AI in 6G: Two Competing Approaches Are Emerging for the 6G Core Network

At The 3rd Generation Partnership Project (3GPP) Technical Specification Group Service and System Aspects (TSG SA) Working Group #2 Meeting #175 that took place in May 2026, the industry advanced two competing approaches for Artificial Intelligence (AI) integration under Key Issue (KI) #18 of the 6G core architecture study.

The first approach, Solution Variant #18.1, proposes embedding new agent-based Network Functions (NFs) within the core. These functions would receive intent from users or applications and orchestrate existing NFs as callable tools, positioning AI as a native control layer, and enabling autonomous decision-making and real-time service orchestration. This solution was submitted by Chinese operators such as China Mobile, China Telecom, China Unicom, Huawei, ZTE, CATT, Xiaomi, OPPO, vivo, HONOR, Futurewei, and Korean and Indian operators including ETRI, LG Uplus, SK Telecom, and IIT Bombay. This approach effectively requires new core network upgrades and perhaps even rearchitecting it, in a “clean-slate” approach that is pursued by vendors and operators from the East.

The second solution, Solution Variant #18.3, suggests creating a separate AI domain that connects to the core using a dedicated Translator Function (TF). The main goal is to keep the existing core NFs “deterministic and AI-free.” In this setup, AI acts as an optimization and orchestration layer, which helps avoid large changes to the core and keeps critical NFs predictable. This proposal came from Nokia, Ericsson, AT&T, T-Mobile USA, Qualcomm, Deutsche Telekom, Apple, Verizon, SoftBank, Google, NVIDIA, and SK Telecom. Notably, SK Telecom contributed to both approaches, underscoring its preference to maintain flexibility. This Western-dominated group aims to upgrade the 5G Next Generation Core gradually, without major investment or potential service disruptions.

Both variants were approved for parallel study, reinforcing that 3GPP is preserving optionality, rather than converging prematurely as 6G requirements and deployment scenarios continue to evolve. Alongside this architectural split, the industry reached consensus on a mandatory operator control mechanism, requiring that all AI-driven actions pass through an operator-governed control layer, including the ability to override or fully disable AI behavior.

ABI Research views this as the first indication that AI integration in the 6G core is bifurcating into two distinct architectural paths, rather than converging on a single model. This split will influence how much AI-driven capability operators can build directly into the core, shaping their competitiveness within the broader AI infrastructure value chain.

Variant #18.1 is a bigger bet on where telco networks are headed.  It would change the core from a connectivity layer into a programmable orchestration platform that can coordinate network, compute, and AI resources for distributed inference and AI workloads. However, deploying this in practice will require a fundamental shift in telco operations, given that networks have historically been built around deterministic behavior. Variant #18.3 takes a more incremental approach. By separating AI capabilities from core NFs, operators can introduce AI gradually, while maintaining the predictability of previous generations. The trade-off is that AI risks becoming another optimization layer, rather than a core capability of the network, potentially limiting how much value operators can capture from AI-driven services.

The choice will also shape the vendor landscape. An AI-native core favors vendors that can bring together networking, compute, AI software, and orchestration capabilities across the stack. This aligns more closely with Chinese vendors’ current strategies, but adoption will ultimately depend on whether AI-native networks deliver clear improvements in operational efficiency and create new revenue opportunities. A more evolutionary approach will likely remain attractive for operators prioritizing interoperability and multi-vendor ecosystems, but it may slow their transition toward a future where AI is deeply integrated into the network.

Operators will naturally avoid committing prematurely to either Variant #18.1 or Variant #18.3 while both remain under active 3GPP study, especially because there is no market consensus on which option to follow. The advancement of both approaches, alongside SK Telecom’s decision to co-sponsor both variants, demonstrates that even leading industry players have not yet determined a single path for AI integration in the 6G core. Therefore, operators should prioritize vendors that can support both architectural outcomes, rather than those requiring early commitment to a specific AI placement. However, operators may have limited flexibility in practice as Western operators do not have access to Chinese vendors and vice-versa. A 3GPP decision favoring a single approach may disrupt the plans of the opposing vendor ecosystem and require operators to align with an architecture they did not initially prioritize. Most likely though, both options will be fully studied and specified, and it will be left to the individual operator and vendor to implement.

This scenario resembles the transition from 5G with Non-Standalone (NSA) to Standalone (SA), where Western operators regretted their initial choice to launch 5G New Radio (NR) immediately and postpone the core network upgrades for later. Even today, 7 years after 3GPP Release 15, SA upgrades are ongoing for many national networks that chose to launch 5G NSA, with many interoperability and integration issues. A similar situation could emerge for 6G core AI features, with operators that invest proactively in AI-native architectures such as Variant #18.1 able to gain a competitive advantage through greater network automation, intelligence, and the ability to support new AI-driven services, while late adopters may face more costly and complex transitions.

Operators should prioritize operator control, which will likely remain important regardless of which architecture advances and require vendors to show strong AI governance, such as human override options and audit trails. From there, operators should evaluate vendors on efficiency gains, service orchestration improvements, and new enterprise revenue.

Finally, operators and vendors should treat agent interoperability as an immediate concern. Unlike KI#18, there is no 3GPP deadline forcing the resolution of agent-communication standards such as Anthropic’s MCP, Google’s A2A, or Cisco-led AGNTCY. Operators should make vendor-neutral agent interoperability a procurement requirement, rather than waiting for the core debate to resolve first.