobileeconomy/. Accessed: Jul. 2026. ITU-R, “Future Technology Trends of Terrestrial IMT towards 2030 and Beyond,” ITU-R Report M.2516-0, International Telecommunication Union, Geneva, Nov. 2022. ITU-R, “Framework and Overall Objectives of the Future Development of IMT for 2030 and Beyond,” ITU-R Recommendation M.2160-0, International Telecommunication Union, Geneva, Nov. 2023. ITU-R WP 5D, “Draft New Report ITU-R M.[IMT-2030.TECH PERF REQ] – Technical Performance Requirements for IMT-2030 Radio Interface(s),” Working Party 5D, International Telecommunication Union, Geneva, Feb. 2026; draft completed by WP 5D and submitted to Study Group 5 for approval (expected Dec. 2026). VOLUME XX, 2026

[5]

Rakuten Group, Inc., “FY2025 Full Year Financial Results Presentation,” Rakuten Group, Inc., Tokyo, Feb. 2026. [Online]. Available: https://global.rakuten.com/corp/news/press/2026/0212_01.html. Accessed: Jun. 2026. [6] NGMN Alliance, “AI Surge and Its Implications for 6G,” v1.0, NGMN, Frankfurt, Feb. 2026. [7] TM Forum, “Autonomous Networks Levels Evaluation Methodology,” IG1252 v3.0.0, TM Forum, May 2025. [8] 3GPP, “Study on 6G Use Cases and Service Requirements – Stage 1,” 3GPP TR 22.870 v2.0.0, 3rd Generation Partnership Project, 2026. [9] H. M. Yohannes, G. Dukundane, M. Nazeeruddin, D. Chaubey, and S. Behera, “A Three-Stage Architecture-Aware Framework for LLM-Enhanced Kubernetes Resource Optimization,” in Proc. Int. Conf. Computing and Automation Engineering (ICCAE), Sydney, Australia, 2026, pp. 1–6. [10] D. Brown, “Rakuten Mobile Adopts eBPF to Strengthen Anomaly Detection and Security in Cloud-Native Telecom Networks,” eBPF Foundation, Aug. 2025. [Online]. Available: https://ebpf.foundation/rakuten-mobile-adopts-ebpf-to-strengthenanomaly-detection-and-security-in-cloud-native-telecom-networks/. Accessed: Jun. 2026. [11] D. Soldani et al., “eBPF: A New Approach to Cloud-Native Observability, Networking and Security for Current (5G) and Future Mobile Networks (6G and Beyond),” IEEE Access, vol. 11, pp. 57174–57202, Jun. 2023, doi: 10.1109/ACCESS.2023.3281480. [12] AI-RAN Alliance, “AI-RAN Architecture,” Version 1.2, AI-RAN Alliance, 2024. [13] AI-RAN Alliance, “State-of-the-Art AI/ML Applications for Improvements in RAN Performance,” WG1 AI-for-RAN Technical Report, AI-RAN Alliance, 2024. [14] AI-RAN Alliance, “AI on RAN: Enabling Monetizable Differentiated Connectivity for AI,” Working Group 3 White Paper, AI-RAN Alliance, 2024. [15] NTT DOCOMO, “5G Evolution and 6G White Paper,” Version 5.0, NTT DOCOMO, Tokyo, Jan. 2023. [16] C. Chatzieleftheriou and E. Liotou, “A Survey on AI for 6G: Challenges and Opportunities,” IEEE Open J. Commun. Soc., vol. 7, pp. 3189–3220, 2026. doi: 10.1109/OJCOMS.2026.3677293 [17] X. Zheng et al., “AI-Native 6G Physical Layer with Cross-Module Optimization and Cooperative Control Agents,” arXiv:2601.02827v2 [eess.SP], Jan. 2026. [18] S. Salmi et al., “AI-Native O-RAN Architectures for 6G: Towards Real-Time Adaptation, Conflict Resolution, and Efficient Resource Management,” TechRxiv (preprint), Sep. 2025. doi: 10.36227/techrxiv.175825547.74922399/v1 [19] O-RAN Alliance nGRG, “Evolution of Near-Real-Time RAN Intelligent Controller Towards 6G,” Technical Report nGRG-RR2025-04, O-RAN Alliance, Jan. 2026. [20] O-RAN Alliance nGRG, “dApps: Distributed Applications for Open RAN,” Technical Report (contributed research report) nGRG-RR2025-05 v2.0, O-RAN Alliance, Jan. 2026. [21] O-RAN Alliance nGRG, “Scalable and User-Centric RAN Architecture: Service Requirements and Design Considerations,” Technical Report nGRG-RS02, O-RAN Alliance, Jan. 2026. [22] H. Harkous et al., “Flat UP: Toward RAN-Core Convergence for the 6G User Plane,” IEEE Commun. Mag., vol. 63, pp. 62–68, May 2025. doi: 10.1109/MCOM.003.2400403 [23] ETSI, “Zero-touch Network and Service Management (ZSM); ZSM Framework for Network-as-a-Service (NaaS),” ETSI GR ZSM 019 v1.1.1, ETSI, Sophia Antipolis, Jan. 2026. [24] ETSI, “Zero-touch Network and Service Management (ZSM); Study on the Utilization of Agents in Autonomous Networks,” ETSI GR ZSM 020 v1.1.1, ETSI, Sophia Antipolis, Jan. 2026. [25] ETSI, “Zero-touch Network and Service Management (ZSM); Reference Architecture,” ETSI GS ZSM 002 v1.1.1, ETSI, Sophia Antipolis, Aug. 2019. [26] NGMN Alliance, “Agentic AI-Based Operating Models,” V1.0, NGMN, Frankfurt, Mar. 2026. [27] IEEE Communications Magazine Editorial Board, “Advanced AI for Zero-Touch Service Management Networks (Special Issue),” IEEE

Commun. Mag., vol. 64, no. 1, Jan. 2026. [Online]. Available: https://ieeexplore.ieee.org/xpl/tocresult.jsp?isnumber=10812619. Accessed: Jun. 2026. [28] H. Sun et al., “Advancing 6G: Survey for Explainable AI on Communications and Network Slicing,” IEEE Open J. Commun. Soc., vol. 6, pp. 2196–2223, 2025. doi: 10.1109/OJCOMS.2025.3534626 [29] C. Provvedi, L. Seidenari, B. Picano, and R. Fantacci, “Intent-LLM: A Framework for Automated Network Configuration Through Code Generation,” IEEE Trans. Cogn. Commun. Netw., vol. 12, pp. 7246– 7257, 2026. doi: 10.1109/TCCN.2026.3683230 [30] M. A. Ferrag, A. Lakas, and M. Debbah, “6G Needs Agents: Toward Agentic AI-Native Networks for Autonomous Intelligence,” IEEE Open J. Commun. Soc., vol. 7, pp. 7254–7282, 2026, doi: 10.1109/OJCOMS.2026.3707904. [31] M. Souppaya, K. Scarfone, and D. Yaga, “Zero Trust Architecture,” NIST Special Publication 800-207, National Institute of Standards and Technology, Aug. 2020. [32] ETSI, “Zero-touch Network and Service Management (ZSM); Closed-Loop Automation Security Aspects,” ETSI GR ZSM 017 v1.1.1, ETSI, Sophia Antipolis, Jan. 2026. [33] E. Choudhary, A. K. Yadav, R. Kumar, and M. Liyanage, “A Practical Approach to Transitioning 5G-AKA Toward Fully and Hybrid Post-Quantum Security for 5G and Beyond Communication,” IEEE Open J. Commun. Soc., vol. 7, pp. 2980–2995, 2026. doi: 10.1109/OJCOMS.2026.3677962 [34] NIST, “Module-Lattice-Based Key-Encapsulation Mechanism Standard (ML-KEM),” FIPS 203, National Institute of Standards and Technology, Aug. 2024. doi: 10.6028/NIST.FIPS.203 [35] NIST, “Module-Lattice-Based Digital Signature Standard (MLDSA),” FIPS 204, National Institute of Standards and Technology, Aug. 2024. doi: 10.6028/NIST.FIPS.204 [36] M. Altintaş et al., “When Beneficial Intelligence Turns Hostile: A Survey of Adversarial Threats in AI-Native 6G,” IEEE Open J. Commun. Soc., vol. 7, pp. 3468–3512, 2026. doi: 10.1109/OJCOMS.2026.3678511 [37] S. N. Karahan, M. Güllü, S. Çimen, A. Yazar, and M. S. Osmanca, “AI-Driven Network Slicing for F5G-A Networks: A Framework With Unsupervised Discovery, Temporal Dynamics, and Explainable AI,” IEEE Open J. Commun. Soc., vol. 7, pp. 1–15, 2026, Art. no. 3679199. doi: 10.1109/OJCOMS.2026.3679199 [38] 3GPP, “Study on 6G Scenarios and Requirements,” 3GPP TR 38.914 v1.0.0, 3rd Generation Partnership Project, Jun. 2026 (approved at RAN #112). [39] Nokia, “6G Baseline Coverage Improvements over 5G in Existing Spectrum,” 3GPP Workshop on 6G, Doc. 6GWS-250004, 2025. [40] Qualcomm, “Qualcomm Vision and Priorities for 6G RAN,” 3GPP Workshop on 6G, Doc. 6GWS-250068, 2025. [41] Japan Ministry of Internal Affairs and Communications (MIC), “Radio Wave Use Homepage – Frequency Allocation,” [Online]. Available: https://www.tele.soumu.go.jp/resource/e/search/myuse/use0303/batc h.pdf. Accessed: Jun. 2026. [42] O-RAN Alliance nGRG, “Generative AI in Network,” Technical Report nGRG-RR-2025-02, O-RAN Alliance, 2025. [43] NGMN Alliance, “Framework for Network Simplification – An Operator View,” v1.0, NGMN, Frankfurt, Feb. 2026. [44] D. Soldani and A. Manzalini, “Horizon 2020 and Beyond: On the 5G Operating System for a True Digital Society,” IEEE Veh. Technol. Mag., vol. 10, no. 1, pp. 26–31, Mar. 2015. doi: 10.1109/MVT.2014.2380581 [45] GSMA Open Gateway. [Online]. Available: https://www.gsma.com/solutions-and-impact/gsma-open-gateway/. Accessed: Jun. 2026. [46] Cloud Native Telco Initiative (CNTi), The Linux Foundation Project. [Online]. Available: https://lfnetworking.org/cloud-native-telcoinitiative-cnti-driving-cloud-native-adoption-in-telecom/. Accessed: Jun. 2026. [47] Nephio, “Cloud Native Network Automation,” The Linux Foundation Project. [Online]. Available: https://nephio.org/. Accessed: Jun. 2026.

27

VOLUME XX, 2026

[48] Sylva, The Linux Foundation Project. [Online]. Available: https://sylvaproject.org/. Accessed: Jun. 2026. [49] Cilium, “eBPF-based Networking, Observability, Security,” The Linux Foundation Project. [Online]. Available: https://cilium.io/. Accessed: Jun. 2026. [50] O-RAN SC, The Linux Foundation Project. [Online]. Available: https://o-ran-sc.org/. Accessed: Jun. 2026. [51] kagent, “Bringing Agentic AI to Cloud Native,” The Linux Foundation Project. [Online]. Available: https://kagent.dev/. Accessed: Jun. 2026. [52] O-RAN Alliance. [Online]. Available: https://www.o-ran.org/. Accessed: Jun. 2026. [53] CAMARA, The Linux Foundation. [Online]. Available: https://camaraproject.org/. Accessed: Jun. 2026. [54] Ericsson, “Network for AI Experiences,” Ericsson AB, Stockholm, 2025. [55] IOWN Global Forum, “Toward a Unified Metric for Energy Efficiency and Sustainability in Next-Generation Networks,” IOWN GF MetricEES v1.2, IOWN Global Forum, Jan. 2026 (reviews fragmented energy-efficiency metrics from ATIS/ITU-T/ETSI and proposes their harmonization). [56] Government of India, Department of Telecommunications, “Bharat 6G Vision,” DoT, New Delhi, Mar. 2023. [57] Regulation (EU) 2024/1689, “Artificial Intelligence Act,” Official Journal of the European Union, 2024. [Online]. Available: https://eur-lex.europa.eu/eli/reg/2024/1689/oj/eng. Accessed: Jun. 2026. [58] Regulation (EU) 2016/679, “General Data Protection Regulation (GDPR),” Official Journal of the European Union, 2016. [Online]. Available: https://gdpr-info.eu/. Accessed: Jun. 2026. [59] Regulation (EU) 2025/327, “European Health Data Space (EHDS),” Official Journal of the European Union, 2025. [Online]. Available: https://health.ec.europa.eu/ehealth-digital-health-and-care/europeanhealth-data-space-regulation-ehds_en. Accessed: Jun. 2026. [60] Directive (EU) 2022/2555, “Directive on Measures for a High Common Level of Cybersecurity Across the Union (NIS 2 Directive),” Official Journal of the European Union, 2022. [61] Regulation (EU) 2024/2847, “Cyber Resilience Act,” Official Journal of the European Union, 2024. [62] United States, “Secure and Trusted Communications Networks Act,” Public Law 116-124, 2020. [63] United States, “CHIPS and Science Act,” Public Law 117-167, 2022. [64] Amazon.com, Inc., “Annual Report on Form 10-K for the Fiscal Year Ended December 31, 2024,” U.S. Securities and Exchange Commission, Washington, DC, Feb. 2025. [65] TM Forum, "Service Quality Management API," TMF657 v4.0, TM Forum. [Online]. Available: https://www.tmforum.org/open-digitalarchitecture/open-apis/service-quality-management-managementapi-TMF657/v4.0. Accessed: Jul. 2026. [66] TM Forum, “Service Ordering Management API,” TMF641, TM Forum. [Online]. Available: https://www.tmforum.org/resources/standard/tmf641-serviceordering-management-api/. Accessed: Jun. 2026. [67] SNS JU Technology Board, “Towards 6G-Enabled eHealth,” Smart Networks and Services Joint Undertaking, Brussels, Feb. 2026. [68] A. Dawid and Y. LeCun, “Introduction to Latent Variable EnergyBased Models: A Path Towards Autonomous Machine Intelligence,” arXiv:2306.02572 [cs.LG], Jun. 2023. [69] M. Assran, Q. Duval, I. Misra, P. Bojanowski, P. Vincent, M. Rabbat, Y. LeCun, and N. Ballas, “Self-Supervised Learning from Images with a Joint-Embedding Predictive Architecture,” arXiv:2301.08243 [cs.CV], Jan. 2023. [70] CNCF, Cloud Native Computing Foundation. [Online]. Available: https://www.cncf.io/. Accessed: Jun. 2026. [71] SNS JU Smart Networks Working Group, “The AI/ML Landscape for Smart Networks and Services: Taxonomy, Standards and Innovation Pathways,” SNS JU Reliable Software Networks Working Group, Dec. 2025. [Online]. Available: https://smartnetworks.eu/wp-content/uploads/2025/12/SNS-JU-AI-MLLandscape.pdf. Accessed: Jun. 2026.

28

[72] Cloud Native Telco Day EU 2026. [Online]. Available: https://colocatedeventseu2026.sched.com/overview/area/Cloud+Nati ve+Telco+Day. Accessed: Jun. 2026. [73] Anthropic (now governed under the Linux Foundation Agentic AI Foundation), “Model Context Protocol (MCP) Specification,” Revision 2025-11-25, 2025. [Online]. Available: https://modelcontextprotocol.io/. Accessed: Jun. 2026. [74] 3GPP, “Security Architecture and Procedures for 5G System,” 3GPP TS 33.501, Release 18, Jun. 2024. [75] 3GPP, “Telecommunication Management; Charging Management; 5G System; Services,” 3GPP TS 32.290, Release 18, Jun. 2024. [76] 3GPP, “Lawful Interception (LI) Architecture and Functions,” 3GPP TS 33.127, Release 18, Sep. 2024; requirements in 3GPP TS 33.126. [77] Cloud Native Computing Foundation, “SPIFFE and SPIRE: Secure Production Identity Framework for Everyone,” The Linux Foundation. [Online]. Available: https://spiffe.io/. Accessed: Jun. 2026. [78] M. Jones, A. Parecki, and B. Campbell, “OAuth 2.0 Protected Resource Metadata,” IETF RFC 9728, 2025. [79] NIST, “Stateless Hash-Based Digital Signature Standard (SLHDSA),” FIPS 205, National Institute of Standards and Technology, Aug. 2024. doi: 10.6028/NIST.FIPS.205 [80] 3GPP, “System Architecture for the 5G System (5GS),” 3GPP TS 23.501, Release 18, Mar. 2024. [81] 3GPP, “Procedures for the 5G System (5GS),” 3GPP TS 23.502, Release 18, Mar. 2024. [82] Google, “Agent-to-Agent (A2A) Protocol Specification,” 2025. [Online]. Available: https://a2a-protocol.org/. Accessed: Jun. 2026. [83] OCUDU Ecosystem Foundation, The Linux Foundation. [Online]. Available: https://ocudu.org/. Accessed: Jun. 2026. [84] 3GPP SA2, “Study on 6G System Architecture – Intent Handling (Key Issue #18),” 3GPP SA2 UID 910038, 6G architecture study, 2026. [85] 3GPP, “Extended Reality (XR) in 5G,” 3GPP TR 26.928 v16.1.0, 3rd Generation Partnership Project, Dec. 2020. [86] AI-RAN Alliance, “AI for RAN – Use Cases and Capabilities,” WG1 Use Case Compendium, AI-RAN Alliance, 2025. [87] European Commission, “EU Toolbox for 5G Security,” NIS Cooperation Group, Brussels, Jan. 2020. [88] European Union, “Corporate Sustainability Reporting Directive (CSRD), Directive (EU) 2022/2464,” Official Journal of the European Union, 2022. [89] 5G Americas, “5G Advanced Overview,” White Paper, July 2025. [Online]. Available: https://www.5gamericas.org/wpcontent/uploads/2025/07/5G-Advanced-Overview.pdf. Accessed: Jun. 2026. [90] OpenAirInterface. [Online]. Available: https://openairinterface.org/. Accessed: Jun. 2026. [91] Open Network Automation Platform (ONAP). [Online]. Available: https://onap.org/. Accessed: Jun. 2026. [92] New Energy and Industrial Technology Development Organization (NEDO). [Online.] Available: https://www.nedo.go.jp/english/index.html

VOLUME XX, 2026

DAVID SOLDANI (Senior Member, IEEE) received the M.Sc. degree (magna cum laude) in engineering from the University of Florence, Italy, in 1994, and the D.Sc. degree in technology (Hons.) from the Helsinki University of Technology, Finland, in 2006. Throughout his career, he has held prestigious academic positions, including Visiting Professor at the University of Surrey, U.K. (2014), Industry Professor at the University of Technology Sydney (UTS), Australia (2016), and Adjunct Professor at the University of New South Wales (UNSW) (2018). In his professional roles, he has served as Chief Information and Security Officer (CISO), SVP Innovation and Advanced Research at Rakuten; Chief Technology and Cyber Security Officer with Huawei Asia Pacific; Head of 5G Technology at Nokia; and Head of the Central Research Institute and VP Strategic Research and Innovation in Europe at Huawei European Research Center. He is currently SVP Advanced Research and Innovation with Rakuten Mobile Inc., Tokyo, Japan. David can be reached online at: LinkedIn profile.

PETRIT NAHI is Chief AI Consultant in Rakuten Mobile’s AI and Data Division and Senior Member of Technical Staff in the CTO Office. He built and led Rakuten Mobile’s original Data Science team, which evolved into the AI and Data Division, establishing the company’s data and AI foundations for largescale network automation and autonomy. He has played a key role in applying AI/ML to live Open RAN network automation, supporting predictive and reactive models across approximately 150,000 cells and Rakuten Mobile’s progress toward Level 4 energy autonomy. His doctoral research at Queen Mary University of London focused on agent-based control of mobile network coverage, an early form of AI-driven network control. He has held technical, product, executive, and board-level roles across Rakuten Mobile, NETSCOUT, Tektronix Communications, Newfield Wireless, and AIRCOM, and formerly served as Chairman of the Board of Kosovo Telecom.

systems. His research interests include artificial intelligence, machine learning, foundation models, AI for telecommunications, cybersecurity, and open-source AI technologies. FRANCESCO MONACO received the Master of Science (M.Sc.) degree in Computer Engineering from Politecnico di Torino, Italy, in 2022, where he has also collaborated as a Research Assistant. He joined Rakuten Mobile Inc., Japan, in 2024, where he currently serves as an Architect/Tech Lead and Operations Manager in the Next Generation Advanced Research Department. He has been a key contributor to the conception, development, and large-scale deployment of the Sauron eBPF Platform across the Rakuten Mobile network. His interests include high-performance networking, programmable data plane, cloud-native technologies, and production-scale network observability systems. FRANCIS JEBAMANI is the Head of AI Products & Engineering at Rakuten Mobile Inc., with over two decades of experience in the telecommunications industry. His career spans Radio Access Networks (RAN), Multi-access Edge Computing (MEC), Core Networks, and AI-driven products, with a focus on delivering intelligent solutions that enhance network performance, energy efficiency, operational efficiency, and end-user experience. Before joining Rakuten Mobile, he spent nearly five years at Reliance Jio, where he contributed to the architecture, engineering, and scaling of the company’s nationwide 4G Core network from the ground up. He currently leads the development of AI-powered products that accelerate intelligent automation and evolution toward autonomous networks. He also played a key role in achieving TM Forum Autonomous Networks Level 4 for Energy Efficiency. His professional interests include AI for telecommunications, network intelligence, autonomous network operations, and energy-efficient AI solutions.

AWN MUHAMMAD is Head of Technical Standards at Rakuten Mobile, with over 18 years of experience in mobile networks, including RAN planning, feature development, deployment, and standardization. He represents Rakuten Mobile as Alternate Director on the O-RAN ALLIANCE Board and actively contributes to O-RAN ALLIANCE and 3GPP RAN activities, with a focus on Open RAN, energy efficiency, 5G Advanced, and 6G evolution.

Nikhil Dwivedi received the B.Tech. degree in Electrical Engineering from Delhi Technological University (DTU), Delhi, India, and the M.Tech. degree in Artificial Intelligence from the Indian Institute of Technology Jodhpur (IIT Jodhpur), India. He is currently a Research Engineer in the Next Generation Advanced Research Department at Rakuten Mobile Inc., Japan. His work focuses on developing artificial intelligence and deep learning solutions for next-generation telecommunications and cybersecurity, with an emphasis on foundation models, trustworthy AI, and translating research into production-scale 29

VOLUME XX, 2026