AI in the RAN Hits the Scale Test: AT&T and Verizon Move From Proof Points to Economics

Executive briefing memo | 19 October 2026 | Prepared for network strategy, RAN planning, and vendor management leadership

Bottom line up front

The AI-in-the-RAN debate has quietly changed shape. Speaking at RCR Wireless News’s Telco AI Forum 2026, Adam Loddeke, AVP of RAN technology at AT&T, and Cindy Dong, associate director of RAN planning at Verizon, described deployments that are already producing measurable results — improved downlink spectral efficiency, better coverage, faster planning and configuration cycles. Neither executive framed the challenge as whether the technology works. Both framed it as where AI creates enough value to justify deployment at scale, and how to manage it once it is there.

That is a materially different conversation from the one the industry was having a year ago, and it has direct consequences for Open RAN operators, vendors, and anyone building a business case around intelligent RAN. The proof-of-concept era is closing. The portfolio-management era is opening.

What was actually said

Two operator positions emerged from the forum, and they are complementary rather than identical.

AT&T is running AI broadly across the network lifecycle. Loddeke said the operator is already using AI at scale, primarily through applications and rApps spanning planning and forecasting, deployment and configuration, and network optimization and lifecycle management. Separately, AT&T has been trialing Ericsson’s AI-native link adaptation feature embedded directly in the RAN and has seen what Loddeke characterized as “significant improvements” in throughput and spectral efficiency.

Verizon is pursuing a dual RAN AI strategy. Dong described a split architecture: centralized rApps operating in a non-real-time environment, combined with local AI integrated into the gNodeB for real-time optimization. Verizon’s work with Ericsson has demonstrated downlink spectral efficiency improvements from AI-based link adaptation, and a separate field trial produced coverage improvements.

Read together, the two positions describe the same structural insight from different angles. AT&T emphasizes the lifecycle — the long tail of planning, configuration, and optimization work where AI can be applied today with existing tooling. Verizon emphasizes the latency split — the recognition that some intelligence must sit centrally where it can see the whole network, and some must sit in the base station where milliseconds matter. Neither operator is treating AI as a single product. Both are treating it as a portfolio allocation problem.

Why this is the story that matters this week

The candidate news pool this week included several legitimate items: VIAVI’s AI-RAN and 6G test portfolio showcased at India Mobile Congress, 1Finity’s continued India push around AI-native RAN and Open RAN, Ericsson’s new CEO identifying AI RAN as a growth opportunity alongside satellite-terrestrial convergence, and a broad TelecomTV roundup of AI-native telco activity. Each is relevant. None of them answers the question that actually determines whether AI-RAN becomes a durable part of the network or a stranded lab investment.

The AT&T and Verizon commentary does, because it comes from the people who hold the budget. When an AVP of RAN technology says the technology works and the question is now economic justification at scale, that is a signal to the supply chain about what the next twelve to twenty-four months will be spent on. It is also a signal about what will get cut.

The three constraints operators just named

The forum commentary surfaced three constraints that will shape deployment decisions. Each deserves separate treatment in any operator’s planning cycle.

1. Value density, not coverage

Scaling changes the challenge. The phrase used in the reporting — that moving beyond trials will require open architectures, automation, and disciplined lifecycle management across distributed networks — is doing a lot of work. It implies that the marginal return on AI features is not uniform across the footprint.

Link adaptation gains in spectral efficiency are most valuable where spectrum is scarce and traffic is dense. Coverage improvements from AI-driven optimization are most valuable at the edge, where the cost of adding a site is highest. A nationwide rollout that applies the same AI feature uniformly across all cells will produce an average benefit that looks unimpressive against the average cost. A targeted rollout that concentrates AI where the marginal value per cell is highest will look very different.

This is the core economic insight operators are converging on: AI in the RAN is not a coverage product, it is a value-density product. That has implications for how features are licensed, how they are measured, and which cells get them first.

2. The real-time / non-real-time split is an architecture decision, not a marketing one

Verizon’s dual strategy — centralized rApps in the non-real-time RIC alongside local AI in the gNodeB — is the clearest public articulation yet of where the intelligence boundary is being drawn in practice.

The non-real-time layer handles policy, forecasting, and slower-loop optimization. The gNodeB handles link adaptation and other functions where the control loop cannot tolerate backhaul latency or centralized round-trips. This is not a new idea, but it is now being described by an operator as an operating model rather than a research concept.

For Open RAN specifically, this matters because the boundary determines which interfaces carry the traffic and which vendors own which function. Operators that push more intelligence into the real-time and near-real-time layers will need RIC platforms and xApps that can keep up. Operators that concentrate intelligence in rApps will need robust non-real-time data pipelines and clean lifecycle management. The two paths have different cost structures and different vendor dependencies.

3. Lifecycle discipline is the hidden cost center

Both operators emphasized lifecycle management, and the reporting language is explicit: scaling requires “disciplined lifecycle management across distributed networks.”

This is where AI-RAN business cases most often break. A single AI feature trialed in a handful of cells is a science project with a clean cost. The same feature deployed across tens of thousands of cells becomes a software lifecycle problem: model versioning, retraining cadence, regression testing, rollback procedures, performance monitoring, and vendor support boundaries. Each of those is a real line item, and none of them appear in the spectral efficiency gain quoted from a field trial.

Open architectures help here, but only if the operator has the automation to exploit them. Disaggregation without automation converts a hardware integration problem into a software integration problem. That is progress, but it is not free.

What this means for Open RAN operators

The AT&T and Verizon positions are not explicitly Open RAN stories, but they are Open RAN-relevant in three concrete ways.

First, the rApp ecosystem gets a clearer demand signal. AT&T’s description of AI applied across planning, forecasting, deployment, configuration, optimization, and lifecycle management is essentially a specification for the rApp market. Operators evaluating rApp vendors should now be asking a sharper question: not whether the app demonstrates a gain, but whether it can be operated across a distributed footprint with versioned models and measurable rollback. Vendors that cannot answer that question will not survive procurement.

Second, the intelligence boundary becomes a procurement lever. Verizon’s dual strategy gives operators a template for negotiating with vendors. If the real-time function stays in the gNodeB and the slow-loop function sits in the non-real-time RIC, the operator can source those two layers separately — provided the interfaces are genuinely open. That is the commercial argument for open interfaces that has always been stronger than the ideological one.

Third, the testing and validation layer becomes strategic. VIAVI’s IMC 2026 showcase — AI RSG for RAN simulation and validation across 4G, 5G, emerging 6G, Open RAN, and NTN environments, with scenario generation scaling from thousands of cells and UEs to city-scale environments, plus AI Inference Test to measure whether responses stay accurate, consistent, and timely under live RAN load — describes exactly the tooling category that lifecycle discipline requires. Operators scaling AI in the RAN will spend more, not less, on validation. That is a budget line worth planning for now.

The vendor read-across

Ericsson sits on both sides of this story. Its AI-native link adaptation is the feature both operators cited for spectral efficiency gains, and its new CEO has publicly identified AI RAN as a growth opportunity. That is a strong position. It is also a concentration risk for operators: the same vendor supplying the RAN and the embedded AI function reduces the competitive pressure that keeps feature pricing honest.

This is where the Open RAN argument reasserts itself, but with a twist. The value of multi-vendor RAN in an AI context is not primarily about swapping radios. It is about preserving optionality on where intelligence lives and who supplies it. An operator that can move a function between the gNodeB, the near-real-time RIC, and the non-real-time RIC without renegotiating its radio contract has real leverage. An operator that cannot has a single-source dependency dressed up as an architecture.

Meanwhile, the broader vendor landscape continues to position around this. 1Finity’s India Mobile Congress messaging centered on AI-native RAN and Open RAN alongside all-photonics networking. VIAVI is building the validation layer. Ericsson is building the embedded feature layer. The question for operators is which of these layers they want to own, which they want to rent, and which they want to keep contestable.

Decision points for the next two quarters

The forum commentary implies a set of decisions that operators should be forcing now rather than later.

Define the value-density map. Before committing to a nationwide AI feature rollout, model the marginal benefit per cell for each candidate feature. Spectral efficiency gains and coverage gains have different geographic distributions. A feature that pays in dense urban cells may not pay at the edge, and vice versa. The rollout plan should follow the map, not the marketing deck.

Lock the intelligence boundary in contracts. The real-time versus non-real-time split should be an explicit term in RAN and RIC procurement, not an implementation detail discovered during integration. Operators should require that functions assigned to the non-real-time layer be portable across vendors.

Budget for lifecycle, not just licensing. Model retraining cadence, regression testing, and rollback as recurring costs from day one. The operators that succeed at scale will be the ones that treated AI features as software with a maintenance burden rather than as hardware capabilities with a one-time price.

Invest in validation capacity. The ability to simulate city-scale environments, inject controlled anomalies, and measure inference accuracy under live RAN load is becoming a prerequisite for scaling. Operators without that capability will discover problems in production.

Preserve multi-vendor optionality where intelligence lives. The strategic value of Open RAN in this phase is not radio interchangeability for its own sake. It is the ability to re-source intelligence functions as the economics become clearer.

The skeptical view

It is worth stating the counterargument plainly. Neither operator disclosed the magnitude of the gains in absolute terms, the size of the footprints involved, or the cost side of the equation. “Significant improvements” in throughput and spectral efficiency is a directional statement, not an economic one. Coverage improvements from a field trial do not automatically translate into avoided capex at scale.

There is also a real possibility that the value concentrates in a minority of cells and that the nationwide business case remains marginal for several years. The reporting itself acknowledges that scaling requires open architectures, automation, and lifecycle management — three things the industry has been promising to deliver for some time and has delivered unevenly.

The honest read is that AI in the RAN has moved from “does it work” to “where does it pay,” and the answer to the second question is not yet known. What has changed is that the question is now being asked by the people who control the budget, with real field data in hand. That is progress, and it is the right kind of progress.

What to watch next

The next signal will be disclosure of deployment scope rather than trial results. Watch for operators naming cell counts, feature licensing models, and measured cost-per-cell for AI functions. Watch for rApp vendors publishing lifecycle management capabilities alongside performance claims. Watch for the testing vendors to report demand patterns that reveal how many operators are actually scaling rather than trialing.

And watch the intelligence boundary. Every public statement about where AI sits — in the gNodeB, in the RIC, in the orchestration layer — is a signal about which vendor relationships operators intend to keep contestable. That is where the Open RAN story will be decided in this phase, not in the radio.

Sources

• AI in the RAN moves toward the scale test — RCR Wireless News
• VIAVI Showcases AI-RAN, 6G and Next-Gen Network Testing at IMC 2026 — TimesTech
• AI-Driven Telecom Networks: VIAVI on Testing and Network Automation — Electronics Media
• The week in AI-native telco (10-16 Oct 2026) — TelecomTV
• 1Finity Connects with Customers, Partners and Industry Leaders at India Mobile Congress 2026 — IT Voice
• SatCos and price hikes on customer minds as Ericsson begins new era — TelcoTitans