The Operator's Vision: China Unicom's Case for Building AI-Centric Networks

China Unicom framed 5G Capital's six-year arc as a shift from proving 5G could work commercially to proving networks can now operate as infrastructure for AI itself. The networks thus far have been built for people and smartphones and optimized for coverage, capacity, and downlink experience. However, this network build is no longer sufficient once agents, robots, low-altitude devices, and industrial systems turn on at scale. Agents need to upload massive amount of input tokens, sensing data and more. Live, immersive experiences will warrant stable uplink. Robots & Autonomous systems will demand low, deterministic latency. This is distant from existing networks which are designed around best-effort downlink service.

China Unicom's Beijing operation for example now defines “high uplink” as a minimum cell uplink bandwidth of 100 MHz and a minimum user uplink speed of 20 Mbps, achieved by refarming older spectrum for 5G-Advanced use. The telco has already lit up more than 10,000 sites with that capability. Its network operations have shifted from manual, periodic tuning to AI-driven management at a scale human teams could not match. The telco shared that tens of thousands of base stations are re-optimized every week instead of every few months, with half of all network workflows now AI-driven. China Unicom is also shifting its billing model from charging by GB toward a model built around AI agents and tokens, backed by a rapidly growing share of capital investment now going into compute rather than pure connectivity. This can be a benchmark for several telcos around the world as Tokenomics take over Trafficonomics.

Another interesting takeaway was to make a deliberate move away from prioritizing premium users toward what was described as “network equity” essentially guaranteeing every user baseline communication, access, and payment capability first, and only then layering differentiated experience on top. Even if that means throttling premium users' heaviest traffic when the network is under strain. It is a departure from how we have thought about, designed and marketed mobile networks for two decades, to make it more scalable for not only people but the digital agents at the same time. This will require a completely differential thinking and implementing when millions of agents begin to talk to each other, make nanosecond decisions at scale. This highly deterministic AI architecture will require equally capable network architecture with completely redefined and dynamic KPIs for agents and humans coexisting in the same network.

Some of the key thinking points and questions which will define the networks evolution and how it would look like in five years from now are:

• How should network capabilities evolve for services that need uplink and downlink, high bandwidth and low latency, and mass-market scale and guaranteed performance simultaneously, rather than optimizing for any single metric?
• How should the industry jointly define what a network needs to deliver, given that requirements now depend as much on device design, model architecture, and business model as on radio engineering?
• How does AI, in turn, transform network operations themselves from operations to healing?

The Vendor's Playbook: Huawei's Engineering Case for the Agent Economy

Huawei's brought a grand vision with their contribution centred around innovating in an era of “traffic-token synergy,” where the real unit of value on a network is shifting from the gigabyte to the token.

 The Token Value Greater than Traffic GB Value

According to Huawei, a gigabyte of data might sell for roughly 2 yuan, but that same gigabyte, expressed as the AI tokens it can carry, represents something closer to 300–400 million tokens, and generating that many tokens consume enough electricity to imply an underlying value on the order of 170 yuan. That gap between what connectivity currently charges and what the traffic riding on it is worth was presented as both a warning and an opportunity. For example, in the mobile-internet era, connectivity providers captured a tenth of total value created, behind platforms and device makers and without a new business model, token-era economics could compress that share further rather than expand it.

To capture more of that value, three concrete technology bets are to be considered for the next phase of network engineering design:

• Pooling capacity across multiple frequency bands and treating them as one flexible resource, rather than optimizing each band in isolation.
• Commercializing “high uplink” at scale this year through more efficient use of existing spectrum, rather than waiting for new spectrum allocations.
• Coordinating scheduling across cells and bands so a whole network behaves like a single cell and thus closing the experience gap at cell edges and indoors that has always been the weakest link in mobile performance.

Further building on China Unicom’s equity principle that protecting a premium user's experience at the edge of a cell, without degrading everyone else's, can consume eight times the network resource of doing the same for a typical user. And this under realistic spectrum configurations, a single cell can only support a handful of such premium users before ordinary users' experience visibly suffers. That capacity ceiling, more than any policy preference, is why equity by design becomes an engineering necessity rather than a marketing choice. In future, a persistent coverage gap in higher, wider-bandwidth spectrum relative to today's mid-bands would need newer innovations such as a new antenna technology as the path to closing it on the way to 6G. The focus on new standards for latency and experience quality in “AI mode” will be critical instead of retrofitting with standards and metrics built for voice and video.

Where the Rest of the Industry Is Heading

Beyond the two hosts, the wider room full of leading device makers, chipmakers, standards bodies, satellite operators and market observers, converged on a strikingly consistent set of trends.

On devices, there was broad agreement that on-device AI agents are moving from novelty to mass-market infrastructure in the space of a single year, with leading smartphone makers citing user bases in the hundreds of millions and multi-fold annual growth. These AI Agents are augmenting human capability and judgment, not substituting it. The roadmaps are tightly locked between chip makers and device makers to foster today’s standalone on-device assistants toward always-on, cross-device companions capable of proactive planning over the next several years.

On network design, chipmakers and standards organizations postulated the same shift: away from measuring networks by peak speed and raw latency and toward measuring whether an AI agent's task completes reliably. This will help gravitate industry to pay via performance or outcome-based pricing rather than one-size-fits-all connectivity.

In addition to the “equity-over-priority” principle, industry should also design network to edge with robust security, guardrails to ensure reliability and avoid instability.

Counterpoint's Perspective: From AI Driven Networks, Devices to AI Cities - the Next Frontier

Our own contribution to the roundtable made the case that this same transition from connectivity, to intelligence, to autonomy out identically one layer up, in the city itself, just on a longer timeline.

Rather than a single jump from “connected” to “autonomous,” we track the transition through five concrete stages:

• Connected City (2021–2025) — ubiquitous broadband, city-scale IoT sensors, smart-city pilots and the first network-slicing foundations. This era is complete.
• Data-Driven City (2022–2024) — analytics dashboards layered onto traffic, energy, and public-safety systems.
• Generative AI City (2024–2026) — where we place the industry today, with LLM-powered citizen services and “GovGPT” assistants enabled by 5G-Advanced edge inference.
• Agentic AI City (2027–2029) — multi-agent systems autonomously running traffic, utilities, and emergency response, with humans setting the goals.
• 6G-Enabled Autonomous City (2030 and beyond) — built on physical AI and a digital-twin operating system for the whole city.

Cities are the bedrock of creating economic value.AI Smart Cities built on a foundational telco and cloud network can take the entire AI ecosystem to the next level across multiple use-cases. For example, the traffic congestion alone costs cities an estimated 3–5% of GDP every year, and AI-managed intersections in leading cities have already cut congestion by 30%. Our sharpest proof point for what a mature deployment looks like is Abu Dhabi, effectively the world's first AI-native government: more than 120 live AI use cases across ten-plus government sectors, a citizen-facing AI assistant app reaching over 900,000 residents, and AED 27 million in verified annual cost savings with more than 60 of those use cases deployed in a single year. The blueprint behind it is replicable: build a sovereign AI cloud, then a unified platform, then onboard government entities onto it, then push services out to citizens through one app.

We see cities converging on three proven governance models rather than one universal path:

• Sovereign AI City — top-down and government-led, with the state building the cloud and mandating adoption. Abu Dhabi, Singapore, and Riyadh anchor this model.
• Telco-Led AI City — operator-driven, with the telco rolling out infrastructure and network-as-a-service APIs city-wide. Shenzhen, Seoul, and Tokyo lead here.
• Ecosystem-First City — a coalition of government, telco, technology platform, and academia building an open platform together. Amsterdam, Barcelona, and Dubai exemplify this approach.

In our 2025 AI City rankings, Singapore took the top spot with a score of 84, Seoul followed close behind at 82, and China placed more cities in the global top fifteen than any other country. Our conclusion for the room was the same one we'd make to any city or operator reading this: the window to build this capability is narrow, roughly 2025 to 2027, and telcos in particular are being asked to evolve from connectivity providers into AI orchestration partners — the operators who move first on network APIs, edge AI and sovereign cloud will define what an intelligent city looks like for the next decade. Everyone else will be renting that future from them.

Key Takeaways

• Uplink, not downlink, is now the industry's shared design priority agents, robots and live AI experiences all need the network to send data, not just receive it.
• Networks are becoming a compute and intelligence layer, not just a pipe and increasingly need to be run by AI, not just carry it.
• Tokens are emerging as a parallel currency to gigabytes, forcing a rethink of how connectivity captures value in an AI-driven traffic mix.
• “Equity” is replacing “priority” as the operating philosophy to guarantee the basics for everyone before layering on premium experience.
• Cities are tracing the same connected-to-autonomous curve as networks, one layer up, through three distinct governance models to become future AI or Autonomous city rather than via a single playbook.
• The success of the key stakeholders will be stemming from whoever builds the AI orchestration layer on top of connectivity, not simply building the fastest pipe.

5G and AI, the two key fundamental technologies might have arrived at the same moment by coincidence, not design but this collision is a significant opportunity for every stakeholder. However, this is reshaping design, architecture, and business models at every layer from networks, devices, to cities.

Further related research:

MWC 2026: How 5G-A and AI Build the Bridge to 6G in Mobile AI Era

Counterpoint Conversations: Huawei's 5G-Advanced Powering the Mobile AI Revolution

GenAI Smartphones’ Terrific Traffic: Are Telecom Networks Prepared?

AI 360 Pulse - Industry Trends, Intelligence and Impact, May 2026

GenAI Smartphones’ Terrific Traffic: Are Telecom Networks Prepared?

Future-Proofing Mobile Networks for AI Era with 5G Advanced

If you are interested in subscribing to the full report or would like to learn more, please contact us at [email protected].