What 6G will do, and who is asking for it

Three of the eight capabilities depend on a party outside the carrier

Part 6 sorted the candidates by a release, a band and a device, and for 5G that was enough. It is not enough here. Three of the eight capabilities called 6G cannot work without a partner the carrier does not own, or a permission no carrier can grant itself (Exhibit 1).

The eight appear in 3GPP’s study description and in every vendor’s 6G paper. They are satellites built into the standard, sensing, a new band around 7 GHz, three different things called AI, energy, and the migration from 5G with the core it lands on. We scored each on five dependencies rather than three, and a cross marks one that is needed, unresolved and nobody’s job.

Read across, the map turns the roadmaps upside down. The capabilities furthest along lean on someone else, while the two the operators argue hardest about, energy and the migration, are wholly inside their own fence. It is easy to see why. Those two decide the next capital cycle.

Satellite service is the only capability with a retail price

Where there is no tower, a phone can keep working by talking to a satellite as if the satellite were one. Whether that is simple or hard depends on where the base station is, on the ground or in orbit, and the standard allows both (Exhibit 2).

A mobile base station is a radio and a computer, and the question is which goes up. In the Release 17 design the computer stays on the ground and the satellite only relays the phone’s signal to a ground station; 3GPP calls that a transparent payload. In the Release 19 design, RP-240775, the whole base station is on board, which 3GPP calls regenerative. It costs more, and it is the only one that works when no ground station is in view, which over an ocean is most of the time. Either way, the orbit sets the delay.

For 6G, 3GPP has decided that the satellite is part of the design from the first release. The study description, RP-251881, tells the working groups to aim at a harmonised radio design for terrestrial and non-terrestrial networks. Still open are unpaired satellite spectrum, operation without satellite positioning and the core’s satellite principles.

So what does it sell for? T-Mobile priced T-Satellite at $10 a month at its launch on 23 July 2025, and in July 2026 told investors that satellite accounts for 0.0002% of its network usage. Rogers charges C$15 a month, One NZ folds texting into its plans, and Verizon gives texting away. No American carrier has disclosed satellite revenue or a take-rate.

The serious money is going somewhere else. Apple committed up to $1.1 billion to Globalstar for 85% of its network capacity. SpaceX did not wait for 3GPP; by independent measurement its satellites fly conventional LTE base stations with proprietary adaptations. In September and November 2025 it agreed to buy EchoStar’s spectrum, $19.6 billion in all, to build next-generation satellites with what its president called a step change in performance. By mid-2026 it was reported to be considering a retail mobile service of its own.

Satellite service is the only 6G capability with a retail price, and the price is for coverage insurance. The native design is a commercial move as much as a technical one, in our reading: it keeps the subscription with the carrier and makes the satellite operator a supplier. Every vendor and every operator that has spoken supports it. The party with $19.6 billion of spectrum has not spoken.

Sensing has a government customer and no permission from any regulator

The signals that carry calls also bounce off things that are not phones, so a network that listens for echoes can detect a drone or count a crowd. That is radar, and the standard has so far specified only its simplest form (Exhibit 3).

Distance comes from the delay of the echo and speed from its frequency shift, and the finer the detail you want, the wider the signal you need. A 20 MHz channel separates objects about 7.5 metres apart; 400 MHz gets down to 40 centimetres, which is why sensing is tied to the wide channels of part 2. In monostatic sensing one base station listens for its own echo. In bistatic sensing a second site listens, which lets ordinary sites act as one large radar. That is how the AT&T and Ericsson demonstration outside a stadium in July 2026 tracked drones at 300 to 400 feet.

What 3GPP has written down is a good deal narrower than the demonstrations. The first normative work item, RP-261566, approved at RAN#112 on 11 June 2026, specifies base-station monostatic sensing for drone detection only, with no impact on devices. For 6G, TR 38.914 requires detection and tracking of drones, people, vehicles and automated guided vehicles. Still open are the waveform, the reference signals, who operates the sensing function and who may see its output.

The vendors agree on the headline and split on the physics. Samsung’s trial with Verizon in Dallas in September 2026 drew crowd-density maps from ordinary phones over 40 MHz of CBRS, and Ericsson expects defence and public safety users first. On the waveform, Samsung says “OFDM already provides strong sensing capabilities” and Ericsson does not think a new one is needed. A United States Department of Defense programme lead said OFDM is not optimal for sensing. 3GPP’s decision to keep OFDM settles the question for Release 21. The defence buyer, if it wants better, will have to buy something else.

Who is paying? So far, the state. Outside China every sensing deployment is a trial. Inside China, China Mobile had more than 500 sensing-capable base stations on 42 low-altitude routes by April 2025, and in February 2026 the ministry ordered the operators to track drones below 300 metres in ten pilot cities. Verizon’s chief technology officer said in September 2026 that businesses who saw its drone-detection work would be glad to pay for it. Lockheed Martin plans a drone-detection service on Verizon’s network for 2027. Those buyers buy by tender, with a specification, a liability clause and a multi-year term, against the dedicated radar they use today.

Sensing has the clearest first customer and the least clear permission. Because the customer is the state, sensing is a procurement business, and no regulator has yet said whether a licence to carry traffic is a licence to watch. 3GPP has left the phone out of sensing for now, which keeps the market smaller than the demonstrations imply. Our integrated sensing series takes the cost and the buyers further in part 5.

The first 6G band is a centimetre-wave band, and the terahertz ambition has gone

The new band adds capacity by using spectrum nobody has used for mobile before, around 7 GHz, in channels two to four times wider than today’s. Higher frequencies lose more signal, and the industry has moved its ambition from frequency to antennas (Exhibit 4).

The loss between a tower and a phone rises with the square of the frequency, as part 2 sets out, so doubling the frequency costs about 6 dB. Against the 3.5 GHz the 5G build was priced on, 7.125 GHz costs 6.2 dB and 15.35 GHz costs 12.8 dB. The D-band at 140 GHz, which the first 6G papers were written around, costs 32 dB. The way back is an antenna that is larger in wavelengths, which narrows the beam. A typical 5G panel has 64 transceivers behind 192 elements; the 7 GHz prototypes of 2026 have 256 ports behind 1,024, better than five times the elements. 3GPP has drawn its line at 52.6 GHz, and sub-terahertz was never proposed.

The retreat can be measured, and Samsung wrote down both ends of it. Its 2020 white paper called terahertz inevitable and set a peak rate of 1,000 gigabits a second. In February 2026 the same company called 7 GHz X-MIMO a core 6G technology and reported a peak of 3 gigabits a second outdoors with KT in Seoul. The draft minimum requirement in part 3 is 36 gigabits a second in 600 MHz, a twenty-eighth of the 2020 figure. No vendor has said it was wrong about terahertz; the 2026 papers just leave it out. For a carrier, one claim matters more than any peak rate. Ericsson’s head of networks says 6 and 7 GHz can be matched in the downlink from the existing 3.5 GHz grid, and no operator has confirmed it.

• For the band. All seven equipment and chip vendors; GSMA, for up to three times today’s spectrum; NGMN, for 6 to 15 GHz. The European regulators’ group names 540 MHz of upper 6 GHz as the primary 6G band by 2030, and the American executive branch backs 7.125 to 7.4 GHz.
• Against it, or sharing it. Apple, Broadcom, the cable industry and the Wi-Fi industry on 6 and 7 GHz, with NCTA citing Wi-Fi “carrying nearly 90% of mobile data”. Ofcom shares upper 6 GHz between Wi-Fi and mobile and supports only 7.125 to 7.25 GHz at WRC-27. Aviation opposes 4.4 to 4.8 GHz.
• Paying today. Nobody, for 7 GHz. For mid-band capacity in general, 8.45 million T-Mobile 5G broadband customers and 6.2 million Verizon fixed wireless subscribers pay $35 to $70 a month.

The first 6G band costs 6 dB and a thousand-element antenna, a price the industry can pay where 32 dB was not. Whether it is cheap depends on one unconfirmed claim, that it fits the existing site grid. Whether it is available depends on regulators who have given the same frequencies to Wi-Fi in the United States and split them in Britain.

Three things are called AI, and the standard needs none of them

The phrase AI-native runs three different things together. One uses a trained model to run part of the radio better. The second puts half a model in the phone and half in the network, and the third replaces the radio’s computer with a graphics processor. They have different owners, different standards and different critics (Exhibit 5).

The first is settled and small. Release 19’s work item, RP-240774, made normative a framework for one-sided models, which live wholly in the phone or the network. The second is the hard one. A phone describes the radio channel to the network in compressed form. A two-sided model puts the compressor in the phone and the decompressor in the base station, and the two halves have to be trained together. So a phone from one company and a base station from another must share a trained model. The Release 20 work item, RP-251870, lists three ways: a fully specified reference model, a standardised encoder structure with parameters exchanged, or a standardised dataset format. The third, running the base station’s signal processing on graphics processors, is not a 3GPP matter.

The 6G study’s own decision is a sentence the marketing never quotes: the 6G radio and RAN design shall ensure that the 6G system can also operate without AI/ML. Which is another way of saying that AI is optional by design, and that “AI-native” describes products built on it. That has not stopped anyone. Nokia Bell Labs says a learning radio could set up bespoke waveforms, constellations and pilot signals. Its trial with NTT DOCOMO and SK Telecom at 4.8 GHz in November 2025 reported throughput gains of up to 100%. Qualcomm says 6G will be the first AI-native wireless system. Ericsson’s researchers say a non-AI fallback will always be there and that one-sided uses should come first.

The third dispute is about money and microseconds. NVIDIA paid $1.0 billion for 2.9% of Nokia at $6.01 a share, and Nokia claims 1.5 times the capacity and twice the spectral efficiency on its own testing, with a commercial release in 2027. Ericsson says GPUs are one option for 6G and AI RAN and not the only one. Its technical objection is that a million-parameter model cannot run in the 500 microseconds a base station has to process each slot. Samsung does not believe GPUs are a prerequisite for 6G. Verizon’s chief technology officer said that if you divide performance by cost, one day it makes sense, and today it does not. Light Reading reports one GPU card at 300 W against one custom accelerator at 40 W, better than seven times the power.

The standard has answered the question the slogans ask. 6G must work without AI, so AI-native is a property of products rather than of the specification. The decision we would watch is the two-sided model, because 3GPP’s three routes are three answers to who pays for the training and who owns the result. The GPU argument lies outside the standard and inside the capex line of part 4.

Energy is the one condition every operator has put on buying anything else

Energy is a condition of sale rather than a feature a customer sees. A network spends most of its life nearly idle, and the saving on offer is in what it does then (Exhibit 6).

The operators have said the same thing for three years. NGMN’s 2023 position statement asks for absolute energy reduction, and Orange lists energy efficiency as a core requirement of 6G. The filings explain why. Vodafone used 6,006 GWh in its 2025 financial year, and 94% of it went on networks and technology centres.

Energy is the only one of the eight on which operators and vendors say the same thing, and the only one on which the standard has fixed a method rather than a figure. The catch is that the vendors’ savings are savings on capacity a 5G site is not using, which at 6% utilisation is most of it. None of them says what a loaded 6G site will draw.

The migration decides the capex, and December 2026 decides the migration

A 5G phone, a 5G site and a 5G core have to reach 6G without a second network, and how they get there is the most argued question in 3GPP this year. The two plenaries of 7 to 11 December 2026 are where it comes to a head, and they are the first hinge on the watch-list in part 7 (Exhibit 7).

The decided part is that 6G stands alone. TR 38.914 requires a standalone RAN architecture and, in the baseline, no aggregation of any kind between 5G and 6G, the opposite of how 5G launched on its 4G anchor. What makes standalone affordable is multi-RAT spectrum sharing, MRSS. One carrier in a band the operator holds is served to 5G and 6G phones from one radio, so 6G can start before any new band is bought. RAN#113 added that the two must share the downlink control channel efficiently, because two separate control channels in a 10 MHz carrier cost over 14% and up to 23% in overhead.

The open part is whether anything beyond MRSS is allowed. The options are a 6G-anchored dual connectivity that adds a 5G leg, a 5G-anchored one, and a dual registration that splits 5G voice from 6G data. RAN#113 postponed the question to RAN#114 on 7 to 11 December 2026 and asked the architecture group to rule on dual registration by SA#114 the same week. The core has its own December. The SA plenary set SA2#178 in November as the last date for an architecture decision, and its chair minuted that the timeline would be in danger without one.

This is where the operators have written the most and agreed the most. NGMN’s 2023 statement said 6G must not inherently trigger a hardware refresh of 5G RAN infrastructure. Vodafone’s Luke Ibbetson put it more plainly: the fundamental rationale is to avoid the need to go and replace radio infrastructure. At RAN#113 Deutsche Telekom withdrew from a twelve-operator paper on the alternatives, on the grounds that none of the options are needed.

• No forced hardware refresh in existing bands. NGMN in 2023, 2025 and 2026; Vodafone, AT&T and Verizon. Ericsson says 6G will be software-upgradable on certain hardware, which it has not named; Nokia offers upgrades on new GPU hardware.
• MRSS as the only baseline. NGMN in June and September 2026; twelve operators at RAN#113. 3GPP has made it the baseline and left the alternatives open. Ericsson calls MRSS “highly efficient” and multi-vendor MRSS on one radio “complex to realize in practise”.
• An evolved core, with connectivity independent of AI agents. Seventeen operators at SA#113. 3GPP assumes the 5G core framework as the starting point, has three directions on the table and a decision due in November 2026. Ericsson wants an evolution of the 5G core, Nokia an AI-aware core, Huawei and ZTE an agentic core.
• A business case before capex. NGMN in 2025 and 2026; Orange; AT&T. This is not 3GPP’s to give. The vendors’ answer is the 5G record: 3.3 billion subscriptions, T-Mobile’s postpaid phone ARPU up from $46.04 to $50.37 in six years, AT&T’s down 0.3% in the last year.

The vendors agree in public and hedge in the detail. Ericsson’s chief executive calls 6G an evolution of 5G, while its RAN blog describes a standalone deployment without any form of aggregation. Qualcomm’s representative told MWC that 6G is not an evolution, it is a revolution. Why the hedge? Ericsson’s chief executive conceded, as reported, that softwarisation may structurally hurt the top line, and that sentence explains it.

What did not make the list

Some of what the early roadmaps promised has gone. Reconfigurable intelligent surfaces were never proposed for the 6G study, sub-terahertz is out of scope above 52.6 GHz, and ambient IoT carries on in 5G-Advanced. Immersive communication is in the report as a composite target and in every vendor’s paper as a scene with glasses in it, but no operator has put it forward as the business case. Meta’s numbers suggest why: $2.2 billion of Reality Labs revenue against a $19.2 billion operating loss for 2025.

In 5G every feature was supplied by the equipment vendor, licensed by the regulator and sold by the carrier. In 6G the three capabilities furthest along each depend on a party outside that triangle, and each of those parties can take the value without the carrier. The capabilities that stay inside the triangle are the ones the operators are fighting for in the standard. They are winning the written part of that fight. Whether they win the hardware part is a December decision.

The terms, briefly

• NTN. Non-terrestrial network: 3GPP’s term for satellites and high-altitude platforms in the mobile standard. Transparent means the satellite relays to a base station on the ground; regenerative means the base station is on board.
• ISAC. Integrated sensing and communication: using the network’s own transmissions as radar. Monostatic sensing listens for its own echo; bistatic uses a second site as the receiver.
• Upper mid-band, FR3, centimetre wave. Industry names for roughly 7 to 24 GHz; 3GPP refers to “around 7 GHz”. Giga-MIMO, X-MIMO and GigaMIMO are vendors’ names for antennas of a thousand or more elements built for it.
• MRSS. Multi-RAT spectrum sharing: one carrier in one band served to 5G and 6G phones from one radio. The operators’ chosen baseline.
• Two-sided model. A trained model split between the phone (encoder) and the base station (decoder), whose halves must be trained together. Interoperability between companies is the open question.
• Standalone. A 6G network with its own core and no 5G anchor. 5G launched the other way, riding on 4G.