Huawei Sizes the Agent Problem at 900 Billion Devices and Rebuilds the Core Plane for It

Huawei is telling carriers that the mobile core network has to stop thinking about sessions and start thinking about tasks, and it has put a number on how large the change will be. The vendor cites its own Intelligent World 2035 report for a forecast of 900 billion AI agents worldwide by 2035, roughly 100 times the world’s population, and argues that the present core architecture cannot carry that load.

The argument rests on a specific set of service cases. Agent Over New Radio covers portable AI devices that are limited on compute, battery, size and cost, where the network has to absorb end-to-end latency to keep the experience usable. Wide-Area Computing Network connects computing resources through the mobile network so tasks can be scheduled toward whichever site can run them, and Huawei says the point is standardized computing services, and it says the alternative is bespoke arrangements per case. Agent Computing Network handles the case where a task spans several specialized agents across domains and ecosystems, which the company says no single agent can complete on its own.

That third case is where the architecture claim sits. Huawei proposes a Data & AI Plane to run alongside the control and user planes, described as the intelligence layer that lets the core understand intent, make decisions and confirm completion. Interfaces shift from signaling toward intent, so terminals state what the task needs and the network works out how to deliver it.

The company’s framing of the shift is from moving bits to moving tokens. That phrase is doing a lot of work, and the evidence it offers is a scenario list, and it has published no traffic data. What is checkable is the product direction: control plane, user plane and now an added plane, which is a change in what operators have to buy and how they slice it.

Others in the industry are pulling toward the same plane. The AI-RAN Alliance moved to implementation blueprints and edge business models this week, and RCR Wireless reported the alliance’s work on shared reference architectures. GSMA Intelligence, working from its own modelling, forecasts 6G reaching roughly 2.4 billion connections by 2035 with 5G still at about 60 percent of global mobile connections at the end of that year, and its head of data modelling said capacity will create the initial need but operator economics will set the pace. GSA’s 5G private networks data shows 2003 deployments across 88 countries, mostly in industrial and enterprise settings, which is the traffic base the agentic scenarios are meant to land on.

Huawei has not published the traffic figures behind its own forecast, and it describes the AI-native core as a direction, with no shipping product named.

Supplier’s note: Core network upgrades of this kind reach us as a spares and board question, because the racks stay and the boards move. We hold carrier ethernet switch line cards, service provider edge router boards, rectifier power modules and cooling fan trays in stock in Shenzhen and Hong Kong, with ready-to-ship units leaving inside two working days and full pallet orders taking one to two weeks where a specific board revision has to be pulled from allocation. Where an OEM board revision is on allocation we quote cross-reference equivalents matched to the same form factor and list the firmware revision for every unit before dispatch. Our installation teams report that the recurring field problems are dust-clogged fan trays in humid sites, power modules drifting under unstable grid input and mismatched cards after a slot change, so every order ships with a spare fan tray set, a matched power module and a labelled card map. Telecom operators, systems integrators doing network integration work and enterprise network teams are the usual buyers, and orders arrive as mixed pallets sized for 10 to 40 racks, priced per unit once freight and duty are quoted.