ZTE and Ucell Cut Network Operating Costs by 10.6%: How an AI Agent Optimises Mobile Network Energy Consumption

Base stations in a mobile network run around the clock, and the electricity bill arrives around the clock too, regardless of whether there is any traffic on the network or not. Ucell, Uzbekistan's state mobile operator, partnered with ZTE to deploy an AI agent for dynamic equipment power management. The result: total network energy consumption dropped by 10.6%, and for a national operator of that scale the annual savings run into millions of dollars, with no tariff increases and no degradation in call quality.

Electricity is the single largest line item in any mobile operator's operating expenditure. Base station equipment consumes nearly as much power at three in the morning, when almost no one is on the network, as it does during peak hours. Multiply that gap by thousands of stations and 365 days, and you have money literally disappearing into thin air. This is a solved problem, and the Ucell case shows exactly how to solve it.

Where the operator's money was leaking

A mobile network is not a single server in a data centre. It is thousands of base stations spread across an entire country. Each station consists of several components: power amplifiers, antenna units, signal processing modules. All of them consume electricity continuously, at full capacity, even when traffic at a given station is close to zero: late at night, in sparsely populated areas, during the hours when subscribers are asleep.

For Ucell, as for any national operator, electricity forms the bulk of operating costs. Engineers knew that real network load is uneven: there are peak hours and there are hours of deep idle. But the traditional logic of network management is simple: keep everything on at full power so you never miss a traffic spike. It is reliable, but expensive. Paying for peak readiness during hours when there is no peak, that is exactly the cost line that needed to change.

The problem was compounded by the fact that manual power management of stations at national-network scale is physically impossible. You cannot sit an engineer at a console to manually switch thousands of stations between power modes based on live traffic. What was needed was a system that does this on its own, accurately and in real time.

Why traditional approaches did not solve the problem

Fixed energy-saving schedules had existed in telecoms for years. The classic approach: define static time windows, say 2:00 to 5:00 AM, during which part of the equipment switches to a low-power mode. This delivered modest savings but created a new problem: the schedule had no awareness of real traffic. On a holiday night or during a major public event, load could be high precisely during the designated quiet window, and a station put to sleep by schedule would drop service quality.

The other option, threshold rules: if load falls below X%, move the component into economy mode. But a threshold is a blunt instrument. It does not account for the fact that traffic might spike sharply seconds later, and it has no way to wake equipment up in advance based on predicted subscriber behaviour.

Both schemes were reactive and static. The network needed a tool that sees demand patterns ahead of time, makes decisions about each component's operating mode in real time, and guarantees that equipment wakes up before traffic actually rises. That is exactly where the AI agent came in.

How the AI agent was designed

The agent was built around a single central task: match the current and predicted load on each base station to its energy state and decide in real time what mode each component should operate in. To do this, the agent continuously collected traffic data across the entire network, analysed historical consumption patterns, and factored in time of day, day of week, seasonality, and the local characteristics of each individual station.

The key function was predictive wake-up: the agent does not simply put components to sleep when load drops, it calculates when load is about to rise and proactively brings equipment out of economy mode in advance. The response time is milliseconds, enough that subscribers notice no difference in service quality.

The agent managed not entire stations but individual components: specific antenna units, amplifiers, processing modules. This enabled granular savings: some components sleep while others serve real traffic. The principle is to scale consumption strictly to operational intensity, rather than keeping everything at maximum just in case.

A quality-protection layer was built in from the start: if the agent detected that moving a component into economy mode posed a risk to network performance metrics (latency, packet loss, coverage), it left that component running. Savings were never allowed to come at the expense of subscriber experience.

How the rollout worked at Ucell

Deployment happened in stages. In the first phase, the agent ran in observation mode: it collected traffic and energy data, built load-pattern models for each coverage zone, but did not yet control any equipment. This allowed the team to accumulate statistics and validate forecast quality before giving the agent real authority.

In the second phase, the agent was connected to live management on a pilot group of stations where risk was lowest: low-traffic areas with predictable overnight patterns. Ucell and ZTE engineers monitored network quality metrics in real time, comparing them against a control group of stations running without the agent. Once they confirmed that service quality was not degrading and consumption was falling, they rolled the system out across the full network.

For the engineering team this did not mean retraining from scratch. The agent operates autonomously and requires no constant manual intervention. Engineers moved from manually configuring power modes to monitoring the agent and handling exceptions, situations that fall outside standard scenarios.

Results

A 10.6% reduction in energy consumption for a national operator is not an abstract percentage. Electricity makes up a significant share of the operating budget, so a double-digit saving here translates directly into millions of dollars that can be reinvested in network development rather than utility bills. Meanwhile, Ucell subscribers noticed no change in service: the agent knows how to save invisibly.

There is an additional effect that is harder to quantify but no less important: reducing the load on equipment during idle hours extends its service life. Components that spend less time running at full capacity last longer, which lowers the total cost of ownership of the infrastructure across its entire lifecycle.

How to apply this to your infrastructure

The Ucell and ZTE case is not limited to telecoms. Any infrastructure that runs around the clock under variable load, from data centres to production lines and warehouse complexes, contains the same opportunity for savings. Here is where to start:

• Measure your real load profile. Look at how your equipment's actual consumption changes across the day and week. The gap between peak and minimum is the savings potential you are currently paying for unnecessarily.
• Identify components that can be managed independently. The more granular the control over individual nodes, the more precisely the agent can optimise consumption without risking overall system performance.
• Start with an observation phase. Run the agent in data-collection and forecasting mode before giving it real control. This reduces risk and gives engineers confidence in the quality of the agent's decisions.
• Build quality protection in as the top priority. The agent must know which performance metrics cannot be compromised under any circumstances. Savings that come at the cost of service quality end up costing more than no savings at all.

If this case sounds like what's happening in your company, our manager can help: he'll analyze your business and niche for free and point out where an AI agent would bring a real result in your case. Message the manager