How ZTE Plans to Rebuild Mobile Networks for an Agent-Driven World

The Agent Economy Demands New Infrastructure

Cui Li, Chief Development Officer at ZTE, stood before the GTI Summit 2026 in late June and framed a problem that most telecom executives have yet to articulate clearly: mobile networks were built to connect people and devices, not to serve billions of autonomous AI agents exchanging instructions, tokens, and real-time inference results across cellular links.

According to ZTE, the rise of agent-driven connections - software entities that negotiate, plan, and act on behalf of users - will require networks to deliver not just bandwidth but "reachable intelligence, guaranteed experiences, and monetizable services." That shift, the company argues, demands a fundamental rethink of uplink capacity, latency floors, experience determinism, and coverage topologies. It is a bet that the next generation of wireless infrastructure will be defined less by peak throughput and more by the reliability and predictability required to run distributed AI workloads at the edge.

At DailyTechWire, we have tracked similar arguments emerging from Seoul to Shenzhen: that 6G is less a speed upgrade and more an architectural pivot toward AI-native networks. ZTE's public framework - a "2+4" structure combining two strategic priorities and four evolution directions - offers one of the clearer roadmaps yet for how a Tier 1 vendor intends to translate that vision into silicon, protocols, and field trials.

The "2+4" Framework: Strategic Priorities and Evolution Directions

ZTE's approach anchors on two strategic priorities. The first is 6G and AI convergence, treating machine learning not as an add-on but as a first-class citizen in the radio access network (RAN) and core. The second is Space-Air-Ground Integrated Network (SAGIN), a term that encompasses satellite direct-to-consumer links, low-altitude unmanned aerial vehicle (UAV) networks, and emergency communication overlays - essentially, the recognition that terrestrial cells alone will not cover the surfaces and altitudes where agents and sensors will operate.

Layered on top are four evolution directions. Ultimate spectral efficiency aims to squeeze more bits per hertz out of finite spectrum allocations, a perennial telecom goal now sharpened by the need to serve uplink-heavy AI workloads. AI massive connections targets the billions of low-power, intermittent endpoints - think wearable AI glasses, industrial sensors, and embodied robots - that will generate short, bursty traffic. Integrated sensing-communication-computing-intelligence bundles radio sensing (for positioning, environment mapping) with communication and on-device inference, a convergence that Chinese vendors have championed more aggressively than their Western counterparts. Finally, trustworthiness and security acknowledges that agent-to-agent transactions, especially in finance, healthcare, and industrial control, will demand cryptographic guarantees and tamper-evident audit trails at the network layer.

This taxonomy is not unique to ZTE - variants appear in whitepapers from Huawei, Samsung, and NTT DOCOMO - but the emphasis on agents as the primary use case is notable. Most 6G roadmaps still lead with holographic telepresence or extended reality; ZTE is betting that the killer app will be invisible software entities negotiating rides, scheduling manufacturing runs, and coordinating supply chains over cellular links.

Four Network Capabilities for the Agent Era

To support agent services, ZTE has identified four capability upgrades. Enhanced uplink addresses the asymmetry in today's networks, which prioritize downlink for video streaming. Agents uploading sensor data, intermediate model weights, or live video for remote inference will reverse that traffic pattern. Ultra-low latency targets single-digit millisecond round-trip times, a requirement for closed-loop industrial control and real-time multi-agent coordination. Deterministic experience shifts the focus from average throughput to tail latency and jitter, ensuring that the 99th percentile user - or agent - receives predictable service. Ubiquitous coverage extends the network envelope to rural areas, airspace, and maritime zones where traditional macro cells are uneconomical.

These are engineering problems with known solutions - beamforming, edge caching, time-sensitive networking - but the challenge lies in delivering all four simultaneously at scale and acceptable cost. The temptation in telecom has always been to optimize for one metric (peak speed, coverage, latency) at the expense of others; agent workloads will punish that trade-off.

AI-RAN Architecture: Decoupling "AI for RAN" and "RAN for AI"

One of the more concrete disclosures from the summit was ZTE's AI-RAN design philosophy, which decouples two functions that vendors often conflate. "AI for RAN" refers to using machine learning to optimize radio resource allocation, beam steering, and handover decisions - essentially, making the network itself smarter. "RAN for AI" means provisioning the RAN to carry AI inference traffic efficiently, handling token streams, model checkpoints, and gradient updates as first-class data types.

ZTE's architecture allows these two functions to evolve independently. For "AI for RAN," the company mandates AI application-specific integrated circuits (ASICs) to ensure power efficiency and deterministic inference times in base stations. For "RAN for AI," it offers optional xPUs - a catch-all term for GPUs, FPGAs, or other accelerators - that can be deployed where economics justify the flexibility. The rationale is that not every cell site will need to run large language models locally, but every site will benefit from AI-driven scheduling. By making ASICs mandatory and xPUs optional, ZTE aims to balance performance, energy consumption, and total cost of ownership across diverse deployment scenarios.

This design choice also reflects a broader tension in the AI infrastructure market: purpose-built versus general-purpose silicon. ZTE is signaling that for the base layer of network intelligence, custom chips win on efficiency; for edge inference workloads, operators should retain the flexibility to choose accelerators based on their service mix.

Space-Terrestrial Integration and Field Trials

The SAGIN priority translates into three near-term testing scenarios. Direct-to-consumer satellite connectivity, the capability that has captured public attention through Starlink and Apple's Emergency SOS, is one strand; ZTE is exploring how to hand off sessions between terrestrial cells and low-Earth-orbit (LEO) satellites without dropping agent transactions. Low-altitude UAV networking targets drones used in logistics, agriculture, and inspection, where the network must track fast-moving endpoints across cell boundaries. Emergency communication layers in backup satellite or high-altitude platform station (HAPS) links when terrestrial infrastructure fails - a use case that gained urgency after recent natural disasters in Asia-Pacific.

The technical crux is seamless handover and service orchestration. An AI agent managing a delivery fleet needs to maintain a single session as drones move from urban macro cells to rural satellite coverage; any interruption breaks the control loop. ZTE has not disclosed the protocol extensions required, but industry work in 3GPP Release 19 and beyond suggests new signaling for non-terrestrial network (NTN) integration and enhanced mobility management.

Field trials are underway, though the company offered no specifics on geographies or partners. The fact that these tests are happening in 2026 - well ahead of the 2030 target for 6G commercial launch - indicates that vendors are using late-stage 5G-Advanced (5G-A) as a proving ground for 6G concepts, a pattern we have observed across the region.

Scenario-Based Validation: Embodied AI, Glasses, and Industrial Agents

ZTE highlighted four application domains for pilot projects. Embodied AI refers to robots with physical actuators - warehouse bots, humanoid assistants, surgical systems - that require ultra-reliable low-latency communication (URLLC) and precise positioning. AI glasses, the wearable form factor that Meta, Apple, and Chinese startups are racing to ship, demand always-on uplink for environmental context and downlink for real-time overlays, all within a power envelope measured in milliwatts. In-vehicle AI encompasses both autonomous driving (where the car is the agent) and in-cabin assistants (where the agent mediates between passengers and services). Industrial agents coordinate machine tools, logistics robots, and quality-control systems in factories, a domain where deterministic latency and security are non-negotiable.

These scenarios share a common thread: they assume intelligence is distributed across cloud, edge, and endpoint, with the network as the connective tissue that must carry not just data but also timing guarantees, security tokens, and quality-of-service commitments. Validating that end-to-end is harder than lab-testing a new modulation scheme; it requires real environments, real agents, and real failure modes.

ZTE's emphasis on "replicable and scalable industry value" suggests the company is wary of one-off demos that generate headlines but do not translate into deployments. The telecom industry has a long history of showcasing futuristic use cases - remote surgery, tactile internet - that never achieve commercial traction because the economics or ecosystem dependencies do not align. By anchoring pilots in embodied AI and industrial agents, ZTE is targeting markets where early adopters (logistics operators, manufacturers) have both the budget and the operational pain points to justify premium connectivity.

The GTI Collaboration and Ecosystem Strategy

ZTE framed its 6G work within the context of the Global TD-LTE Initiative (GTI), a carrier-led organization that historically championed Time Division Duplex (TDD) spectrum and later became a vehicle for 5G-A trials and OpenLab validation. The company emphasized its role in urban pilot projects and ecosystem co-building, language that signals a strategy of de-risking 6G investment through pre-commercial collaboration with operators.

Four collaboration directions were outlined. New network capability research targets agent communication protocols, token flow scheduling, and cross-domain experience assurance - essentially, the software stack above the physical layer. Cost-efficient AI-RAN architecture was covered above. TN-NTN integration testing focuses on the space-terrestrial handover scenarios. Scenario-based application validation brings in vertical industry partners to test embodied AI, glasses, vehicles, and industrial agents in live networks.

This structure mirrors the approach Chinese vendors used to accelerate 5G: define requirements with anchor operators (China Mobile, China Telecom, China Unicom), build prototypes in controlled testbeds, then scale through GTI and 3GPP standardization. The difference in 2026 is that the geopolitical landscape has fragmented the ecosystem; ZTE and Huawei face export controls and market access barriers in North America, Europe, and parts of Asia-Pacific. GTI, as a result, has become a more explicitly China-centric forum, with membership weighted toward operators in Belt and Road markets.

For ZTE, that means its 6G roadmap will likely gain traction first in Southeast Asia, the Middle East, Africa, and Latin America - regions where Chinese vendors already hold dominant RAN share and where operators are more willing to adopt SAGIN and AI-RAN features ahead of Western incumbents. Whether that early deployment translates into global standardization influence remains an open question; 3GPP is still the arena where interoperability is decided, and the organization's consensus-driven process can blunt any single vendor's advantage.

Risks and Open Questions

ZTE's framework assumes that agent-driven traffic will materialize at scale within the 6G timeframe (2030-2040). That is not a given. While AI agents are proliferating in software-as-a-service platforms and enterprise workflows, their migration to mobile networks depends on business models, regulatory clarity, and user acceptance. If agents remain largely cloud-bound - orchestrated within data centers rather than distributed across endpoints - the case for enhanced uplink and ubiquitous coverage weakens.

The cost-efficiency claims for AI-RAN also warrant scrutiny. Deploying ASICs and optional xPUs at every cell site adds capital expenditure and operational complexity. Operators are already under margin pressure; unless AI-RAN delivers measurable revenue uplift - through premium service tiers, new enterprise contracts, or cost savings in network operations - adoption will lag. ZTE has not published total cost of ownership models or return-on-investment case studies, leaving the business case largely theoretical.

Finally, the SAGIN vision requires regulatory coordination across spectrum (terrestrial versus satellite), orbital slots, and cross-border handover agreements. The international regulatory environment for NTN is still evolving, and conflicts over spectrum allocation between terrestrial mobile and satellite operators have been fierce. ZTE's ability to deliver seamless space-terrestrial integration depends on policy outcomes it cannot control.

What It Signals for 6G's Direction

Despite those uncertainties, ZTE's public commitment to a 2+4 framework and its early field trials offer a useful data point on where at least one major vendor believes 6G value will concentrate. The shift from connectivity to "reachable intelligence" is more than marketing; it reflects a genuine architectural question about whether networks should remain agnostic bit-pipes or evolve into platforms that understand and optimize for AI workloads.

Other vendors will make different bets. Ericsson and Nokia have emphasized network energy efficiency and open RAN; Samsung is pushing terahertz spectrum and reconfigurable intelligent surfaces; NTT DOCOMO has championed AI-RAN but with less emphasis on SAGIN. The diversity of approaches is healthy - 6G is still early enough that no consensus architecture has emerged - but it also means interoperability and standardization will be battlegrounds.

For operators and enterprises watching from the sidelines, the takeaway is that 6G planning is already underway, and the decisions being made in 2026 and 2027 - on spectrum, on RAN architecture, on satellite partnerships - will shape what is possible in the 2030s. Agent-driven services may or may not become the dominant use case, but the infrastructure to support them is being designed now. Whether ZTE's vision proves prescient or premature will depend on how quickly AI agents move from cloud orchestration to the mobile edge - and whether the economics of serving them at scale ever pencil out.