Ericsson counts 60,000 AI-RAN nodes and eyes inference revenue

About 60,000 network nodes are already running AI-RAN, by Ericsson's own count, and the number is climbing fast. That figure came from CTO Erik Ekudden at TIP FYUZ 2026 in Seattle, where he argued operators don't need to wait for 6G to make their networks AI-capable. Much of what's required, he said, already sits in roadmaps or in the field.

The concrete proof point: Ericsson built AI hardware into its radios earlier this year to handle link adaptation and beamforming. The first models running on that existing kit have lifted spectral efficiency by 20% to 25%. Not bad for a software upgrade, essentially.

How open is the stack?

Ericsson keeps hardware, software and models separate, so operators can mix their silicon. The options include Intel, Arm, Nvidia or Ericsson's own chips. That framing reads as vendor magnanimity, but it also reflects reality — no operator wants one supplier owning the AI pipeline end to end.

Ekudden sketched what he calls an "intelligent fabric," which must give each agent, user and application its own tailored experience. In practice that demands better uplink and ubiquitous performance. Satellite will complement terrestrial networks, he added, and indoor performance matters just as much as the outdoor blanket.

Inference is already the main AI workload, he noted, and someone has to decide what runs on a sensor, on a device, and in the network or data centre. That distribution question will shape operator architectures for years.

What changes in the core?

Ericsson's Service Management and Orchestration platform can now host core network applications for three jobs: preventive healing, capacity planning and threat mitigation. Positioning error, Ekudden said, can be cut by a factor of ten. AI-enabled IMS could also help detect fraud.

The commercial pitch goes beyond engineering. Operators have a third revenue stream beyond bits per second per watt and network APIs, Ekudden argued. They can sell inference itself, lease GPUs, or deliver AI services to enterprises and governments.

"That's really the new world that we are already in, in today's 5G era," he said.

Fair enough. Whether enterprise customers pay for inference at telco scale remains unproven, but the option at least exists on paper.

Does the autonomy level still matter?

Ekudden raised a mildly contrarian point on network autonomy certification. Certifying at Level 3 or 4 autonomy perhaps matters less than it used to, he said, with Level 4 becoming the minimum requirement anyway. The better test, in his view, is whether a network improves end-to-end outcomes rather than a single process. A network that optimises one link beautifully but degrades the session is not much of an achievement.

Operators are already benefiting from differentiated standalone 5G services and from employing AI radio, he noted. That's the present tense part of the story, and it's real.

When does 6G enter the picture?

Not soon, by Ekudden's own timeline. AI-native 6G is still a few years off, he said. Future opportunities cluster around selling inference and sensing, which suggests the revenue story bridges the 5G-Advanced and 6G eras rather than waiting for a standards flag to drop.

For now, the 60,000-node count and the 20% to 25% spectral-efficiency gain are the numbers operators can actually bank. The inference marketplace remains a forecast with a growing hardware base behind it.