Seventeen major operators submitted a joint position paper ahead of the recent Madrid 3GPP meeting, urging the standards body to keep 6G core design evolutionary rather than force a costly rip-and-replace. Their intervention exposes a deepening rift between Western and Chinese vendors over how AI and signaling should be embedded in the next-generation core network.

What are the three proposed architectures?

At the SA2 working group meeting in Prague, delegates examined three competing candidate directions for handling AI functionality inside the 6G core. All three build on the existing 5G service-based architecture (SBA) but diverge sharply on non-access stratum (NAS) signaling — the secure pipeline between device and core that has anchored mobile networks since 2G (GSM).

• Direction 1 – Separate AI domain. A dedicated AI domain for intent handling sits independent of the packet-switched network, with the AMF routing NAS signaling while AI fulfillment runs over the service-based interface (SBI). Backers include Nokia, T-Mobile USA, Verizon, NVIDIA, Deutsche Telekom, Apple, Qualcomm, Ericsson and NEC. Implementation requires a UE AI domain client integrated into devices.
• Direction 2 – AI in 6G network functions. AI integrates through dedicated 6G network functions following standardized 3GPP procedures, with flexible NAS routing either standalone or combined with core functions such as the AMF. NTT DOCOMO, Samsung, LG Electronics, NEC, IIT Bombay, Vodafone, Apple and China Telecom (partly) support this approach.
• Direction 3 – AI/agent-handled connectivity. AI agents integrate deeply into network procedures with dynamic coordination of capabilities and tool invocation, using a signaling routing function (SRF) independent of the AMF. Huawei, HiSilicon, China Mobile, ZTE, vivo, CATT, ETRI, Ewha Womans University and OPPO (partly) align here. It introduces the most procedural disruption.

Why are operators pushing back?

Operators are still absorbing capital costs from 5G standalone deployments that have yet to deliver proportional commercial returns. Their Madrid paper stressed upgradable paths, investment protection and a migration they can afford.

Reworking NAS would require modem redesign, core security overhaul and device backward-compatibility planning. Most operators consider such a step avoidable, since AI could in principle sit on top of existing 5G core interfaces.

The Western camp — T-Mobile USA, Verizon, Deutsche Telekom, Nokia, Ericsson, NEC, NVIDIA and, to some extent, Apple — has converged on Direction 1. Anchoring the core to the 5G SBA preserves 5G core investments and lets intent handling mature alongside existing services.

Where could a compromise emerge?

Delegates at Prague discussed a partial merge of Directions 1 and 2. Operators would keep a separate AI domain while introducing modular NAS routing to prevent potential chokepoints in the AMF. The tweak would let NAS routing operate standalone or combined with core functions like the AMF.

Direction 3 remains the clean-slate alternative. It proposes native agentic routing and an independent signaling function, replacing fixed signaling with dynamic intent coordination. Backers argue it sets the stage for autonomous AI-driven operations, but it deviates from the lean, evolutionary 5G SA migration principle that 3GPP adopted as the baseline for 6G.

What are the commercial stakes?

Even if 3GPP settles on Direction 1, architectural disruption shifts out to the terminal. Direction 1 does not only adjust how a modem connects to the RAN. It standardizes UE AI domain client interfaces to the core by creating an application-layer intent client and a 3GPP AI domain client.

Three placement options sit on the table:

• Baseband firmware: tightest 3GPP integration, but tied to silicon cycles at Qualcomm, MediaTek and Samsung.
• Operating system layer (Android/iOS): faster software updates, but requires new APIs to translate application intent to baseband hardware.
• App layer: quickest to deploy, but lacks the low-level integration needed for deterministic performance.

If specifications lean toward an operator-controlled framework, device-side AI agents, hyperscalers and over-the-top ecosystems risk lock-out. An OTT AI agent running cloud gaming or enterprise workloads might need an additional operator translation function to adapt network parameters on the fly.

Can 3GPP keep a single global standard?

Past generations occasionally resolved deadlocks by leaving contentious variations as optional vendor implementations outside 3GPP specifications. For 6G, the architectural divergence is too deep for that workaround.

The gap between evolutionary and revolutionary camps risks a splintered global standard. Operators could end up running two divergent cores: a revolutionary single-vendor solution in Chinese domestic markets and an evolutionary multi-vendor architecture elsewhere.

Post-Madrid discussions must finalize an architecture by November. If a merger of Directions 1 and 2 emerges, the industry can keep a unified 6G standard; if not, the next G risks shipping in two incompatible flavors.