AI-Led 2G Shutdowns: Lessons from Hong Kong

China Mobile Hong Kong has approval to switch off 2G on 23 June 2026—the last of Hong Kong’s four major operators to do it. What grabs me isn’t the date. It’s the tiny number behind it: less than 2.3% of the operator’s mobile base still uses 2G.

That “last 2.3%” is where network modernization gets messy. It’s also where AI in telecommunications stops being a buzzword and starts being the difference between a clean retirement of legacy gear and a slow-motion customer and operational headache.

Hong Kong’s 2G sunset is a compact case study: a dense, tech-forward market with high 4G/5G adoption, yet still a meaningful tail of old SIMs, old handsets, and low-bandwidth devices. If you run network operations, plan spectrum, manage IoT fleets, or own customer experience, the lesson is simple: turning off legacy networks is a data problem before it’s an RF problem.

What Hong Kong’s 2G phaseout actually signals

A 2G shutdown isn’t just “we’ll use the spectrum for 5G.” It’s a structural shift in how a telco operates.

In the source news: China Mobile HK is the territory’s largest operator with 8.5 million mobile connections (end-September), and it already turned off 3G earlier in 2026 after usage fell to 0.25%. Competitors moved earlier on 2G: 3 Hong Kong (Sept 2021), SmarTone (Oct 2022), and HKT approved in Nov 2024.

So why does this matter beyond Hong Kong?

The “long tail” isn’t small operationally

Even when it’s only a few percent of subscribers, the tail is rarely uniform. It’s usually a mix of:

• People with old handsets or inactive second SIMs
• SMEs running legacy POS terminals or basic telemetry
• Roamers and niche devices that behave differently than smartphone traffic
• Security- and compliance-sensitive use cases (alarms, elevators, call boxes)

That mix is exactly why operators that treat shutdowns as a schedule-and-comms exercise get burned.

Sunsetting 2G is now a modernization playbook, not a one-off

By late 2025, the industry has done enough sunsets (2G/3G in many markets) that executives expect these programs to be repeatable. That creates pressure for automation: fewer bespoke war rooms, more continuous monitoring, and clearer accountability.

That’s where AI-driven network optimization fits naturally.

The hard part of shutting down 2G: finding who still depends on it

If you want a smooth 2G sunset, you need to identify dependency precisely: which subscribers, which devices, which locations, and which time windows still use 2G.

The challenge is that legacy dependency hides in ordinary-looking data:

• 2G attach events that spike only at certain hours
• 2G fallback driven by coverage holes, not preference
• Old SIM profiles that never upgraded, but still occasionally authenticate
• M2M devices that only “phone home” once a day

Where AI helps first: segmentation that’s actually actionable

Most companies get this wrong by segmenting customers only by plan type or handset model. For a sunset, you need segmentation by behavior.

A practical AI segmentation approach combines:

• Network telemetry (attach/handovers, location area updates, paging success)
• Device identity signals (TAC ranges, radio capability, firmware fingerprints)
• SIM lifecycle data (age, provisioning type, last swap, roaming status)
• Customer/account context (consumer vs SMB vs enterprise fleet)

A good model doesn’t just label “2G user.” It ranks:

1. Risk of service impact if 2G goes away
2. Likelihood to migrate with light-touch nudges
3. Most effective intervention (SIM swap, handset upgrade, IoT module replacement, coverage remediation)

Snippet-worthy rule: A 2G shutdown succeeds when you treat it like a churn-prevention program backed by network science.

People Also Ask: “Why not keep 2G for emergencies?”

Operators and regulators sometimes keep 2G around for coverage and voice reliability. But keeping it alive has real costs: spectrum inefficiency, vendor lock-in, energy use, site complexity, and operational overhead.

The better trade is typically: retire 2G, strengthen low-band 4G/5G, and use targeted solutions for edge cases (including dedicated IoT migration plans and better indoor coverage).

AI-driven network optimization during the transition window

Once an operator announces the date, the network enters a transition period. Usage patterns change because:

• Some customers migrate early
• Others wait until the last moment
• Certain devices start failing and generate support calls

AI earns its keep here by turning transition into a managed runway.

Predictive operations: forecasting where the problems will show up

During sunsets, the operational question is: Where will we see “silent failures” first?

AI models can forecast risk by cell and by neighborhood, using signals such as:

• Rising 2G reselect attempts in specific clusters
• Increased call setup failures on 2G in areas with aging equipment
• Abnormal roaming attach behavior (tourist-heavy zones)
• Atypical IoT traffic bursts tied to specific enterprise accounts

This supports a shift from reactive firefighting to proactive fixes:

• Deploy targeted in-building solutions for known trouble spots
• Adjust neighbor lists, reselection parameters, or coverage layers
• Prioritize customer outreach by geography and account value

Closed-loop assurance: “trust, but verify” your migration

An underappreciated risk is thinking your migration is complete because your dashboard says 2G traffic is down.

Closed-loop assurance uses AI to continuously test assumptions:

• Did 2G users truly migrate, or did they churn?
• Did their experience improve on 4G/5G, or did they start falling back to weak coverage layers?
• Did enterprise IoT devices re-register successfully after SIM changes?

The reality? A clean sunset isn’t “traffic hits zero.” It’s complaints, failures, and abnormal events hitting normal levels.

Customer migration: AI can reduce cost-to-serve (and protect revenue)

China Mobile HK has started notifying affected customers to upgrade SIMs and, where applicable, handsets. That’s the visible part of the work. The expensive part is handling the exceptions.

Use AI to prioritize outreach that customers will actually respond to

A blunt SMS blast or generic retail push wastes money and annoys people who already migrated.

Instead, AI can drive a tiered plan:

1. Low risk / high likelihood to self-fix: simple digital comms, self-serve SIM swap appointment
2. Medium risk: proactive call center outreach with scripts informed by device/account data
3. High risk / high value (SMB/enterprise IoT): account-manager-led migration plan with device inventory and test windows

If you’ve ever sat in an ops review, you know the metric that matters: contacts avoided. AI helps reduce inbound calls by making outreach relevant and timed.

People Also Ask: “What about IoT devices stuck on 2G?”

This is the part many operators underestimate. IoT migration needs its own track:

• Identify fleets via IMEI ranges and traffic patterns
• Confirm radio capability (2G-only vs multi-mode)
• Plan module replacement cycles and certification constraints
• Offer migration pathways (LTE-M, NB-IoT, or regular LTE/5G) based on power, coverage, and throughput

AI supports this by detecting likely IoT endpoints, clustering them by behavior, and flagging “must-touch” accounts months earlier.

A practical AI playbook for legacy shutdowns (2G, then the next one)

If you’re planning a 2G shutdown—or watching your regulator push you toward one—here’s a playbook I’d actually use.

1) Build a “legacy dependency graph,” not a static list

A spreadsheet of remaining 2G subscribers becomes stale fast. A dependency graph updates continuously and ties:

• Subscriber/account → SIM → device → cell sites → usage windows

AI helps by inferring relationships (especially for IoT) and flagging anomalies.

2) Define success as fewer incidents, not just less traffic

Set operational KPIs that reflect reality:

• 2G attach events trending to zero by region
• Call setup success and data session success on 4G/5G for migrated users
• Complaint rate and truck rolls per 1,000 migrations
• Enterprise IoT re-registration success within defined windows

3) Run “migration A/B tests” in controlled geographies

Before the full shutdown, simulate it:

• Select a small area
• Encourage migration
• Observe: failures, fallbacks, coverage gaps, support load

AI models get better when they see real transition behavior, not just historical steady-state data.

4) Use predictive maintenance to reduce noise during the sunset

During a sunset you don’t want extra outages from aging kit. Use predictive maintenance models to:

• Identify baseband/RF components trending toward failure
• Schedule preventative swaps ahead of the high-touch migration period
• Reduce false alarms and prioritize work orders

This connects directly to the broader AI in Telecommunications theme: modernization is smoother when the network is stable and observable.

What to take from China Mobile HK’s timing

Being last to turn off 2G in a market isn’t automatically good or bad. But it does raise a strategic point: if your competitors already retired 2G, you’re likely supporting a higher share of the market’s remaining legacy devices—and that can distort your operational priorities.

The smarter stance is to treat 2G retirement as a catalyst:

• simplify the network
• re-farm spectrum with confidence
• reduce energy and maintenance load
• and re-invest in better 4G/5G coverage layers where it counts

If you do it with AI-driven network optimization, you also end up with something more valuable than a shutdown: a reusable modernization engine you can apply to the next transition, whether that’s 3G refarming, IoT fleet upgrades, or advanced 5G standalone optimization.

The next 18 months (from now through June 2026) will be the interesting part: not the switch-off moment, but the quality of the migration runway leading up to it.

If you’re planning your own legacy sunset, what’s your biggest unknown right now—device discovery, IoT migration, coverage gaps, or customer outreach capacity?