Essential, But Only at the Edges
The “edge” between a device-side encoder and a network-side decoder, where essential patents are most likely to live.
In one line: 6G will use AI models that the standard says it won’t standardize, yet the standard is being forced to specify how those models pair up. That narrow zone is where the next big patent fights will happen.
Every time your phone connects to a network, it relies on hundreds of patents you will never see. They are called standard-essential patents (SEPs): patents that a company cannot avoid infringing if it builds a product that follows a mobile standard such as 4G, 5G or, soon, 6G. Their owners must license them on FRAND terms, meaning fair, reasonable and non-discriminatory. A patent is “essential” if it reads directly on the text of the standard.
That system rests on a quiet assumption: the standard tells you exactly how the technology works.
AI is about to test that assumption.
The tidy story, and why it fails
Here is the version most people tell. Mobile standards specify how a radio must behave. AI arrives, the standards body (3GPP) declines to standardize neural networks, and patents on those networks fall outside the standard. Case closed.
It is a good story, and I think it is wrong. The standard cannot stay out of the model, and the calendar shows why.
Four years, three steps
The Release 18 study on AI/ML for the air interface (the radio link between phone and network) was the first time 3GPP applied AI directly to core air-interface problems. It set up graded levels of collaboration between network and device, starting from none at all. The study showed real performance gains, but it also exposed serious problems with interoperability (whether one vendor’s model works with another’s) and generalization (whether a model trained in one setting works in another).
Release 19 then moved to normative specifications, meaning binding rules, for AI-based beam management, positioning and CSI prediction. CSI, or channel state information, is the report a phone sends so the network knows how good the radio channel is. The core work was completed in 2025, while performance evaluation and device conformance work run on through 2026.
The pattern was consistent: specify the frame around the model and leave the model alone. That works when the model sits on one side of the link. A beam predictor inside a phone needs nobody’s cooperation, and neither does the CSI prediction that made it into Release 19, which runs on one side while the other stays unchanged.
Then the model needed a partner
CSI compression breaks the pattern. An encoder in the device shrinks the report, and a decoder in the base station must unpack it. The two halves must work as a pair, and they may come from different vendors who never meet until deployment.
Get Shikkha Srivastava’s stories in your inbox
Join Medium for free to get updates from this writer.
Release 19 left that problem open. So in June 2025, 3GPP started a Release 20 work item on two-sided models, scheduled to continue through 2027. Its scope covers model pairing, inter-vendor training collaboration, signaling, monitoring, fallback and performance evaluation. It includes a model-pairing handshake and data collection for training.
Read that list as an essentiality analyst. None of it is a neural network’s internals, yet to make two vendors’ models interoperate, the specification has to say something about how they are paired, trained and checked. That is claim-chart material in a place where none existed a year earlier.
The standard stopped telling you how, and interoperability forced it to start again, a little at a time. Essentiality will be decided in that reluctant middle.
The 6G clock is now official
At its plenaries in Singapore in June 2026, 3GPP approved the first timeline for early 6G specifications. The package is due for approval in March 2027, the architecture stage freezes in June 2028, and the functional specifications freeze in December 2028. The final ASN.1 freeze follows in March 2029.
Money is following the dates. South Korea’s ETRI has launched a 47 billion won (about $31 million) AI-RAN programme running from April 2026 to December 2030. It brings together the country’s three big operators, domestic vendors and four universities to build AI-RAN software on Releases 19 and 21 and test it in a digital-twin network.
Software is being written and money committed before the 6G specification is final. For patent owners, the window from now to late 2028 is not a waiting period. It is the drafting period.
What I would tell a client
Draft for the seams, not the weights. The claims most likely to be essential describe what crosses the interface: inputs, how outputs are quantized and reported, how a model is identified, when it applies, what triggers monitoring and fallback. Claims on architecture or training tricks will struggle to read on anything normative.
Prepare for an evidence fight. You cannot chart a neural network against a clause. Disputes will turn on test-bed results and on whether alternative models reach the same performance. Start building a reproducible record now.
Ask who pays before anyone else does. In a two-sided system, the chipset maker, the device maker and the network vendor each supply part of the function. The licensing chain that settled in 4G and 5G does not map neatly onto a shared encoder-decoder pair. Whoever frames the question first will shape the answers.
Press apportionment. FRAND royalties should reflect the value of the standardized contribution. In AI-RAN, much of the gain may come from proprietary data, compute and tuning that nobody standardized. Implementers will say the standard is only plumbing. Owners must show the plumbing is what unlocked the gain.
The gap nobody owns
FRAND commitments attach to patents. But a feature can quietly depend on a dataset, a pretrained model or a reference design, protected by trade secret or copyright and untouched by any IPR policy. A standard can end up functionally dependent on something its licensing framework never reaches. I would rather see that raised at a standards body now than argued in court in 2032.
The better question
The old test asked whether a claim reads on the standard. The better test for AI-native radio asks whether it reads on anything the standard is forced to control.
Today that set is small, and it sits at the edges: the pairing handshake, the signaling, the monitoring and fallback triggers. It is growing, and the 6G decisions of the next two years will decide how far it reaches into the model.
