Ericsson Puts a 10 Percent Spectral Efficiency Number on AI-RAN and Says It Is Already Billing for It

Ericsson chief Per Narvinger used the company’s third-quarter results call to defend the vendor’s custom-silicon strategy for AI-RAN, and he put a number on the table: a 10 percent gain in spectral efficiency from AI-native link adaptation, already running across more than 60,000 base stations with three operators. The figure is the most concrete AI-RAN revenue claim any vendor has made so far, and Narvinger went further by translating it into a market value reference, pointing out that SpaceX just paid roughly $8 billion for spectrum and that 10 percent more efficiency on that same spectrum would translate to $800 million of value.

The Q3 numbers behind the claim show Ericsson posting seven percent organic sales growth in a market Narvinger described as very flattish. The growth was helped by pre-emptive purchasing from operators worried about AI-linked component price rises, and Ericsson warned that those component costs will pressure margins through 2027. The flat market backdrop is what makes the AI-RAN number matter: Ericsson is looking for upsell on installed base, with no new spending cycle to wait for. Narvinger said the firm has deployed the AI-native link adaptation solution with three operators, with Optus named among them, and that AT&T, Bell Canada and T-Mobile are testing it.

On the question of whether Ericsson has closed the door on putting GPUs into its baseband units, Narvinger was direct. He said the company is not philosophically against anything and that disaggregated software and hardware gives Ericsson a lot of flexibility in the future. He then anchored the present-tense argument: right now it is an extraordinary asset to have Ericsson’s own ASICs, because it makes it possible to deliver functionality fast and be very cost competitive. The next-generation hardware may or may not deploy on GPUs, but any GPU pivot would require a site visit to deploy new hardware, whereas the ASIC-based AI-native link adaptation can be pushed to basebands already in the field today.

Dell’Oro Group’s latest RAN tracker, covering the first half of 2026, puts the top five suppliers as Huawei, Ericsson, Nokia, ZTE and Samsung, together holding 96 percent of worldwide macro RAN revenue. The research firm said the market continued to stabilise in the third quarter after two steep years of decline that erased roughly $8 to 9 billion of RAN revenue between 2021 and 2024, and it expects flat growth to hold through 2026. Revenue rankings were unchanged in the first quarter: Huawei first, Ericsson second, Nokia third. That is the backdrop against which Ericsson is turning AI-RAN into a billing line: there is no new RAN spending cycle to wait for, so the vendor is looking for upsell on installed base.

The strategic gap between Ericsson and Nokia on this point is widening. Nokia has moved to put GPUs into its baseband units as part of its AI-RAN platform work, while Ericsson is betting that the fastest route to AI-RAN revenue is to make existing silicon smarter and leave the GPU door open. Narvinger framed the 10 percent efficiency gain as a differentiator when competing in the market, and said some of that value should come Ericsson’s way. That is an explicit upsell pitch, not aspirational research.

The structural change at Ericsson announced alongside the results is the dissolution of the enterprise wireless division, home to private 5G and neutral-host solutions, into the networks business. Asa Tamsons, senior vice president and head of the group, confirmed she will depart the role and continue as an executive advisor during a transition. She listed the division’s achievements as 200 percent business growth since 2020, nearly 40,000 enterprise customers, about 3,000 channel partners, and deployments across industrial verticals.

Ericsson framed the move as a way to streamline R&D and innovate around indoor connectivity, including for Starlink-style satellite integration with terrestrial networks. Outside observers have noted the similarity to Nokia’s retreat from its private wireless campus edge unit.

Supplier’s note: We stock Ericsson Radio System basebands, including the Baseband 6648 and the AI-capable Radio Processors that host the AI-native link adaptation Narvinger referenced, and we can quote against legacy ERS basebands where operators are being asked to add the AI link-adaptation feature onto sites already in the field. The 60,000-base-station installed base he cited maps to the same population of Radio Processors we cross-reference for customers running mixed Ericsson and Nokia RAN estates, and the upgrade path for those units is a software load, skipping the hardware swap, which shortens lead time from the usual 12 to 16 weeks for new baseband to roughly four weeks for a feature-key rollout. For the GPU baseband strategy Nokia is pursuing, we are not yet holding the GPU-equipped variants in distribution, and any project quoting them should expect a 2027 availability window at the earliest. On the dissolved enterprise wireless side, the Ericsson Private 5G and neutral-host portfolios we carry for industrial customers remain available through the networks business unit, and we have seen no change to the product list or the support model in the two weeks since the announcement.