The vendors and the architecture bets

Open RAN has an eight-year paper trail, and its latest participant count is lower than the one before

The open radio access network has been going long enough to measure: eight years of documents, from a founding release to a membership page. The most recent number in that trail is smaller than the one before it (Exhibit 1).

Five operators announced the O-RAN Alliance on 27 February 2018: AT&T, China Mobile, Deutsche Telekom, NTT DOCOMO and Orange. It merged two earlier groups, the xRAN Forum and the C-RAN Alliance, and was set up as a German entity that August. By the first board meeting, in Shanghai on 28 June 2018, twelve operators had signed its constitution.

Eight years on, its own overview, dated July 2026, reports over 220 participants, over 150 published specifications and eleven working groups. It counts twelve PlugFests since 2019, with 31 participants across nine labs at the spring 2026 event, and it records that its specifications are being transposed into national standards through ATIS, ETSI and TTA. During 2026 its software community is moving under Linux Foundation Networking. On the face of it, a body doing what it was set up to do.

The participant count needs its history beside it, though. The March 2025 edition of the same overview reported over 300 participants. The membership page, read on 2 October 2026, listed 28 operator members and 194 contributors, which is 222 in all. The alliance has not said whether the fall is attrition or a change in who gets counted, and we cannot tell from outside. What its own documents do say is that the number fell by about a quarter in sixteen months while the specification count went up.

What do the alliance’s own documents establish? That open interfaces are specified, tested and transposed, and that an operator-founded body does the work. What they do not establish is share. The overview counts documents and participants, and no primary source we could check gives a figure for the open share of deployed RAN. So a 6G architecture that assumes open interfaces is assuming two things at once: a specification base, which exists, and a market share, which has to be shown separately.

AI-native is two years old as an alliance and three releases old in 3GPP

AI-native has existed for two years as an industry body and for three releases as a 3GPP work area, and the alliance and 3GPP tell different stories (Exhibits 2 and 3).

SoftBank launched the AI-RAN Alliance on 26 February 2024 at Mobile World Congress, with ten other founders: Amazon Web Services, Arm, DeepSig, Ericsson, Microsoft, Nokia, Northeastern University, NVIDIA, Samsung Electronics and T-Mobile USA. Its stated goal was to make mobile networks more efficient, cut their power consumption and retrofit existing infrastructure, so that telecommunications companies could open new economic opportunities with AI. It has grown quickly. The alliance reported 75 members in February 2025 and passed 100 in July 2025, and on 26 February 2026 it reported 132 and seated Qualcomm, SK Telecom and Vodafone on its board (Exhibit 2).

When we read its members page on 2 October 2026 it listed 150 organisations: ten on the executive board, three on the general board and 137 general members. Amazon Web Services, a founder, was no longer among them. The alliance works in three groups, named for the three relationships it sees between its two subjects: AI for RAN, AI and RAN, and AI on RAN. The showcases around its 2026 congress, as SDxCentral reported them, were a Nokia radio line, a Samsung virtualised RAN test with Vodafone on Intel hardware, and a SoftBank edge deployment with NVIDIA.

In 3GPP the same subject has identifiers, and identifiers can be checked. Release 18 listed artificial intelligence and machine learning as one of its three new topics. Release 19 made it normative, in a work item on AI and machine learning for the NR air interface, RP-240774, which the ATIS webinar called the largest project in the release. It also opened a study on AI for mobility, RP-240082. The work item covers one-sided models, meaning models that live wholly in the phone or wholly in the network, for beam prediction and positioning. Release 20 goes a step further with a work item on two-sided models, RP-251870, in which a phone and a base station from different companies have to share a trained model. Release 20’s fourteen 6G study items include the radio study whose report, TR 38.914, was approved in June 2026. The normative decision belongs to Release 21, which is scoped from March 2027.

The study description sets the boundary, and it is a telling one. It requires the 6G radio and RAN design to ensure that the 6G system can also operate without AI and machine learning. So AI is optional by design in the standard, and AI-native describes products built on top of it rather than the standard itself. That is as far as the normative text goes: one frozen work item, one two-sided work item in progress, and a 6G study report waiting for its work items (Exhibit 3).

The AI-native commitment that can be checked is one Release 19 work item, one Release 20 work item and one study report. The alliance list shows who has signed, and most of the industry has. The showcases show trials. A vendor roadmap lives in the gap between the two, and a buyer is entitled to ask for that gap to be filled with a work-item number or a deployment rather than a membership.

Three vendors spend a fifth to a quarter of revenue on research, against flat network sales

Standards work and alliance memberships are paid for out of research budgets, so it is worth asking how big those budgets are and what they are up against. The four largest radio vendors publish enough to compare, though not all of them in the same shape (Exhibit 4).

The three vendors that disclose a research line spend between a fifth and a quarter of every sale on it (Exhibit 5). Three things follow from that.

• Direction. Ericsson’s R&D fell, from SEK 53.5 billion to SEK 48.9 billion in 2025, and Nokia’s rose, from EUR 4,512 million to EUR 4,855 million. Network sales fell at both, by 5% and 4% as reported, and Ericsson expects the RAN market to be flat in 2026. That is the market a 6G radio has to sell into.
• Geography. The largest research budget in the table, Huawei’s CNY 192.3 billion, belongs to a company whose equipment is shut out of the American market this series covers. So the biggest budget and the budgets that sell into this market are different money.
• Scale. SEK 48.9 billion is the whole of Ericsson’s research, across every product, and the 6G share is not disclosed. A vendor roadmap that implies a research programme beyond what these lines can fund is making a claim its own annual report does not support.

Three questions sort a roadmap from a slide

Put the three sets of documents together and a buyer has three questions to ask any vendor.

1. Architecture. Which 3GPP work item or O-RAN specification does the claim rest on, by number? A claim without one is a plan, and a plan is not a commitment.
2. AI. Which operator has published a deployment? A trial at a trade show is a trial, however good the demonstration.
3. Timeline. Which research budget pays for it, and what does the same annual report say about the market the product is meant to sell into?

A roadmap that answers all three is a document. One that answers none of them is a slide.

Read this way, the bets are narrower than the words. Open is a specification base with an operator-founded body behind it, an undisclosed share, and a participant count lower than it was. AI-native is one frozen work item, one in progress, a study description that keeps AI optional, and a membership list that has grown and lost a founder. Both are real. Both are early. Neither, in our view, is a reason to move the dates in part 1.

The terms, briefly

• Open RAN. A radio access network whose interfaces between radio, baseband and controller follow published specifications, so that equipment from different suppliers can be mixed. The O-RAN Alliance writes the specifications, which extend the interfaces 3GPP defines.
• AI-RAN. The alliance’s umbrella for three things. AI for RAN uses AI to run the radio better. AI and RAN shares the same hardware between AI workloads and the radio. AI on RAN runs applications on that hardware at the edge.
• Work item and study item. In 3GPP a study produces a technical report and no specification text. A work item produces normative text. RP-numbered documents are the RAN plenary’s approvals of each.
• R&D intensity. Research and development expense as a share of revenue. It measures how much of each sale is reinvested, and says nothing about what any single programme costs.