AI Billions in Telecom: What 2025 Means for Ops

A funny thing happened in telecom this year: the biggest AI announcements often sounded like “cloud strategy” or “network modernization” rather than “AI strategy.” But when you listen closely to the themes that shaped 2025—vendor shake-ups, satellite finally turning into an operational reality, and operators tightening up internal programs—the common thread is simple: telcos are treating AI as infrastructure, not an experiment.

That framing matters if you sit anywhere near supply chain, procurement, network operations, or customer experience. Because once AI becomes infrastructure, your requirements change: procurement turns into a capacity-planning problem, supplier risk expands into GPUs and power, and your operating model needs controls that are stricter than the “pilot-and-pray” era.

Mobile World Live’s end-of-year podcast episode touched a nerve by calling out “billions pumped into AI” alongside Nokia’s moves and broader operator trends. I’m going to use those talking points as a practical case study: what telecom’s AI spend in 2025 tells us about how to buy, build, and govern AI in 2026—especially through a supply chain & procurement lens.

2025 proved AI is now a telecom capex category

The key point: By the end of 2025, AI spending in telecom stopped behaving like software budget and started behaving like network capex—multi-year, capacity-driven, and constrained by physical supply.

You can see the shift in how leaders talk:

• Not “Which model do we use?” but “Where do we run it—edge, core, or public cloud?”
• Not “Can we automate this process?” but “Can we trust the automation in regulated, safety-critical environments?”
• Not “Do we have data?” but “Is our data supply chain clean enough to operate AI continuously?”

Why this matters to procurement teams

Procurement isn’t being asked to “buy AI.” It’s being asked to secure scarce inputs that make AI operational:

• Compute capacity (GPUs, accelerators, inference-optimized servers)
• Data platform commitments (storage, pipeline tooling, feature stores)
• Network-grade observability (telemetry, log pipelines, event correlation)
• Power and space (data centre power, cooling, rack density constraints)

If you treat those as normal IT purchases, you’ll get burned on delivery timelines, price volatility, and vendor lock-in.

Snippet-worthy stance: If AI is part of your network, then compute is part of your supply chain.

Nokia’s 2025 shake-up: what it signals about AI operations

The key point: Vendor reorganizations and portfolio reshuffles aren’t “corporate drama.” They’re signals about where suppliers think the margin will be: AI-driven network performance and automation.

The podcast flagged a Nokia shake-up as one of the year’s defining items. Whether you buy from Nokia, Ericsson, Samsung, or a mixed stack, the procurement implication is the same: AI is becoming embedded inside network products rather than bolted on.

Here’s what changes when AI moves inside the product:

• You’re not only buying radios or core functions; you’re also buying closed-loop automation logic.
• Upgrade cycles start to look like model lifecycle cycles (more frequent, more risk).
• Contracts need to define who owns failure modes (vendor AI decision vs operator policy).

Contracting for “AI inside the network”

If you only change one thing in 2026 contracting, make it this: treat “AI features” like you treat any mission-critical capability—with clear SLOs and audit hooks.

Practical clauses to push for:

1. Operational SLOs tied to outcomes (e.g., anomaly detection precision/recall thresholds, mean time to detect, mean time to remediate).
2. Explainability and traceability requirements for network-impacting actions (what data was used, what policy fired, what rollback exists).
3. Change-control and model-update governance (release notes for model changes, pre-prod validation, rollback windows).
4. Telemetry access guarantees (you can’t manage what you can’t observe; black-box AI becomes a supplier risk).

Satellite “gets real”: AI meets hybrid network planning

The key point: As satellite connectivity becomes a serious component of telecom service delivery, AI becomes the planning layer that makes hybrid networks manageable.

Satellite “getting real” isn’t just about new coverage maps. It introduces procurement and planning complexity:

• New suppliers and dependency chains (space segment, ground stations, terminals)
• Different performance characteristics (latency profiles, weather sensitivity)
• New regulatory and security boundaries

Where AI actually helps (and where it doesn’t)

AI earns its keep in hybrid networks when it’s used for prediction, allocation, and risk control, not for vague “automation.”

Good use cases:

• Traffic prediction to decide when satellite backhaul is economically justified.
• Dynamic policy control for routing and quality-of-service across terrestrial + satellite.
• Predictive maintenance for terminals and ground infrastructure (failure probability, spare parts planning).

Bad use cases:

• Replacing engineered policy with “let the model decide everything.” Hybrid networks punish that quickly.

Supply chain tie-in: satellite adds hardware logistics and spares management to connectivity planning. That pushes AI into classic demand forecasting and inventory optimization—exactly the theme of this topic series.

Operators rolling back DEI: why it shows up in AI risk

The key point: Internal policy swings matter because AI programs are people programs—data access, model approvals, and operational accountability depend on stable governance.

The podcast referenced US operators winding back DEI efforts. Regardless of your view on that trend, it’s relevant to AI because governance capacity is fragile. When organizations reshape internal programs, it can ripple into:

• Who reviews AI impacts (bias, safety, customer harm)
• How vendor tools are evaluated
• How frontline teams are trained to override automation

From a procurement standpoint, this isn’t political; it’s operational. If governance weakens, vendor risk increases. You’ll need tighter acceptance criteria, more robust testing, and clearer escalation paths.

Memorable one-liner: AI doesn’t fail like software. It fails like a process—with people on the hook.

Smartphones “get thin”: a quiet driver of AI supply chain pressure

The key point: Device differentiation and hardware cycles influence telecom AI because they shape traffic patterns, edge workloads, and customer expectations.

Thinner phones might sound cosmetic, but device trends often correlate with:

• More on-device AI features (voice, camera, assistants)
• Higher uplink/downlink bursts (media creation, real-time personalization)
• More customer support complexity (AI features break in weird ways)

That feeds back into network planning and customer operations, both of which increasingly rely on AI. If customer experience teams deploy agentic AI to handle support, the quality of device telemetry and entitlement data becomes critical.

Procurement lens: data is the new spare part

In supply chain terms, telcos historically stocked spares: routers, optics, handsets. In 2025, the scarce resource is often high-quality, timely data:

• Device capability data
• Service entitlement data
• Network event streams
• Ticket and interaction histories

If your data “inventory” is incomplete or delayed, AI performance drops, and you’ll overbuy compute to compensate. That’s a real cost pattern I’ve seen: teams scale infrastructure to mask data issues.

Where the “AI billions” are really going (and what to do about it)

The key point: The biggest AI budgets in telecom are flowing into five buckets: compute, cloud platforms, data engineering, security, and operationalization.

If you’re building your 2026 roadmap, assume these are the real line items you’ll be negotiating:

1. Inference capacity (often larger than training in steady-state telecom ops)
2. Data pipelines and streaming (near-real-time telemetry is table stakes)
3. Model management and monitoring (drift detection, evaluation, audit logs)
4. Security and privacy tooling (PII controls, isolation, threat monitoring)
5. Integration work (OSS/BSS integration, workflow automation, knowledge bases)

A procurement checklist that actually works

Here’s the checklist I recommend for telecom AI purchases (vendor tools, platforms, or managed services). It’s short because teams will use it.

• Outcome definition: What metric changes in 90 days? (MTTR, truck rolls, call handle time, churn risk)
• Data readiness: Which sources are required, and what’s their current quality/latency?
• Cost model clarity: $/inference, $/1K events, storage egress, retraining costs, support tiers
• Portability: Can you export models, prompts, and logs? Can you run the core logic on different compute?
• Controls: Human-in-the-loop options, rollback, approval workflows, audit trails
• Supplier resilience: Roadmap stability, dependency on single cloud/GPU vendor, support response times

If a supplier can’t answer those cleanly, you’re not buying a solution—you’re buying a future firefight.

What to expect in 2026: the telecom AI “operating model” year

The key point: 2026 will reward telcos that industrialize AI—treating it like a living production system with supply chain discipline.

Three predictions I’m comfortable putting my name on:

1. Network AI will be measured like radio performance. If vendors can’t prove outcomes, features won’t renew.
2. Compute procurement will look like energy procurement. Multi-year commitments, hedging, and capacity planning will become normal.
3. Customer ops will split into two lanes: fast AI-assisted self-service and high-touch “human escalation” teams. Companies that blur the lanes will see higher churn.

Next steps for supply chain & procurement leaders

If you’re responsible for procurement, vendor management, or operational readiness, here’s a clean plan for Q1 2026:

1. Inventory your AI dependencies: compute, data sources, vendors, and critical workflows.
2. Standardize acceptance tests: model accuracy is not enough—add resilience, auditability, and rollback tests.
3. Renegotiate contracts around operational outcomes: tie renewals to measurable improvements.
4. Build a supplier risk map for AI infrastructure: single points of failure in cloud region, GPU supply, or proprietary telemetry.

The reality? Telecom is turning AI into plumbing. Once it’s plumbing, it has to be reliable, governable, and procurable at scale.

If you’re planning your 2026 AI roadmap, ask one forward-looking question: when AI makes a bad call in production, do you have a contract, a control, and an operating process that can contain it in minutes—not days?