A radio standard usually describes how a handset and a tower exchange bits. Qualcomm’s chief executive wants the next one to describe where the computing happens, which turns an engineering document into an argument about who owns the inference.
Cristiano Amon told the 2026 Mobile Network Conference this week that 6G will be AI native from the design stage rather than a network with machine learning bolted on afterward. The network, in his description, would anticipate how a radio channel will change before it changes, and would decide whether a workload runs on the phone, at the cell site or in a data center. Qualcomm is aiming at commercial launch in 2029.
Three details carry the weight in that sentence, and each of them costs money. The first is prediction. A radio that forecasts channel conditions can spend less capacity on the signalling that today tells a handset how the air between it and the tower is behaving. That is a real gain and a familiar research goal. What is new is placing the model inside the standard rather than inside a vendor’s proprietary scheduler. A standardised predictor needs standardised training data, standardised model interfaces, and a rule for what happens when two vendors’ models disagree about the same channel. None of that exists yet.
The second is compute placement. Amon’s framing puts the network in charge of deciding whether inference runs on the device, at the edge or in the cloud. That is a load-balancing question in a laboratory and a commercial question in a live network, because the three locations have different owners, different cost structures and different regulators. A handset maker wants to sell silicon. An operator wants to sell connectivity and, increasingly, compute. A hyperscaler wants the workload. Writing “the network decides” into a standard settles a negotiation that has not happened.
The third is timing. Qualcomm says the first 6G standard will be complete in 2029 and that commercial service begins in 2029. Those two dates normally sit two or three years apart. A standard is frozen, chips are designed against it, networks are built, handsets follow. Compressing the gap means operators buy equipment specified by a document that does not exist yet, which is what happened at the start of the 5G cycle and which produced a first generation of gear that could not be upgraded without sending a technician to the site.
Qualcomm has one piece of hardware evidence. In September the company said it had completed interoperability testing of base station and device prototypes in the U6G band with carriers, using the most recent 3GPP definitions. Prototypes prove that a link closes. They do not prove a supply chain, a price or a deployment model. Spectrum is the constraint no vendor controls, and it decides whether the 2029 date means anything. The upper 6 GHz range is attractive because it offers wide channels near existing mid-band sites, and contested because each administration allocates it on its own timetable. A vendor can design silicon before a standard is frozen. It cannot design silicon without a band. Without harmonised allocations, a 2029 launch becomes a launch in whichever countries moved first, plus a handset that covers the band in three of them and falls back elsewhere.
There is a plausible reading in which AI native is mostly a positioning statement. Every generation of mobile technology arrives with a phrase that promises more than the standard delivers. 4G was going to be a fabric of applications. 5G was going to be industrial. Both arrived, unevenly, and the marketing ran years ahead of the revenue. Amon’s claim deserves the same scepticism, particularly because a network that allocates compute needs something worth allocating it for, and the applications that would justify the complexity have not been named.
The more interesting reading concerns Qualcomm’s own position. The company sells modems and the patent portfolio that surrounds them. If value in a network shifts toward inference, toward models that predict channels and place workloads, then the portfolio that matters is weighted differently, toward machine learning rather than waveform design. Getting the phrase into the standard’s language early is how a component vendor protects a royalty stream it has not yet earned. What Amon has done is commit his company in public to a date and a design philosophy, which gives customers something to plan against and competitors something to attack. None of this will be settled by a keynote, though. The standards work will run for years, the spectrum decisions belong to the governments that own them, and operators still carrying a 5G investment they have not recovered will approach a sixth generation with unusual caution.
The test is simple. If 6G arrives in 2029 with models inside the radio and compute allocated by the network, the phrase will have meant something. If it arrives with a faster air interface and a slide deck, the industry will have spent four years arguing about a word.
