Turning AI Billions into Better Telco Networks

Nokia didn’t spend 2025 “talking about AI.” It spent the year reorganizing around it, hiring leadership to push it, and publicly tying future competitiveness to big AI investment. That’s not a branding exercise—it’s a signal that telecom is treating AI as core infrastructure.

For operators, vendors, and the procurement teams that fund them, this matters for one reason: AI is starting to show up as a line item with measurable operational outcomes—fewer truck rolls, faster root-cause analysis, tighter energy control, and better customer experience. The podcast recap of 2025 (Nokia shake-up, satellite getting real, device differentiation, and “AI billions”) is a useful lens for what’s actually changing inside telco organizations.

This post is part of our AI in Supply Chain & Procurement series, so I’ll take a stance: the winners in telecom AI won’t be the companies with the flashiest demos; they’ll be the ones who can procure, govern, and operationalize AI at scale—across networks, IT, and field ops—without losing control of cost, risk, or performance.

2025 proved AI budgets are now network budgets

AI spending in telecom shifted in 2025 from “innovation funding” to capex and opex logic. That’s why Nokia’s AI push matters: when a major network supplier treats AI as a strategic pillar, it forces every operator to ask the same uncomfortable question—what parts of my network operations still run on manual workflows and static rules?

Here’s the practical translation of “AI billions” into telco reality:

• Network AI (operations): anomaly detection, predictive maintenance, automated ticket triage, and closed-loop optimization.
• Planning AI (build): better demand forecasting for capacity, smarter RAN parameter planning, and site rollout prioritization.
• Service AI (revenue and CX): churn risk, proactive care, and personalized offers.

If you’re in procurement, the real shift is that you’re no longer buying “a tool.” You’re buying capability plus data access plus operational change—often across multiple vendors.

The myth that held teams back

Most companies get this wrong: they treat AI as a software procurement. In telco, AI is closer to a supply chain program—data supply, model supply, compute supply, and governance supply.

If your AI vendor can’t tell you:

• what data it needs,
• where it will run (cloud, edge, on-prem),
• how it integrates into OSS/BSS and ticketing,
• how accuracy will be monitored month-to-month,

…you’re not buying AI. You’re buying a science project.

Nokia’s AI push is a case study in vendor-driven operational AI

A “Nokia shake-up” paired with a high-profile AI investment narrative tells you something specific about the vendor landscape: network vendors are repositioning as AI operations partners, not just equipment providers.

That creates opportunities and risks for operators.

Opportunity: faster time to value in network optimization

Operators are under pressure to do more with the same (or smaller) teams. AI is increasingly used to compress the time between “issue appears” and “issue resolved.” In practice, that means:

• Earlier detection: spotting performance degradation before customers complain.
• Better prioritization: ranking incidents by customer impact and revenue risk.
• Guided remediation: recommending actions with confidence scoring.

When this works, the business result is simple: lower mean time to repair (MTTR) and fewer repeat incidents.

Risk: opaque bundles and vendor lock-in

When vendors bundle AI into network deals, procurement can lose visibility into:

• what part is license vs. services,
• what’s reusable across domains (RAN, transport, core),
• what data rights you’re giving away,
• what switching costs look like in year 3.

My view: telcos should treat AI components like a portfolio, not a single purchase.

A practical way to structure it:

1. Data foundation (internal): data models, event pipelines, and access controls you own.
2. Domain AI apps (mixed): vendor solutions where they’re strong (RAN optimization), internal builds where you need differentiation (customer care triage).
3. MLOps and governance (shared): monitoring, drift detection, auditability, and change control.

That portfolio approach reduces the chance you end up paying “AI tax” on every contract renewal.

Satellite “getting real” changes telco planning and procurement

The podcast’s point that satellite is “getting real” is more than a coverage story. It’s a planning and sourcing story.

Direct-to-device, LEO backhaul, and hybrid connectivity push telcos toward multi-network operations, which increases complexity in:

• performance management,
• service assurance,
• customer experience consistency,
• partner SLAs.

AI becomes the coordination layer—because humans can’t manually manage high-volume events across terrestrial + satellite domains without missing things.

What supply chain & procurement teams should do now

If your network roadmap includes satellite partnerships in 2026, procurement needs to get ahead of three issues:

• SLA language for hybrid networks: define availability, latency bands, and escalation paths across parties.
• Telemetry requirements: specify what network data you must receive, at what granularity, and how long you can retain it.
• Shared incident ownership: codify how faults are attributed when the customer just sees “no service.”

AI can help here—but only if you contract for the data needed to make AI-driven assurance work.

AI in telecom operations is now a supply chain problem

Here’s the thing about AI in telecommunications: model performance isn’t a one-time event. It decays—new devices appear, traffic patterns shift, network configs change, and training data gets stale.

That means AI operations depends on a supply chain of its own:

• Data supply chain: ingestion, labeling, quality checks, lineage.
• Model supply chain: versioning, validation, rollback plans.
• Compute supply chain: GPUs, edge accelerators, capacity planning.
• People supply chain: SRE + network engineers + data scientists working off the same incident workflow.

In our AI in Supply Chain & Procurement series, we often talk about demand forecasting and supplier risk. In telco AI, you can apply the same thinking:

• Forecast demand for compute, not just network capacity.
• Manage supplier risk for model dependencies (especially third-party foundation models).
• Treat data sources as critical suppliers with quality SLAs.

A procurement checklist for AI network optimization

Use this when evaluating AI for network optimization and 5G management:

1. Outcome definition: What operational metric changes—MTTR, energy per GB, dropped calls, ticket volume?
2. Data requirements: Which counters, logs, traces, and customer-impact signals are needed?
3. Deployment model: Where does inference run (RAN edge, regional DC, public cloud)?
4. Integration points: OSS alarms, ticketing, workforce management, configuration management.
5. Human-in-the-loop: When does AI recommend vs. execute changes?
6. Governance: Audit trails, model drift monitoring, change approvals.
7. Commercial model: licensing tied to nodes, traffic, features, or outcomes?
8. Exit plan: data portability, model portability, retraining support.

If a vendor can’t walk you through items 2–4 clearly, the project will stall—usually after the pilot.

Smartphones getting thin is not a gadget story for telcos

“Smartphones get thin” sounds like consumer tech trivia, but it has two telco implications:

1. Battery and thermal constraints push more intelligence into the network. Devices can’t always brute-force compute-heavy features locally.
2. Device differentiation increases traffic unpredictability. New on-device AI features, camera workflows, and always-on assistants change uplink/downlink patterns.

That makes AI-based network planning more valuable—particularly for:

• capacity forecasting by cell and time of day,
• QoE (quality of experience) prediction,
• traffic classification that respects privacy constraints.

Procurement teams should expect more proposals that bundle “device insights” with network optimization. Be careful: device telemetry is useful, but it’s also a privacy and compliance minefield. Contracting should specify aggregation levels, retention limits, and allowed use cases.

A practical 2026 plan: from pilots to production

The podcast ends with predictions for 2026, but you don’t need a crystal ball to plan well. You need a production mindset.

Here’s what works if you want AI to materially improve telecom operations in 2026:

1) Start with one closed-loop use case that doesn’t scare the org

Good first candidates:

• automated ticket enrichment and routing,
• anomaly detection with explainable alerting,
• energy optimization recommendations (not autonomous changes).

Bad first candidates:

• autonomous parameter changes across the RAN without rollback discipline,
• “single AI platform for everything” programs.

2) Build a shared measurement baseline

Before you deploy anything, establish the “before” numbers:

• MTTR (median and p95)
• repeat incident rate
• truck rolls per 1,000 sites
• energy cost per GB (or per cell)
• NPS/complaint volume tied to network issues

If you can’t measure it, finance will treat it as overhead.

3) Treat AI vendors like strategic suppliers—because they are

That means quarterly business reviews, roadmap alignment, and performance reporting—just like you’d do for critical network equipment suppliers.

And yes, negotiate like it’s telecom:

• price protections,
• service credits tied to measurable outcomes,
• transparency into model updates.

What to do next if you’re buying AI for telecom

AI billions will keep flowing into telecom. Some of it will be wasted. The companies that get value will be the ones that connect AI to operations and treat it like a supply chain discipline: inputs, controls, outputs, and continuous improvement.

If you’re evaluating AI for network optimization, 5G management, or service assurance, start by writing down two numbers you want to move in 2026 (for example: reduce MTTR by 20% and cut truck rolls by 10%). Then procure backward from those outcomes—data, integrations, governance, and only then the models.

What’s your organization’s biggest blocker right now: data readiness, integration complexity, or the commercial model vendors are pushing?