00:50 - Dean has written a report on satellite D2D, and has also followed the US International 6G Call-to-Action (link)- the new initiative by the NTIA to build a coalition for 6G technology security, harmonisation and supply-chain control (Link to Dean's report on Satellite D2D for STL)

4:55 - Geopolitics and sovereignty for subsea cable, 6G, spectrum and more

9:03 - Benoit discusses the FCC's changing definitions of "broadband" and both of us discuss the usefulness of "Gigabit" as a target, and whether changing it is a nod to the satellite industry. (Link to Benoit's post on FCC + Gigabit) (Link to Dean's 2024 article on broadband definitions & satellite)

17:05 - Dean has a random question - could there be a possible wholesale model for satellite broadband, especially if retail competition is lacking?

19:01 - Benoit discusses the French regulator ARCEP's Interconnection Barometer report, and what says about IP Interconnect volumes in particular. Uplink to the cloud seems to be growing, but Dean wonders which subset of data is being collected and analysed. (Link to ARCEP report)

25:47 - Dean discusses the July 16th Indoor Wireless workshop and the implications for regulators, operators and neutral hosts (Link to Dean's Indoor Wireless writeup)

29:30 - Upcoming, Dean is speaking about AI Voice at the September Alianza Navigate 2026 event (link) and is running his own workshop in October (Link to Oct 29 Unthinkable Lab on AI Voice). 

32:00 - Benoit expects to be doing a lot of work around the next phase of the European DNA Digital Network Act policymaking discussion.

Book Ben & Dean for your event! Get us on stage at your conference, or facilitating at your private meeting or executive offsite. Connect & message us on Linked, or via deanbubley.com and fiberevolution.com

I’ve also set up a completely separate part of my Substack, with its own newsletter and sign-up, as I don’t want to distract from my main telecoms policy thinking and tech musings. The Post-Digital Substack homepage is here (with an intro and the first monthly newsletter) and a direct sign-up is here. Apologies that it’s slightly clunky for my existing Tech Musings subscribers.

Enable Post-Digital Futures emails

Each month, I'll publish updates on about 20 post-digital "signals" which can be a mix of company announcements, policy initiatives, R&D and academic research findings and other coverage at events or in the news. The details will be on the website, and I'll do a full newsletter here, plus any other articles such as event/document reports or deep-dives if relevant.

The main P-D site also has some basic explainers, an initial draft of "what this means for telecoms" and a list of interesting-looking events. Not sure if I'll get along to many of them in the short-term, but I definitely want more direct exposure over time.

Note: This is a work-in-progress, and I must disclose upfront that it's a hybrid of my own input, layered onto a lot of curation and background research from AI. Hopefully it doesn't end up in the "slop" category - especially as a fair amount of it is a learning experience for me as well.

I can't claim to be a deep expert in some of the physics, maths and materials science behind "reservoir computing" or some of the cybersecurity aspects of quantum.

But as far as I know, nobody else is attempting to gather together all of these technologies and assess what they could mean for telecoms and networks.

Please reach out to me if you'd like to discuss or disagree with any of what I cover, or if you'd like to suggest events or other engagements. I'm looking forward to all of this evolving - and as an opportunity to get my head around something that not just another version of 180-year digital communications, or 80-year old Von Neumann architecture.

If you want to subscribe to my new Post-Digital Future newsletter, there is a separate sign-up via this page https://deanbubley.substack.com/s/post-digital-futures.

Computing is escaping the computer.

Information is increasingly being processed using light, magnetic states, memory arrays, radio waves, quantum effects, ionic movement and even living biological systems. These technologies are also taking roles inside network, device and compute hardware.

Some of these approaches will become commercially important (a few already are). Others will remain scientific curiosities, specialist components or catalysts for misleading hype.

The central near-term development won’t be the outright replacement of silicon, conventional processors or binary “digital” logic. Instead, it is the appearance of unusual forms of computation or sensing, inside otherwise familiar systems such as:

• Specialist accelerators

• Optical links and switches

• Radio networks and elements

• Sensors and edge devices

• Memory arrays

• Network operational tools

• Conventional cloud and software platforms

• IoT elements and other endpoints

Rather than substitution of existing technologies, most of the innovations being tracked will manifest initially as hybrids or sub-systems, in order to work around existing lock-in and path-dependency effects. Some - notably photonics - are already beyond that point, however.

What “post-digital” means

I use “post-digital” as an umbrella term for computing, sensing, communication and information-processing technologies that do not fit neatly within a conventional binary 1’s and 0’s, electronic-based computing paradigm.

As many regular readers will know, I have long disliked the lazy use of “digital” as a synonym for new, advanced or innovative. Morse’s telegraph, invented in 1838, was digital. Arguably, so were manual systems such as abacuses, which literally involved digits. Terms like “digital transformation” or “digital infrastructure” make me cringe.

In terms of modern compute and networking, “digital” is broadly associated with what are called von Neumann architectures, that separate processing and memory, although the full distinction is a bit more subtle. In those systems, the best way to get data from the processor to memory and back is usually to send it as bits or binary pulses, and process it with transistors as switches. If you bring compute and memory together, quite a lot changes, including the fundamental ways that calculations and analysis occur.

That said, most credible developments will complement mainstream processors, or sensors rather than replace them. CPUs, GPUs, ASICS for networks, and conventional software will orchestrate specialist new components, prepare their inputs and interpret their outputs. In many cases, users may never know whether a particular calculation involved quantum annealing, optical processing, in-memory computation or another unusual mechanism, either directly or in an intermediate network element.

None of the means that digital computing or communications is dying or fully obsolete. Digital remains important. But it is certainly not new, or especially cool. Having a “Chief Digital Officer” is like having an “Chief Electricity Officer”.

…. For more, and for an update on the August “signals”, carry on reading here.

Already subscribed to Disruptive Analysis? Open your subscription settings, select Disruptive Analysis and enable Post-Digital Futures under newsletters.

Enable Post-Digital Futures emails

The main conclusion is that unusual forms of computation are starting to enter otherwise familiar systems: operational tools, packaged accelerators, edge devices, optical links, sensors and conventional software.

There are no near- or even medium-term replacements of traditional technology approaches. The probable commercial trajectory involves specialist acceleration and myriad types of hybrids.

Five developments that matter

1. Quantum computing enters telecoms, but through an operational side door

Signal 021 covers AT&T’s expanded work with D-Wave.

AT&T reported that quantum annealing reduced one specific network-optimisation workload from roughly an hour to less than 15 seconds. The planned applications include outage response, technician routing, network-build planning and traffic management.

Quantum annealing is a specialist approach for searching among large numbers of possible combinations. It is not the same as running general-purpose software on a universal quantum computer - and indeed, various classical approaches to annealing-type optimisation have existed for many years. (I remember a UK company called Parc Technologies I knew many years ago, which created traffic engineering software based on annealing techniques. It was acquired by Cisco)

An interesting aspect of this announcement is the proposed integration with AT&T’s existing agentic-AI and operational tools. If quantum processing becomes commercially useful, users may not encounter a recognisably “quantum” product. They may simply invoke a specialist optimisation capability from familiar software, with the quantum aspects “behind the scenes”.

However, the announcement leaves important questions unanswered:

• How large and difficult was the optimisation problem tested?

• What classical alternatives were also tested and how did they compare?

• Did the result influence a live network or only a planning model?

• How much of the improvement came from modelling, workflow changes or quantum processing itself?

Overall, this is certainly more credible than a generic quantum-network vision, but it is not yet evidence of quantum advantage in live telecom operations.

The wider quantum cluster of signals also moved closer to operational questions, reflecting the relative nearness to adoption of a range of quantum services. Signals 022–025 examine hollow-core fibre for quantum teleportation, routing and spectrum allocation for entanglement networks, a proposed use of IPv6 extension headers for future quantum routing, and a hidden optical impairment termed “hyperloss”.

The last of these may prove especially relevant. Quantum links could require new forms of path qualification, fault localisation, instrumentation and service assurance, because ordinary optical power measurements may not reveal the impairments that destroy quantum correlations and entanglement.

Overall, that points towards a possible future category of quantum-network OSS and assurance, long before there is a mass-market quantum internet.

2. Photonic computing is becoming a systems question

Photonics produced another large cluster in this briefing.

The strongest near-term commercial signal is from Signal 027, with an expected $300 million US government award to GlobalFoundries for silicon-photonics manufacturing, advanced packaging and near-packaged or co-packaged optical links.

This is highly significant as it concerns manufacturing capacity, materials, bonding, packaging and the route towards volume production. (A further US indicator is the new National Security Science and Technology Strategy, which includes advanced computing areas aligned with this newsletter, and which will be reviewed in the next edition).

Co-packaged optics (CPO) moves optical engines close to switches and processors, reducing the length and power consumption of very high-speed electrical connections. Near-packaged optics (NPO) provides a less integrated and potentially more serviceable intermediate step. I’m writing about these for an upcoming STL Partners report.

The first CPO/NPO deployments will be driven by AI clusters and large datacentres, not mobile backhaul or access networks. Nevertheless, the technology could eventually affect high-capacity routers, metro aggregation, telco edge datacentres and upgraded central offices.

Further photonic signals illustrate both progress and the need for caution:

• Signal 028 proposes programmable optical delay and synchronisation, but remains simulation-led.

• Signal 029 demonstrates a photonic accelerator using a non-volatile ferroelectric material, potentially reducing continuous tuning power.

• Signal 030 demonstrates a reconfigurable two-bit optical arithmetic unit at a 20 Gbit/s input rate.

• Signal 031 models a small photonic neural network using components familiar from optical communications.

• Signal 032 packages a photonic matrix with converters, amplifiers and an FPGA to compress images before storage or transmission.

Signal 032 is particularly interesting because it demonstrates a more complete system. The prototype applied photonic compressive sensing to images, with adjustable compression ratios from 2× to 256×.

Potential applications include satellites, drones, industrial cameras, medical equipment and other systems that generate large images but face limits on power, storage or communications. I’ve written before about the impact of advanced compression techniques on data traffic volumes, although I primarily focused on semantic techniques, rather than direct and novel hardware approaches.

The new proposed system is neither a universal replacement for conventional codecs nor the AI-based compression techniques I expect to emerge. Reconstruction is approximate, image quality falls at aggressive compression ratios and part of the workload remains electronic. Comparisons with a general-purpose GPU also flatter specialist hardware.

Nevertheless, this is still more useful than quoting the performance of an isolated optical component while ignoring the rest of the system.

A recurring test for photonic computing should be:

What remains of the speed or energy advantage after lasers, data conversion, memory, calibration, control electronics, packaging and thermal management are included?

3. Neuromorphic and in-memory computing are developing more credible routes to commercialisation

Signal 033 covers EuroNMC, a European neuromorphic collaboration involving 139 researchers, more than 80 institutions and 24 countries. While research networks cannot directly address shortages of scale-up capital, fabrication capacity or anchor customers, they perform useful roles in convening fragmented ecosystems and help raise both awareness and funding.

Neuromorphic computing currently crosses too many layers to develop just through isolated research groups. It requires interaction among materials science, devices, chips, algorithms, software, benchmarks and applications. Shared language and credible benchmarks matter, because “brain-inspired” is often used far more loosely than the evidence justifies.

There is also a more specific telecom connection. Ericsson and Forschungszentrum Jülich have announced work on advanced AI for future mobile networks, using the JUPITER supercomputer and investigating energy-efficient, neuro-inspired approaches. Possible areas include radio-channel estimation and massive-MIMO processing. I posted about this on LinkedIn recently.

It would be wrong to claim that “neuromorphic base stations” are imminent. The most plausible route is via a specialist co-processor for bounded, repetitive and energy-sensitive workloads. This would also align with Ericsson’s general view of optimising RAN compute hardware, rather than just relying on AI-RAN style general-purpose GPUs everywhere.

Signal 034 provides a complementary in-memory-computing example.

Researchers from TetraMem, SK Hynix and the University of Southern California fabricated a system-on-chip containing ten memristor-based neural-processing units and a small RISC-V control processor. It ran a compact visual wake-word model and reported an energy efficiency of 21.3 trillion operations per second (TOPS) per watt.

The significance is less about the peak rate, and more the fact it was a complete chip and addressed “depthwise convolution”, an operation used in lightweight edge models, but poorly suited to many conventional in-memory arrays.

The potential applications are mostly in narrow, power-constrained inference tasks in cameras, drones, gateways and industrial sensors. While not directly built into a telco-scale network, that could alter and significantly reduce the traffic generated by some edge systems, if they transmit events, anomalies or compact interpretations instead of continuous raw data. (That effect is not unique to in-memory computing. Conventional accelerators can also support local processing).

4. The edge impact is about traffic topology, not simply more or less traffic

Several signals point towards more processing capabilities located close to (or embedded directly in) sensors and data sources, which is sometimes called the “far edge”. This includes photonic image compression, in-memory edge inference, fluidic sensing and locally adaptable spintronic devices - alongside other ongoing developments for on-device AI and processing.

A simplistic telecom conclusion could be that local processing reduces total network traffic. But while that may happen in some applications, it is not a reliable general rule. In some instances, this is more reflective of either power constraints, concerns over data security, or simply unreliable networks or intermittent coverage.

Taking these evolutions together, edge system might transmit fewer raw measurements but generate:

• More frequent events and alerts

• A need for sizeable model updates or other software downloads

• New forms of control-loop traffic

• High-value data requiring stronger assurance and error-checking

• New telemetry streams about the accelerator or sensor itself

• Occasional bursts of full-resolution information, for calibration or urgency

• Data replicated across cloud, edge and sovereign environments

This means that new orchestration models will be required which take into account the new computing pathways and capabilities, as well as where information originates, where it is interpreted, which network domains it transits and what network guarantees or feedback loops are required.

5. Biological computing may reach the market by leaving the biology behind

Signal 039 covers The Biological Computing Co., which says it studies living neuron cultures, translating the resulting techniques into ordinary software. Customers run the data outputs on conventional GPUs and cloud infrastructure, rather than maintaining living neurons in a specialised environment.

This is a similar model to some quantum computing-aaS approaches, where the end user accesses the novel platform via cloud APIs, rather than needing to maintain their own on-prem QC system, with cryogenics and other supporting paraphernalia.

The biocomputing company reported improvements to its video-generation model, including roughly 4.4 times lower inference cost and more coherent video.

This suggests that biological experiments might be used as a research instrument for discovering algorithms, while the deployed product remains conventional software. That avoids many of the immediate problems associated with biological hardware: contamination such as limited lifetime, specialist laboratory skills, ethical / regulatory considerations and integration with datacentres or remote edge sites.

We shouldn’t mislabel this. Biology-derived software is not really a biological computer, and we certainly cannot claim that neurons are somehow going to replace GPUs.

Signal 040 addresses a different constraint which illustrates some of the ethical issues with biocomputing: tissue consent and provenance. Donated biological material may have been collected for medical research, without explicit permission for allowing commercial computing, defence applications or AI model development.

If biological computing becomes valuable in future, the origins and permitted use of a given cell line may need to become part of the technology supply chain, especially in countries with more conservative licensing rules. Governance might need to cover consent, provenance, ownership, jurisdiction, handling and eventual disposal.

This may turn out to be a pre-requisite for the concept more broadly, and will need to be promptly addressed.

Further-future signals are intriguing

This briefing and the main P-D website’s selection of signals includes space for some developments that are still in labs, and which are not remotely close to market. These include:

• A 400-particle liquid reservoir computer that detected anomalies

• A fluidic memristor combining ionic movement, memory and processing

• Spintronic neurons trained around real hardware variation

• A nanoscale device switching between distinct magnetic phases

These all represent tests of whether computation can emerge from collective physical dynamics, material states or the behaviour of imperfect devices. These are potentially parts of future mechanisms to reduce energy involved in compute and sensing.

However, only a few such experiments may ever get commercialised. Just because something works at small scales in ideal situations, does not mean that full programmability and assurance are realistic. (Much can be said about various telecom technologies too, even if they go through standardisation).

Future analogue, neuromorphic and physical systems will be judged not only by theoretical efficiency, but by whether they can be calibrated, tested, monitored, debugged and manufactured consistently. That said, keeping an eye on more sci-fi style developments is worthwhile for spotting anything that could prove truly disruptive in due course.

Two points evolving from the previous briefing

The July signals briefing, covering developments from early June, was dominated by quantum-safe security. (Note: that briefing has not been fully vetted, edited and approved by me at the time of this publication. It was added as an extra initial set of data for the website)

The key conclusion was that operators should act on post-quantum cryptography migration, without allowing QKD trials and quantum-internet language to inflate near-term infrastructure expectations. This is something that I have also covered in my telecoms work, including on a panel at last year’s Techritory conference in Riga. (See my write-up here and a video of the session here)

A second highlight and very telecom-relevant example was Signal 012, in which BrainChip applied neuromorphic processing to RF signal classification. This remains one of the clearer routes by which unconventional computing could enter telecom-adjacent systems: bounded, always-on classification close to a sensor or radio. This is especially relevant for domains such as dynamic spectrum-sharing, where it may be important to recognise - and manage - heterogeneous signal types in the same band (for instance cellular signals and vehicle/aviation radar).

Signal 017, which involved the broadcast of neural-network weights over a MU-MIMO radio system, also showed that radio waves can sometimes become part of the computation, rather than merely carrying its inputs and outputs. I’ve also written about this general idea here.

These are still experimental ideas. But they connect post-digital computing with more familiar network functions, rather more directly than many speculative quantum-network proposals.

What should telecom decision-makers watch?

My current shortlist of qualifying questions for telco and vendor CTO offices, or perhaps internal VC and academic-collaboration functions include:

1. Operational integration: Can an unconventional accelerator enter an existing workflow through familiar software and APIs? Are there hybrid pathways that can overcome existing path-dependencies and lock-in effects?

2. Whole-system efficiency: Do the claimed gains still apply in the context of wider systems, which may have various conversions back-and-forth, memory needs, calibration, energy and cooling, security and control and so on?

3. Manufacturing and supply chains: Is there a realistic route from a device demonstration in a lab, to repeatable fabrication, deployment and operation? Are there any obvious bottlenecks in raw materials, regulatory constraints on trade, material-handling etc? (Most places won’t have supplies of liquid nitrogen or helium, for example - or clean rooms)

4. Traffic topology: Does processing move data among devices locally, and around or to/from access nodes? Where would edge locations, cores and datacentres be?

5. Assurance: What new monitoring, testing, fault-management and service mechanisms and metrics are required?

6. Sovereignty and governance: What are the regulatory and policy implications? And who / where controls elements such as foundries, materials, software, biological inputs, intellectual property and specialist skills?

7. Evidence of advantage: Is the technology being compared with the strongest credible conventional alternative, or with a convenient baseline? What are the counterfactual arguments - and what changes between now and any commercial launch?

What comes next

The September briefing will cover developments published during August.

Enable Post-Digital Futures emails

Already subscribed to Disruptive Analysis? Open your subscription settings, select Disruptive Analysis and enable Post-Digital Futures under newsletters.

One item already on the watchlist is The Economist’s briefing on neuromorphic and biological computing. The article itself is not proof of technical progress, but mainstream coverage is a useful signal that post-digital ideas are moving beyond specialist research communities. Another will be the implications of the US National Security Science and Technology Strategy.

There may also be separate Substack notes, or quick posts on my LinkedIn.

The complete August briefing, with all twenty signal cards and original sources, is available at postdigitalfutures.com. You can also browse the Post-Digital Primers or read the evolving assessment of what these technologies could mean for telecoms.

Disclosure: Content is human-edited but mostly AI-assisted. This project is experimental. Double-check facts, links, quotes and interpretations.

Dean Bubley
Disruptive Analysis

Computing is no longer confined to conventional processors, separate memory and familiar “digital” binary electronic logic.

Information is increasingly being processed using light, magnetic states, memory arrays, radio waves, quantum effects, ionic movement and even living biological systems. Some of these approaches may become commercially important. Others will remain scientific curiosities, specialist components or sources of misleading hype.

Post-Digital Futures is my attempt to work out which is which.

I use “post-digital” as an umbrella for emerging computing and information-processing technologies that do not fit neatly within conventional binary, electronic and von Neumann assumptions. It does not mean that digital computing is ending. Most credible developments will complement mainstream processors rather than replace them.

The important near-term story is likely to be hybridisation: specialist accelerators, unusual sensors, new memory technologies and physical processing elements inserted into otherwise conventional systems.

A disclosure: I have long had a dislike for the word “digital”, especially in the context of terms like “digital transformation” or “digital marketing”, or where policymakers refer to “digital infrastructure”. It’s worth remembering that Morse’s telegraph, invented in 1838, was digital. So, arguably, were manual compute systems like abacuses, which literally used digits (fingers) as part of the mechanism.

Digital might be important, but it’s certainly not new, cool, or innovative. Post-digital is.

Enable Post-Digital Futures emails

Already subscribed to Disruptive Analysis? Open your subscription settings, select Disruptive Analysis and enable Post-Digital Futures under newsletters.

What this section covers

The main technology categories include:

• Quantum computing, optimisation, communications and sensing

• Neuromorphic and event-driven computing

• In-memory and analog computing

• Photonic and optical processing

• Spintronics and magnonics

• Physical, ionic and reservoir computing

• Biological and biology-derived computing

• New technologies that do not yet fit comfortably into any established category

The emphasis is not just on faster chips. I am interested in where information is generated, stored, processed and transmitted, and how new computing models could change systems, networks, energy consumption, supply chains and business models.

Why telecoms should care

Telecom operators are unlikely to become major vendors or users of quantum computers or biological processors. That is the wrong test of relevance.

The more useful questions include:

• Could a technology change where traffic originates, terminates or is processed?

• Does it create new requirements for latency, synchronisation, assurance or sovereignty?

• Could it alter the balance between devices, access networks, edge sites, cores and datacentres?

• Will it appear inside radios, routers, sensors, optical systems or OSS / BSS software?

• Does it create new demand for specialist connectivity, infrastructure or managed services?

• Could telecom assets, skills or customer relationships provide useful control points?

• What must operators understand before the technology appears in vendor roadmaps or policy discussions?

The answers will vary enormously by technology and timescale. Some implications are immediate. Others belong in five-, ten- or fifteen-year scenario planning.

What you will receive

The monthly Post-Digital Futures briefing normally contains up to 20 worthwhile signals. These are numbered sequentially so they can be discussed and revisited without confusion.

Each signal includes:

• A plain-language explanation of what happened

• The original source and publication date

• An assessment of maturity and commercial timing

• A telecom relevance rating

• A critical interpretation of what the development does and does not demonstrate

The accompanying website also contains Post-Digital Primers and a regularly updated assessment of telecom relevance.

I will also occasionally cover important conferences, policy developments, corporate activity and signs that a previously obscure technology is entering wider industry awareness.

The editorial approach

This project is designed to find consequential developments, not to fill a quota with incremental academic papers or routine announcements.

I am particularly interested in awkward evidence:

• A negative experimental result that reveals a hidden deployment problem

• A technology that works, but only when important system costs are excluded

• A company entering an unexpected part of the value chain

• A supposedly futuristic technology appearing inside a mundane operational workflow

• A governance, manufacturing or assurance constraint that could matter more than raw performance

I am particularly interested in evidence that shows where these technologies are becoming relevant, investable or strategically important:

• A telecom operator, network vendor or infrastructure provider collaborating with an emerging-computing company

• A government programme, public funding / procurement decision or national strategy that signals serious commitment

• An established technology company incorporating an unconventional computing approach into a product, platform or roadmap

• An investor, corporate venture arm or major funding round indicating that a technology is attracting greater commercial attention

• A supposedly futuristic technology appearing inside a mundane operational workflow

• A promising use-case that could change how networks, datacentres, devices or edge systems are designed

• A negative experimental result or deployment constraint that reveals what must improve before wider adoption

• A technology that works, but only when important system costs are excluded

• A governance, manufacturing or assurance constraint that could matter as much as raw performance

• A development that changes an important assumption without yet creating a product

• Broader coverage of post-digital tech in mainstream publications or events

Claims from companies, universities and governments are not treated as proof merely because they appear in a press release. At the same time, partnerships, investment, procurement and sustained institutional attention are meaningful signals even when they do not yet demonstrate technical or commercial success. An early experiment should not be dismissed simply because it is far from commercialisation, and a strategic announcement should not be treated as validation merely because it comes from a prominent organisation. Weak signals can be useful if their significance and limitations are made explicit.

A disclosure about AI

Appropriately for the theme, much of the content on Post-Digital Futures is human-edited but AI-generated or -assisted. The main site and this newsletter are experimental. Facts, links, quotes and interpretations should be double-checked before use or citation.

AI helps with scanning, classification, initial explanation and drafting. I review and edit the material, and signal-level interpretations are not presented as my own take until I have approved or rewritten them.

I currently use ChatGPT Pro with 5.6 Sol, with a significant history of relevant context, as I have been privately running a post-digital project, research, chat and analysis stream for nearly a year, alongside my other tech and telecom work. I also periodically cross-check against other models, as well as my human contacts and friends.

Who this is for

Post-Digital Futures is primarily intended for people working across:

• Telecom networks and strategy

• Semiconductors and computing infrastructure

• Cloud, edge and AI systems

• Technology policy and regulation

• Investment, corporate development and innovation

• Research commercialisation

• Long-term scenario planning

Enable Post-Digital Futures emails

Already subscribed to Disruptive Analysis? Open your subscription settings, select Disruptive Analysis and enable Post-Digital Futures under newsletters.

You do not need to be a physicist or chip designer. The objective is to make unfamiliar technologies understandable without pretending that difficult concepts are simple or settled.

This work is deliberately telecom-focused at the moment - I am not a mainstream semiconductor analyst, but I do understand networks, both wireless and fixed. Over time I’ll likely expand my own knowledge to apply it more broadly.

You can explore the full project at postdigitalfutures.com.

If you have relevant research, a company announcement, a sceptical correction or an application I have overlooked, reply to a newsletter or contact me through the site.

Dean Bubley
Disruptive Analysis

I’ve commented on some of these through the year, especially on topics such as IP interconnect, “voluntary conciliation” for ecosystem disputes and greater harmonisation of spectrum management. Previously I have discussed the deeply flawed “fair share” concepts, which thankfully didn’t make it into the DNA.

But there is also input from other sectors not normally represented in telecom regulatory and policy discussions. While the DNA is intended to be a modernisation and “simplification” of existing telecoms rules such EECC, aiming to drive more pan-European scale and growth, in reality there are various contentious elements - and a risk of harmful spillover into adjacent markets.

Subscribe now

One of those areas is automotive, which may be “captured” by some of the new policies and regulatory proposals. Clearly, this is also a sector with deep regulation of its own, giving plenty of scope for conflict.

The European Automobile Manufacturers’ Association, or ACEA, represents Europe’s major car, truck, van, and bus makers. There are 17 members, including firms like BMW, Renault, Mercedes, DAF, VW and Stellantis, as well as international companies with large European presence, such as Ford, Toyota and JLR. Notably, it does not include automotive component manufacturers such as Bosch, only the vehicle brands.

ACEA’s paper focuses on connected vehicles. It has numerous interesting points relevant to that set of use-cases, but also exposes some important European Commission category errors, false binaries and the growing risk of unintended consequences.

As context, a lot of the discussion around networks seems to descend into a polarised battle between European large telcos and US cloud/content providers. Some EU officials and member states’ politicians want to push Deutsche Telekom, Orange, Telefonica, Vodafone and others as European “champions” against the likes of Amazon, Google, Apple, Meta and Microsoft.

This has previously led to farcical discussions about so-called “fair share” payments, as well as many of the controversial elements of DMA and DSA such as regulating “gatekeepers” and items such as messaging interoperability (which I also think is deeply stupid). More recently it has cropped up again in sovereignty debates, flawed thinking about edge computing, and the desire to drive “scale” and consolidation in European telecoms, even if that harms competition and smaller / niche providers.

The weight of the automotive industry - and its role in creating, operating and selling networks - changes the binary framing.

To be clear, as a lobbying document, ACEA’s paper also obviously has its own agenda and objectives, some of which seem to transfer costs or responsibilities from vehicle manufacturers (“OEMs”) to telecom operators.

Some of its technical arguments treat current vehicle architecture as though it were a fixed law of nature which cannot evolve. To be fair, the typical timelines for new model design, type-approval and product lifetime make immediate changes very difficult in practice - especially where components are used in safety-critical functions, such as the European eCall emergency system.

There are also some surprising omissions in the report, although it is fairly short overall. It does not reference cloud, AI, or the frequent use of private networks in the automotive sector.

Share

The DNA makes a serious category error

Connected cars represent a useful stress-test for the EU DNA, because they don't fit neatly into categories that Brussels' telecom policy folk tend to assume apply. That is especially awkward here, as in many ways, vehicle manufacture is a stronger and more strategic industrial sector than telecoms for Europe, and many politicians pay attention when it speaks up.

Reading the ACEA document and extrapolating to the broader market, you start to realise how badly the “telco vs. tech-co” argument is flawed. It’s a false binary - essentially a category error and what I call an “assumption of uniqueness”.

In fact, a vehicle OEM may simultaneously take the roles of:

• Wholesale-connectivity customer

• ISP

• M2M service provider

• Emergency call operator

• Application developer

• Controller of embedded eSIMs

• A cloud and data-platform operator

• Road toll / fee collector

• A customer of CDNs

• In some cases an MVNO

• Perhaps a satellite connectivity endpoint

In addition, the same physical connection might carry eCall, telemetry, cybersecurity updates, infotainment, remote diagnostics, theft prevention, cooperative-driving applications, passenger broadband and maybe other functions, all with very different levels of criticality. And indeed, the vehicle may well have two or more separate cellular modems, plus Bluetooth, Wi-Fi or USB connecting to internal devices such as the driver’s phone.

Automotive OEMs also operate private networks in factories, test facilities and logistics sites, maintain extensive connections to dealerships and EV charging networks, run huge compute systems for R&D and production, and some even have direct traffic exchanges and interconnection at peering points and IXPs - a point which I’ll come back to later in this article.

Table: Sample of automotive industry direct IP interconnects at Internet Exchanges

^{Source: ChatGPT from ASN data / PeeringDB}

In other words, car and truck manufacturers’ network relevance cannot be squeezed into a single regulatory category such as “electronic communications service”, “Internet access service” or “content and application provider”.

“Connected vehicle connectivity” is not a single use case. The paper demonstrates why the Brussels framing of Internet companies versus telcos is far too binary for the real multi-stakeholder universe.

ACEA’s recommendations and wishes

The ACEA position paper makes eight main recommendations. At a top level, it wants

• An exemption from switching rules for vehicle connectivity/ISP services, especially where critical and non-critical data shares the same subscription

• Protection against formal and implicit network fees and “fair share” attempts to charge developers / interconnected networks for mobile data traffic

• Clearer eligibility criteria for optimised QoS, slicing or “specialised services”

• More widespread, predictable and affordable 5G coverage

• Availability of alternatives networks, such as satellite connections

• Long and more predictable transition periods for the switch-off of legacy mobile networks, especially 2G / 3G

• Free callback communications for eCall. This is when an public safety call-centre calls back to a vehicle for more details or updates after an initial emergency call

• No mandatory FM or DAB+ radio requirements in new cars

ACEA provider-switching request highlights a problem

ACEA notes that the provider switching issue is partly covered in the DNA proposals under Recital 275, which recognises it may be impractical if an embedded SIM/eSIM supports both M2M functions (eg software updates) and consumer-facing Internet access. But the binding text does not currently create a clear exemption, and it is concerned that the OEM might need to duplicate expensive automotive-grade radios, antennas and software simply to separate these functions.

It’s fair enough to say that the driver should not have complete freedom to change connectivity provider for the vehicle’s cybersecurity systems, eCall support, theft/immobilisation or possible future EU driver sobriety or other safety functions. But equally I think an owner (or secondhand purchaser) really should be able to replace an expensive or poorly-functioning ISP service, and not be locked-in to an obsolete all-or-nothing contract for a vehicle’s 20 year lifespan. In particular, users should have choice to switch after any “free access” period expires - often at the same time as the vehicle warranty.

I can see multiple options here to separate these out - multiple eSIM profiles, separate APN, various capabilities embedded in connection manager software, easy tethering to a smartphone etc. I may cover the details another time.

But the point here raises a wider DNA lesson: The EU is likely to find that any poorly-framed horizontal rules attract dozens of sectoral carve-outs, many of them likely to be discovered post-hoc, as other industries may not have the ability of the huge automotive sector to get their needs expressed upfront like this.

Leave a comment

QoS, slicing, specialised services and interconnection

The section on “specialised services” which can offer guaranteed quality of service is very interesting, for multiple reasons. It also links to the document’s comments on network coverage. Key points:

• There is actually a commercial demand for QoS from some sectors. Despite decades of hype and supply-side push, this might be the first clear demand-side indicator I’ve seen of willingness-to-pay for QoS or slices.

• But, it’s only for some functions, some of the time, in some places. And as discussed above, vehicles have dozens of applications, with varying levels of importance / safety-criticality, bandwidth and latency demands, frequency of use etc

• Therefore there’s no such thing as a single “connected vehicle slice”. It’s an amalgam of many different things. No vehicle OEM is going to pay to prioritise everything, everywhere.

• ACEA quite reasonably wants regulatory clarity that vehicles should be identified as eligible users for safety-related connectivity prioritisation (eg for realtime hazard warnings, driver assistance and maybe autonomous driving)…

• … but that should not affect price or performance of normal “best efforts” connectivity for all the other applications.

• ACEA links network quality and coverage. The report essentially says “we’ll pay for this, but only if it works everywhere” before helpfully suggesting satellite connections as an alternative to terrestrial 5G.

• Not specifically discussed, but some of these functions might be provided by dedicated roadside V2X systems, rather than via a public MNO network

ACEA stresses that specialised services must not degrade the capacity and quality of the general-purpose Internet, so that MNOs then charge manufacturers (or others) to restore previous baseline performance. This is a common argument made for Net Neutrality more generally by Internet-centric companies, so it’s interesting to see the automotive industry taking the same line.

I’ve often said that “specialised services need to be genuinely special” and this seems to underscore that. But any such services need to have regulatory oversight to ensure normal users (or non-special applications) don’t end up having to pay more merely to avoid degradation.

But the real eye-opener is that ACEA makes it clear that quality and therefore network coverage is needed everywhere. If vehicles are using premium slices on public networks for critical functions such as safety alerts, inter-vehicle coordination for functions like “platooning” or critical software updates, there can’t be gaps in coverage.

It mentions rural areas and overall continuity, but it doesn’t specifically mention tunnels, parking structures or indoor reach - another omission, but maybe not huge in this context.

My thought is that this needs really good mapping, perhaps down to the level of individual lanes of a highway. Metrics like “residential population coverage by 5G” are wildly insufficient. Regulators need to get serious about data collection / monitoring and maybe crowdsourcing of actual user-experienced data - which here might mean both for devices inside vehicles and for external antennas. (This is a theme I’ll come back to in my Indoor Wireless work soon).

I’ve often discussed this in the context of “why 5G failed”. It’s because most of the use-cases suggested for network-slicing or guaranteed quality either need ubiquity (ie full coverage, including for 5G SA) or specific localisation and customisation (ie best served by private networks).

As a coverage suggestion, ACEA recommends satellite networks to improve connectivity continuity, particularly in places where terrestrial service is weak. In some ways this makes a lot of sense, as vehicles are ideal NTN endpoints and can support large antenna arrays. Tesla is already discussing support for Starlink - which I suspect is part of the reason for European manufacturers’ interest in ensuring they’re not left behind.

But while that makes sense up to a point, I’m not sure either satellite broadband or especially D2D continity services can easily accommodate all vehicular QoS requests, given spectrum, capacity, latency and uplink limitations. The paper also refers to “un-interruptible” connectivity, which is an unrealistic objective, given tunnels, garages with canopies, mountain roads, forests and dense urban canyons where satellite visibility may be absent or intermittent.

Nevertheless, it is really good to see external validation that coverage is a prerequisite for “differentiated connectivity”.

Network fees and interconnect: the debate is not telcos versus Internet companies

ACEA’s paper welcomes the absence of explicit fair-share or network-fee mechanisms in the DNA proposal, arguing such fees would raise the cost of vehicles, slow innovation, damage European automotive competitiveness and plausibly result in higher prices or reduced functionality for consumers.

But it goes further, warning about de-facto network fees emerging through euphemistically named “commercial negotiations” (or perhaps “voluntary conciliation” in DNA-speak). OEMs might be charged for access to baseline network performance, required to purchase premium connectivity to avoid degradation, or otherwise pressured by operators’ control over access to their retail customers.

This is exposes the inadequacy of a common Brussels narrative, framing the debate as a contest between European telecom operators and large US Internet companies. Once a regulatory principle exists that arbitrary sources or destinations of user-demanded (or IoT-demanded) traffic should contribute directly to access-network costs, there is no obvious way to confine it permanently to a handful of companies without appearing utterly mendacious and hypocritical.

This doesn’t just bring car manufacturers into the picture - it would also capture gaming companies, healthcare companies, industrial IoT providers, banks, media companies, satellite operators, cloud customers and public-sector platforms. All of these send or receive substantial Internet data volumes or require particular performance.

Going further, the proposals around IP Interconnect regulation also take on a different colour. It is not just auto companies that are starting to appear at peering sites with their own ASN numbers. Broadcasters, banks, gaming companies, universities and cybersecurity companies also operate their own interconnections. Some of the discussions around DNA and interconnect could have a huge and unintended set of spillovers, if they came to fruition.

ACEA supports commercial agreements for specialised services, but wants Commission monitoring, transparency and protection against discriminatory behaviour. That seems reasonable, but does not obviously link to the DNA’s proposed ecosystem conciliation mechanism. I suspect that it is more concerned that some operators might provide favourable terms to their own inhouse telematic businesses, or otherwise compete in the wider automotive marketplace for data and connectivity services, or somehow make automotive MVNOs less competitive.

The organisation’s intervention is therefore useful evidence that concern about network fees is not inherently a defence of “Big Tech”. It is a concern about telcos creating a new and unfair tollbooth that could eventually face almost every data-intensive sector.

Subscribe now

Other themes

ACEA identifies a genuine mismatch between automotive and telecom lifecycles. A vehicle which can take 5-10 years to develop and then remain on the road for 20-30 years is a poor fit with telcos retiring network generations and refarming spectrum at relatively short notice. Vehicles stranded with 2G or 3G eCall equipment after networks have closed is obviously problematic.

ACEA wants operators to publish coordinated transition plans several years in advance, or maintain a minimum service for maybe 8-10 years. This seems pretty reasonable to me, especially if governments support a thin de-minimis shared network, rather than requiring each MNO to operate obsolete infrastructure on expensive spectrum assets. I could also see an option for telcos funding retrofits and replacements, the same way US operators will pay for airlines to change old altimeters in order to get access to upper C-Band spectrum.

The eCall callback recommendation is probably the strongest specific suggestion in the paper, and is an issue I identified in a client consulting project about 9 years ago. When an automatic emergency call is interrupted, or when additional detail is needed, a Public Safety Answering Point may need to call the vehicle back. These calls can attract high or unpredictable charges because connected vehicles may use international roaming, global connectivity platforms or unusual numbering ranges such as +882 and +883. In my original research, I found that some PSAPs were even configured to prohibit international calls, although that may have changed.

Unlike some of the broader rhetoric around investment and competitiveness, this appears to be a concrete interoperability and charging failure that the DNA could usefully fix.

Lastly, the FM/DAB request is a great example of the type of unexpected consequence I’ve been talking about. At one level it’s automotive industry over-reach, opportunistically realising that these are communication modes that relate to spectrum policy, broadcast etc - and therefore covered by the DNA and national regulators. But it also uses the regulators’ own language turned back on itself. We know that telecom policymakers like the idea of “tech neutrality”, and so this is basically just saying “radio works over an Internet connection as well as broadcast, so let the market decide rather than mandating one technology in particular”.

Conclusion

I do not agree with all ACEA’s proposals, but it raises genuine issues around eCall callbacks, 2G/3G network shutdowns, baseline versus premium connectivity and the inadequacy of generic “5G coverage” statistics for moving vehicles. That said, its switching suggestions would increase the risks of lock-in to expensive or underperforming services.

It also misses various important elements that could be touched by the DNA, notably the important role of private networks in factories and other sites, plus explicit discussion of IP interconnect, CDNs and AI.

Rules on interconnection or traffic payments that ignore those layers may assign cost or responsibility to the wrong actor. The OEM may not control the relevant interconnection arrangement at all, or it may purchase it indirectly as part of a cloud or CDN service. Conversely, a large manufacturer may operate its own delivery infrastructure and resemble the Internet companies that Brussels policymakers think they are regulating.

It’s also worth noting that the report does not really discuss AI autonomous (or remote) driving in detail, perhaps because the industry know it will always need safe control even if the vehicle is completely offline. One of the founding myths of 5G was that autonomous vehicles were a realistic central use-case. They’re not. Guaranteed QoS is not a substitute for onboard, local intelligence and autonomy.

A universe of unintended consequences

My wider concern is unintended consequences. The DNA attempts to span an extraordinary range of domains, from numbering and emergency communications to spectrum, switching, slicing, interconnection, cloud, private networks and satellite. Each has different technologies, actors, economics and failure modes.

That makes precise scope and definitions absolutely essential for the DNA and any final regulatory instruments that emerge. Otherwise, the EU may spend the next decade constructing sector-specific exemptions from rules that were too poorly conceived in the first place. Automotive is just a starting point here.

If DNA provisions on interconnection, stakeholder conciliation, specialised services etc. evolve into a possible demand for payment to access networks, the affected parties will not stop at Google, Meta or Netflix. Healthcare providers, industrial IoT platforms, gaming companies, banks, security companies and many other enterprises could become the next victims of the “large traffic generator” lie I’ve discussed before.

A specialised-service provision intended to enable innovation may encourage operators to reclassify baseline performance as premium connectivity. A voluntary conciliation process, intended to resolve a few interconnection disputes, may create an implicit sender-pays regime affecting dozens of industrial sectors.

None of this means the EU should avoid reforming communications regulation. The existing EECC predates much of the current evolution of telecoms, cloud, satellite, private networks, shared and dynamic spectrum, industrial systems and embedded connectivity.

The DNA therefore needs a functional test and impact statement for each obligation. Policymakers should ask what exact service or activity is being regulated, whether it is publicly offered or self-provided, and whether it is just “telcos” or a broader set of stakeholders affected.

There needs to be clarity about where market power actually exists, whether rules distinguish safety-critical or consumer-facing applications and whether different workloads can be separated.

There are also yet more signs that the DNA needs to be underpinned by solid metrics, statistics and data-collection and analytics - not just handwaving about “population coverage” or “homes passed”. We need accurate, updated datasets and maps.

It would also mean explicit protection against using interconnection or conciliation rules to manufacture network fees without evidence of a genuine uncompensated cost or market failure.

Above all, we need to get rid of the familiar Brussels framing of telcos on one side and Internet companies on the other. Like the cringeworthy term “OTT”, this thinking is obsolete.

The real ecosystem contains manufacturers, enterprises, connectivity aggregators, MVNOs, cloud providers, towercos, CDNs, satellite operators, private-network users, application developers, public authorities and consumers, with many organisations occupying several categories at once.

The connected car does not fit neatly into the DNA’s regulatory boxes. That is not a defect in a car or its manufacturer. It mean the boxes are wrong.

If you are interested in my speaking and advisory work, please contact me via either LinkedIn or my speaker website

That sentence could express genuine enthusiasm. Or irritation. Or weary resignation. Or a warning that the next words will be “Can I speak to a human?” or “Cancel my order”.

The transcript is identical in every case.

That is a pretty fundamental problem for some of the current generation of voice AI services, whether they are telecom-centric AI-enhanced calls, or more AI-native voice experiences.

Subscribe now

Much of the world of speech-based AI is still built around a simple pipeline: capture audio, convert it into text, hand the text to a language model, then generate a spoken answer. Which means that it may produce a perfectly “reasonable” spoken reply, or generate an accurate transcript, but still misunderstand the conversation. And conventional emotion-detection analysis might not get it right either, if it just looks for “angry words”.

Humans do not just process words. We listen to pitch, rhythm, stress, volume, hesitation, pauses, laughter, sighs and changes in speaking rate. We compare what is said with how it is said, who said it, what happened previously and what is happening around them. (There are also visual cues expressed in-person, or on a video call).

In other words, meaning is partly acoustic. It’s linked to a term known as “prosody“, referring to aspects such as pitch, cadence and pauses in speech.

Technically, prosody is a subset of a larger family of acoustic elements known collectively as “paralinguistics”, which also includes other non-lexical vocal information such as laughter, sighs, huskiness, hesitation, crying, chuckles, throat-clearing, etc.

At the risk of a horrible pun, this all sounds obvious. Yet much of the voice AI industry has spent the last few years discarding the audio as soon as speech-to-text has finished with it.

Listening as well as transcribing

This leads to an obvious question: can voice AI solutions blend in these other facets of speech, ideally in a standardised way, so that other applications and tools (eg telco voice platforms, CRM systems and contact-centre agents) can understand them and act / respond / transcribe appropriately?

A number of AI companies are trying to get closer to genuine speech intent and meaning, and there is also some interesting academic work in this domain. Some of the developments I’m watching include:

• Hume AI has an “expression measurement” API, which detects expressive characteristics of either recorded or live speech. So it might give a rated transcript segment such as “Yes, that’s absolutely fine. {Disappointment 0.81, Annoyance 0.73, Doubt 0.61}”. It also has a set of tools for rating AI voice models on their emotional awareness and paralinguistic abilities.

• OpenAI is using speech-to-speech models, which don’t go via the speech-to-text path at all. In theory this is better both for overall conversational latency, and some of the prosody aspects of Voice AI such as pauses and cadence. It doesn’t mean it can do structured emotional classification or calibrated ratings, although it is able to detect things such as frustration in a contact-centre call.

• Google Gemini’s Flash Live model is explicitly described as having “acoustic nuance detection”, while Gemini’s broader audio models support emotion detection. Gemini Live also includes what Google calls Affective Dialog, adapting its response style and tone to the expression of the person speaking.

Over time, these type of “speech reasoning” models should get better at mapping spoken words and lexical interpretation, with acoustic evidence and external context and analysis. For example, knowing that a call is someone’s third complaint this week might suggest exasperation or sarcasm is quite likely,

A surprising amount of supposedly advanced contact-centre “sentiment analysis” still largely analyses transcripts. Imagine this comment in a customer-service context:

“Wonderful. I’ve only been waiting six weeks.”

A transcript-oriented sentiment system might take that literally. Conversely, a prosody-aware system should realise that the customer probably isn’t offering a testimonial.

Share

Some of the prosody and paralinguistics research long pre-dates the current LLM wave, such as this 2008 paper with the wonderful title of “The Sound of Sarcasm“. Another newly published Chinese research paper, Towards More Expressive Spoken LLMs, is more recent and AI-centric, with the idea of “ParaIntent”, a benchmark with 14 categories of communicative intent. It has a concept for blending explicit words with implicit vocal delivery, and text and audio tokens, which it calls “Acoustic-Lexical Decoupled Policy Optimisation”.

It will be much harder to make that work across a broad range of accents, cultures, noisy calls, poor codecs, disabilities, neurodiversity, deliberate deception (i.e. “acting”) and so on.

Sarcasm is not an emotion label

There are multiple layers here, and some of what gets described as “emotional analytics” is often too basic. Some of it just works on the words / text, and some on basic prosody. But reality is more complex.

Sarcasm is a particularly useful stress-test because it often depends on contradiction. “That was really helpful” may mean the opposite if it follows the fourth failed identity check. “Take your time” can be permission or a complaint. “Yes, obviously” can signal agreement, contempt or an attempt to end the exchange. There is no universal sarcastic tone. It varies by speaker, language, relationship and type of irony. Context often matters more than any single vocal feature.

A sarcasm training dataset (yes, really) may also have real-world limitations. MUStARD, for example, uses clips from TV comedies, but sitcom sarcasm is written, acted and surrounded by visual context. A tired customer disputing a bill is not the same as John Cleese complaining about a Hungarian phrasebook.

Share Disruptive Analysis by Dean Bubley

And even then, the system has to do something sensible as a result...

Now, let’s assume our empathic Voice AI manages to detect sarcasm accurately, and even make an educated guess about the reason for it.

“Thanks so much for your help”. The lexical interpretation is positive. The acoustic interpretation is strongly negative. The context is that someone’s broadband has failed for the third time this week.

This detection phase is a good start. But detecting sarcasm is only metadata, unless it changes behaviour, and whatever is said next. There needs to be a separate loop for what changes because of the lexical / acoustic incongruity.

Maybe the next response should be a clarification question, or a soft query, to double-check “Sounds like this isn’t the outcome you wanted. Let me check what’s happened”. At the very least, the system shouldn’t continue its normal path of trying to upsell, or cheerily ask if the user would like to fill in a survey.

Obviously there is also the chance to offer a transfer to a human. Basically the task involves distinguishing perception vs. interpretation vs. policy and avoiding brittle automations based on dodgy pseudo-psychology.

Subscribe now

Can it be standardised?

This whole area is a work-in progress, with limited standardisation or reliability yet. There are many operational and also regulatory challenges here. What should be recorded as audio? When is a transcript enough?

The emerging vCon, or Virtualised Conversation work at the IETF looks like a useful container, including an analysis object for information such as transcripts, sentiment, summaries and other derived data. I’ve written about vCons many times before, and spoke at an event in Dallas earlier this year.

I think there is a maybe a missing element here - perhaps a standard Spoken Interaction Annotation Format for prosody, affect and paralinguistics. That could incorporate a record of acoustic observations such as unusual pitch movements, cadence, laughter, hesitation or affectations such as frustration, sarcasm, sincerity or humour. It could even try to deduce intent at different phases of the conversation.

It would need to be quite fine-grained and auditable. A CRM field saying “customer personality: sarcastic” is not really helpful. But “at 02min43, model X version Y estimated a 78% chance of sarcasm, based on acoustic and contextual evidence“.

A live version could be even more useful, both for contact-centres and also for some of the types of non-contact centre “AI front-desk assistant” interactions I recently talked about in my AI Voice for SMBs white paper and webinar.

The network’s voice media stream contains useful machine-readable context that a transcript simply doesn’t. There is a genuine role here for telco platforms, CPaaS providers and contact-centre infrastructure vendors, and also probably for more device- and consumer-centric app and device providers. (I’m thinking of sci-fi author Iain M Banks’ infamous and often highly-sarcastic personal drones here).

Some caution is warranted, however. This type of analysis can improve appointment bookings, sales or technical support. But it could also become manipulative if a synthetic, nuanced agent can detect vulnerability, impatience or uncertainty. The line between empathy and optimisation is not especially thick.

I’ll try to coin a new term here: this is “Conversational Error Detection”.

From phone calls to ambient voice

The same issue extends beyond telephony and contact centres. I’ve written before about ambient voice, involving real-world audio capture in a physical setting such as a retail setting, doctor’s office or a transport hub.

That also needs emotion and intent detection. After a difficult customer meeting, a salesperson drops into the chair opposite their boss, and says: “Well, that went brilliantly.”

A transcript-led system writes “meeting successful” into the CRM and schedules a follow-up sales proposal. A better system notices flat or exaggerated prosody, the preceding discussion, perhaps even a sigh, and asks: “Do you mean the meeting went badly?” to clarify and then updates the record.

A maintenance worker beside a rattling machine in a factory snorts, and then says, “Yes, that sounds healthy.” The background noise itself is part of the meaning.

That creates larger opportunities, and - obviously - much larger consent and permission problems.

The telco opportunity is not an emotion API

Telcos and communications providers have a plausible role here, but it is not to build the world’s most emotionally convincing foundation models.

They already control important parts of the signalling and media path for many voice-based interactions: numbers, call routing, codecs, recording, SIP infrastructure, hosted PBXs and contact-centre platforms. They also increasingly have consent-management capabilities.

Telcos can preserve or expose the acoustic information that a transcript discards. They can support real-time, permissioned media access, recording, conversation data containers such as vCons, jurisdictional and sovereignty controls, and manage auditable hand-offs between models and humans. All can contribute to this type of analytical awareness, and ability to spot lexical / acoustic divergence.

That may also (finally) drive a network-quality angle as well. Aggressive noise suppression, low-bitrate codecs, packet loss concealment and other audio processing may remove or distort the cues an AI is trying to interpret. “Voice quality for machines” might not always be identical to voice quality for humans. Even edge processing could be valuable here.

Regulators have noticed, but not consistently

The regulatory position is already interesting, with regional variations.

The EU AI Act prohibits using AI to infer emotions in workplaces and educational institutions, with narrow medical and safety exceptions. The European Commission’s own guidance gives a specific example: a call centre cannot use webcams and voice recognition to monitor the emotions of its employees. Yet the same guidance says that tracking customers’ anger or impatience is not prohibited by that clause.

But “detecting prosody” and “inferring somebody’s emotional state” aren’t necessarily legally or technically identical. As per above, sarcasm isn’t really an emotion. There is happy sarcasm and angry sarcasm. I suspect regulators, lawyers and standards bodies are going to spend a lot of time disentangling categories of observation, inference and affectation here.

In the UK, there is no direct equivalent to the EU workplace ban, but data-protection rules still matter. Voice data can be personal or biometric data, profiling may require a data-protection impact assessment, and consequential automated decisions raise additional questions.

Conclusion: Transcripts aren’t everything

We’ve all got used to AI-generated transcripts of calls. They’re starting to become table-stakes for many B2B interactions, and I suspect they’ll get emulated in the consumer and personal space as well.

But that only tells part of the story. The next generation of voice AI will increasingly separate and then recombine what was said, how it was said, when it was said and what was happening around it. If consent and regulation permits, it should listen to words, prosody, silence and surrounding sound - and link those to broader contextual information.

Perhaps that is the monetisation layer for some operators - I’ll have to give it some more thought, and also discuss it at upcoming events.

On that topic, I’ll be giving a keynote on AI Voice at the upcoming Alianza Navigate 2026 event in September. Sign up here & come talk to me about this. I’ll be genuinely intrigued by anyone with a practitioner’s view on this. Really.

Also - save the date for the next Unthinkable Lab on AIVoice in central London on October 29th. We’ll be asking some tough questions on all this... get in touch if you’d like more details.

Nobody in the technology industry can ignore the huge impact that AI and datacentres are having on the networking sector, turbocharging the previous cloud-driven cycle. Huge requirements for connectivity between and inside datacentres are driving deployments of massive extra capacity, and development of new optical and photonic techniques to manage demand.

Inside DCs there’s huge interest in photonic switching and co-packaged optics between GPUs, which I wrote about last year. Over the wide area, we’re seeing ramp-ups in long-haul fibre backbones, subsea cable builds, metro fabrics and NaaS for scale-out expansion and local interconnect (see my Mplify Alliance event report here). Companies like Nokia Ciena Juniper Networks and others at the component / chipset & laser level are doing very well.

Inference and other classes of XPU, including at edge sites, change things again. At enterprise locations, faster Ethernet, private 5G cellular and Wi-Fi is increasingly linked to AI applications and vertical solutions such as machine vision or robotics / physical AI.

Subscribe now

... but not yet on the access network

But the big unknown is the massmarket access network - fixed or mobile, especially for consumer smartphones, residential access, and other end-points such as vehicles and smart-home solutions.

Will continued GenAI deployment and use - and new AI paradigms such as agentics, edge-offload or robotics - move the needle for normal telco networks? Will it drive usage, congestion, investment - and perhaps extra regulatory focus or a need for more spectrum?

There is a lot of enthusiasm in the telecoms industry at the moment about AI driving future last-mile network traffic, and some of it is justified. Data volumes could certainly increase if we see more personalised video (and more consumption of it), more user-generated AI-enabled media, more cloud inference offload, more machine-to-machine interactions, perhaps agentic data (discussed later) and more uplink from cameras, robots, vehicles, drones, glasses, industrial systems and other forms of “physical AI”.

I wrote recently that uplink growth may be seen first in fixed broadband networks, earlier than mobile.

Traffic bull vs bear cases

In other words, there are perfectly plausible upside “bull” scenarios for access-network traffic. Various forecasts from Nokia Cisco, Ericsson and assorted fellow analysts and consultants have tried to quantify this. In some cases they have used existing observable data and tried to extrapolate. (My review of a recent Nokia study is here)

But whatever the source and growth methodology, that is only half the story.

Too many AI traffic forecasts start with a one-way assumption that “AI usage” automatically becomes a justification for more spectrum, edge infrastructure, slicing, new “kinetic tokens”, more QoS APIs, extra capex, or whatever else there is an interest in selling or regulating.

That chain of logic is partial and perhaps weak, because while AI will certainly change traffic patterns, and increase certain aspects, it may also reduce some types of traffic, shift traffic away from public access networks, compress it, filter it, delay it, localise it, or make networks more efficient at carrying it.

Moreover there are certain trends that we can see on the horizon - such as more on-device AI, or semantic compression - but which haven’t yet arrived. There could be disruptive innovations that lead to extra upside or downside.

So in the interest of balanced discussions, here are 10 reasons why AI might actually reduce network traffic, or at least reduce the need for additional public fixed or mobile access capacity. I’m still not sure on the net outcomes. But I am sure that any realistic analysis needs to consider both sets of drivers and argue convincingly for the balance of effects.

There is no guaranteed one-way ratchet effect here. Remember - voice minutes and SMS volumes kept growing... until they didn’t.

1. On-device inference reduces round trips to the cloud

The most obvious traffic-reduction mechanism is local inference. Phones, laptops, tablets, cars, cameras, home gateways, industrial controllers and AR glasses are all getting better AI-capable silicon, and while not all inference will be local, enough of it will be to matter.

Speech recognition, translation, image enhancement, summarisation, object detection, spam filtering, call screening, notification handling and many personal-assistant tasks can run partly or wholly on the device, rather than requiring every prompt, audio stream, image or sensor feed to be sent to a remote model.

In many cases, this will substitute for existing network traffic and cloud processing, not just minimise new AI loads.

This does not collapse all traffic, because models still need updates, some queries still require external knowledge and reasoning, some tasks need cloud-scale compute, and many applications will use hybrid local/cloud architectures. But it changes the traffic profile. In many cases, the network sees a much smaller transaction: a prompt, a command, a summary, a metadata or JSON object, an event, or nothing at all.

2. AI filters data before it reaches the network

A lot of future AI traffic claims rely on cameras, microphones and sensors sending huge amounts of upstream data, but the more realistic architecture is often the opposite: sensors generate huge raw data volumes locally, and AI decides what is actually worth transmitting.

A security camera does not need to upload 24 hours of empty corridor video if it can upload a short clip when it sees a person, a vehicle, smoke, or a suspect package. A factory vision system does not need to send every frame to a remote cloud if it can identify defects on-site, and a smart-city system does not need to stream every sensor reading if local analytics can classify, aggregate and discard the uninteresting bits.

This principle is not new. During an open day at CERN a few years ago, I remember seeing a rack of hardware discriminators (called triggers) for a particle-accelerator’s sensors, which acted to filter out 99% of useless readings from one of the main experiments in an underground chamber at the LHC.

AI filtering effectively changes the network from a raw-data conduit, to an exception-reporting system. The traffic reduction can be enormous, because most real-world sensor data is boring. AI’s job in many of these systems is not to archive everything, but to throw most of the potential traffic away before it ever hits the access network.

To be fair, there’s a fair amount of this type of pre-processing in existing systems already, without adding AI as well, but once again it should improve the ratio of both new and existing streams - as CERN itself now demonstrates, as it now has neural-network based triggers - see this post.

Subscribe now

3. Semantic compression beats dumb compression

Traditional compression reduces the size of a file or stream, but AI enables something more radical: semantic compression. Instead of sending a video, a system might send “three people entered the room, one wearing a red jacket, no safety helmets detected”. Apart from botanists, nobody really needs every leaf on every tree in a movie scene to be rendered as its 4K original.

Instead of sending a full document, it might send the extracted fields. Instead of sending all Instead of sending a 30-minute meeting recording, it might send the transcript, decisions, action points and disputed items in a vCon record (as me how this fits with AI voice evolution) - and include a link to the raw audio if needed.

This will not apply everywhere, because sometimes people need the original image, clip, document or evidentiary record, and in regulated, safety-critical or legal settings the raw data may need to be retained. But even then, the access network may not need to transport all of it in real time, or transport it at all if local storage, local audit trails or delayed synchronisation are enough.

AI shifts the unit of communication from bytes to meaning, and that creates a major problem for simple traffic-growth extrapolations, because a lot of value can be delivered with fewer bits.

This also has huge implications if content can be semantically compressed on a server, sent across the network efficiently, and then reconstituted on the user device. (Or vice versa for uplinks)

4. Agents may replace browsing, not just add to it

Agentic AI is often presented as a traffic multiplier, with agents constantly fetching web pages, calling APIs, checking databases, negotiating with other agents and generating background activity at machine speed. There is a vision of agents running on your phone or laptop, which are continually going to the web to “do stuff”. Some of this probably comes from early experience of things like OpenClaw.

That may happen to excess, especially with badly-designed agents, repeated research tasks, web automation, testing systems, scraping-like behaviour or enterprise workflows that are allowed to run without proper cost controls. Some will legitimately create more traffic.

But agents can also replace a lot of human browsing or streaming, and in many cases they may retrieve less, not more. A human researching a holiday, insurance policy, phone tariff or medical condition might open dozens of pages, load images, adverts, trackers, recommendation widgets, autoplay videos, cookies, scripts and analytics, while scrolling through irrelevant material and abandoning most of it.

By contrast, a well-designed agent could query structured sources, use cached indexes, pull snippets, discard irrelevant material and return a concise answer. Much of the traffic will be “inside the cloud” as well, with a server-side agent connecting to its peers via IP interconnect and cloud backbones. The user may never load the original pages over the access network, and the agent may avoid much of the advertising and tracking overhead that makes the modern web so bloated.

This is a huge unknown. Some agents will be wasteful and noisy, while others will be brutally efficient. But it is far too simplistic to assume that every agentic workflow adds traffic on top of today’s web activity, because some of it will replace today’s messy, duplicated, media-heavy web activity with leaner machine-readable interactions. The current web is an inefficient traffic generator. AI may make it worse, but it may also route around some of the bloat.

5. AI work may happen inside “the cloud”, not across the last mile

The note in the previous section about agents driving inter-cloud traffic can be generalised.

A lot of other AI traffic will be real, but in the wrong place for many telecoms narratives. Training traffic is mostly inside data centres and between data centres. Large-scale inference may sit inside hyperscaler clouds. Enterprise AI may run in private cloud, SaaS platforms, corporate data centres, sovereign-cloud environments, campus infrastructure or sector-specific platforms. Agentic systems may perform most of their work near the data sources, with only the final output crossing the access network.

That still means more network demand, but not necessarily more public mobile or residential broadband demand. It may mean more 800G and 1.6T optical, more data-centre interconnect, more peering, more private fibre, more cloud-region connectivity, more traffic between GPUs, more east-west traffic and more enterprise WAN redesign and NaaS.

But those are not the same thing as “we need more macro-cellular spectrum”, upgrade to 10G-XGS-PON fibre, or “every cell site needs GPUs”.

This distinction matters because the telecoms industry - and their regulators and policymakers - often uses the word “network” as if all networks are the same. Maybe 80% of all attention from executives, investors, governments and commentators is on residential broadband or consumer 5G, as it’s what we all face day-to-day - and the retail market is huge, competitive and complex. But it’s not necessarily centre-stage for AI.

Share

6. Scheduling and batching can move traffic out of the busy hour

Not all AI traffic is real-time, and so not all AI-related data needs to move immediately. From a network engineering standpoint, peak or busy-hour traffic matters much more than total aggregate traffic.

Model updates, training data, telemetry uploads, map updates, logs, synthetic data transfers, video archives, industrial analytics and fleet learning can often be delayed, batched, compressed or scheduled for times when capacity is cheap and plentiful.

A car can upload diagnostic data when parked at home, a robot can upload logs overnight, a phone can synchronise AI “memories” when charging on Wi-Fi, a camera system can upload evidence clips after local triage, and a home gateway can fetch model updates at 3am.

Again, this is not new or AI-specific, but the nature of AI itself means that (hopefully) optimisation of this sort of timing should improve. AI traffic forecasts therefore need to ask “when?” as well as “how much?”.

This is especially relevant for mobile spectrum arguments, where busy-hour, cell-level, location-specific demand is the real constraint. A large amount of off-peak AI-related data transfer does not create the same spectrum problem as live uplink from thousands of users in a dense urban cell or venue.

7. Better offload / federation will absorb a lot of AI traffic

Even where AI does increase access traffic to and from user devices, it will not all go over public mobile networks. Most AI usage on smartphones already happens indoors, on Wi-Fi, over fixed broadband.

We can reasonably expect AI-enabled connection management tools or OS functions to improve this ratio, especially if coupled with wider adoption of automated access via approaches such as Passpoint and OpenRoaming.

Many enterprise AI systems will use Ethernet, Wi-Fi 7, private fibre, local 5G, industrial wireless or on-premises compute. Physical AI in factories, ports, warehouses, hospitals, universities, airports and campuses is much more likely to use local connectivity than national macro networks.

Various mobile AI use cases will still need wide-area cellular, of course. Vehicles, logistics, field workers, consumer mobility, outdoor robotics and remote monitoring are obvious candidates for these types of application generating traffic.

But the centre of gravity for heavy AI traffic is likely to be indoors, fixed, local, private or cloud-adjacent. That is why I remain skeptical of “AI therefore mobile spectrum” arguments in particular.

To be fair, AI prove to be may be an argument for better Wi-Fi, private networks and cellular neutral hosts, as I’ll be discussing in my July 16th workshop on Indoor Wireless (details and sign-up here). But that’s not so much about data volumes as reliability, latency, sovereignty and security.

8. AI can improve network efficiency or change user behaviour

AI is not just a traffic source. It is also a network optimisation tool.

I’m about to head off to DTW in Copenhagen this week, and I’m sure I’ll be saturated with details of operators and vendors already using AI and machine learning. It has applicability for traffic prediction, energy management, RAN parameter optimisation, fault detection, routing, anomaly detection, customer experience management, capacity planning and spectrum efficiency.

That doesn’t all necessarily reduce traffic, although some forms of network management and demand-shaping will indeed have that effect. And while some of this will be overhyped, or just turn out to be old optimisation techniques with an AI sticker, the direction is real.

If AI helps networks carry more traffic with the same spectrum, same fibre, same routers or same number of sites, then the pressure for additional capacity is lower than a simple demand forecast suggests. This is especially important for mobile, because if traffic growth slows while efficiency increases, then the investment story looks very different - with possible overcapacity risk, as I have written before.

9. Personalisation can reduce wasted media consumption

A lot of today’s bulk consumer data traffic is recommendation-driven waste: autoplay videos, infinite scroll, poorly-targeted adverts, oversized images, preloaded content never viewed or listened to, duplicate downloads and pages stuffed with scripts, trackers, cookies and other undesirable nonsense.

Uploads are equally bloated - why do we back up the 5 lousy video takes or blurry/badly framed images, as well as the nice one we actually want?

AI could certainly make all of this much worse, or much better. One one hand, we could suffer from it generating endless personalised slop, spam video adverts and addictive scrolling.

But it could also make content discovery or sharing more efficient.

A good AI assistant might say: “You do not need to watch these five videos or listen to hours of podcasts. Here are the three points you care about, and here are the clipped highlights compiled”.

I’ve long wanted to have a tool that can, say, watch or listen to all the recent telco results webcasts and answer queries like “give me the bits where they talk about spectrum”. I don’t need the sections where the CFO talks about pensions, or where some sycophantic financial analyst congratulates them on a “great quarter, guys”.

Going further, such a tool might block low-quality content - imagine a personalised “avoid boring waffle” filter!

Now obviously the advertising industry will fight hard to preserve wasteful engagement loops, and content creators might balk at brutal editing. But the downside scenario exists. Any serious traffic forecast needs to model the possibility that AI reduces some categories of human attention-driven traffic, even while increasing others

10. Economic constraints will discipline AI traffic

AI usage and traffic is not just technology. It is economics. Cloud inference, GPU time, energy, API calls and mobile uplink all cost money. Enterprises will not allow uncontrolled agent usage to generate unlimited traffic, and there will likely be consumer protections too.

The first wave of AI usage has been distorted by free tiers, investor subsidies, ludicrous concepts like “tokenmaxxing” and novelty. We already see the push-back on this.

As AI gets priced properly, applications and infrastructure will become more efficient. They will use better model-routing, cache more, compress more, summarise more, use smaller models, run locally, batch requests, and avoid shipping around unnecessary context and data.

This economic imperative will translate to data traffic as well. The idea that AI agents will run everywhere, all the time, generating unlimited public-network broadband and mobile traffic, is not a forecast. It is a cartoon.

Conclusions

I am not saying AI will reduce network traffic consistently or ubiquitously. It clearly will not. AI will increase some traffic, in some places, at some times, over some networks, and it will drive a lot of investment especially in datacentres, fibre routes, interconnect, NaaS and elsewhere.

It may also increase some fixed access traffic and some mobile uplink traffic, especially for video-heavy, sensor-heavy, physical-AI and enterprise applications. But that is not the whole story.

AI could reduce access traffic through local inference, semantic compression, filtering, scheduling, offload and application / behaviour change. It makes everything more unpredictable as there will be a dynamic mix of increase/decrease drivers and dependencies.

The net-net is hard to gauge - and it seems to me most people don’t bother to try. They mostly just focus on the bull case, for obvious reasons. But modelling the upside while ignoring the downside or relegating it to a footnote or sidebar is not enough.

The unanswered question is not simply “will AI increase traffic?”, but where, which network, which direction, at what time of day, on whose infrastructure, at what cost, with what substitution, and with what impact on public fixed or mobile access networks.

Operators and policymakers faced with AI network forecasts need to ask questions about whether both sides of the ledger have been considered equally.

If not, they should consider why that might be, and react accordingly.

This post was originally published on my LinkedIn newsletter. See LI for more comments and responses.

We also cover Indoor Wireless and regulatory shifts towards assuring end-user and application performance, rather than just headline coverage.

Subscribe now

01:10 - Sovereignty as a key theme at telecoms events in Europe, UK & Canada

02:30 - New proposed European cloud public-sector procurement rules could cause problems for telcos. (See Dean’s LI post here)

04:40 - Benoit’s thoughts on European sovereign cloud propositions. and possible re-globalisation of tech companies to map to new geopolitical realities.

13:30 - The Telecom Industry needs to work collectively, across competing subsectors, to define a proper framework for Telecom Sovereignty to avoid getting sidelined by the Cloud/AI/chips focus

14:15 - Benoit talks about policy-meets-data, a recent webinar he did with Ookla (link), and another study about user-experienced quality vs. Internet Exchange location.

21:10 - Dean and Ben discuss Ofcom’s new position paper on mobile quality, and its use of OpenSignal crowdsourced data and methodologies. Also network speeds and latencies (Dean will be reviewing that paper soon on his Substack)

29:30 - What’s next? Dean will be at conferences on NaaS, spectrum, satellite D2D and then DTW (events list here) and will be running his own workshop on Indoor Wireless policy & tech trends on July 16 (details + earlybird here). Benoit will be chasing the dead EU “fair share” network-fees zombie as it tries to reanimate itself around the world

30:10 – Book Ben & Dean for your event! Get us on stage at your conference, or facilitating at your private meeting or executive offsite. Connect & message us on Linked.

01:10 - Sovereignty as a key theme at telecoms events in Europe, UK & Canada

02:30 - New proposed European cloud public-sector procurement rules could cause problems for telcos. (See Dean's LI post here)

04:40 - Benoit's thoughts on European sovereign cloud propositions. and possible re-globalisation of tech companies to map to new geopolitical realities.

13:30 - The Telecom Industry needs to work collectively, across competing subsectors, to define a proper framework for Telecom Sovereignty to avoid getting sidelined by the Cloud/AI/chips focus

14:15 - Benoit talks about policy-meets-data, a recent webinar he did with Ookla (link), and another study about user-experienced quality vs. Internet Exchange location.

21:10 - Dean and Ben discuss Ofcom's new position paper on mobile quality, and its use of OpenSignal crowdsourced data and methodologies. Also network speeds and latencies (Dean will be reviewing that paper soon on his Substack)

29:30 - What's next? Dean will be at conferences on NaaS, spectrum, satellite D2D and then DTW (events list here) and will be running his own workshop on Indoor Wireless policy & tech trends on July 16 (details + earlybird here). Benoit  will be chasing the dead EU "fair share" network-fees zombie as it tries to reanimate itself around the world  

30:10 – Book Ben & Dean for your event! Get us on stage at your conference, or facilitating at your private meeting or executive offsite. Connect & message us on Linked.

On 12^th March 2026, the bipartisan US-China Economic & Security Review Commission (USCC) published a report on “China’s Management of Electromagnetic Spectrum Resources”, researched and written by BluePath Labs.

The study focused on China’s approach to spectrum management, including how centralised authorities coordinate commercial and military requirements for frequency resources, as well as its international outreach in standards creation.

It also makes several recommendations for US authorities in the field of spectrum management, leaning on what it sees are China’s successes and limitations:

• Streamline and speed up the US spectrum allocation process, while preserving its emphasis on market-based dynamics.

• Accelerate R&D efforts on cognitive, dynamic and sharing models of spectrum management and efficient utilisation.

• Engage more broadly in international forums for wireless standardisation and spectrum governance.

However, the report’s detailed analysis and conclusions are something of a mixed bag.

While it notes reasonable observations about institutional friction and roadblocks in US spectrum management processes, the report seems to endorse a mistaken, largely debunked view of the US lagging China in spectrum for 5G / 6G. In particular, it reuses the flawed definitions and calculations of “midband” spectrum promoted by CTIA, which I have criticised in the past.

It also ignores the role of non-traditional cellular networks such as CBRS-powered industrial enterprises and overlooks the US’s extensive work on spectrum-sharing and dynamic allocation mechanisms.

Subscribe now

Regulatory agency coordination and speed

Much of the report details the organisation of China’s various spectrum-using groups, especially the commercial wireless sector and government/defence agencies. It notes that MIIT (Ministry of Industry and Information Technology) is responsible for allocating spectrum bands, acting as a centralised coordination function, and aligning with strategic national goals, such as industrial automation / IoT, smart manufacturing and transportation.

Spectrum management and 5G-related allocations were very prominent in the CCP’s 13^th and 14^th Five-Year Plans (2016-2020 and 2021-2025 respectively. (The 15^th plan was published the same day the USCC report came out. It has much less direct emphasis on spectrum management, focusing more on AI, 6G systems & satellite integration).

Two key points are made:

• China does not use market mechanisms like auctions for spectrum allocations. This helps it accelerate bringing frequencies into use, but reduces some scope for spectrum-related innovation, or aftermarket transactions when market conditions change.

• Coordination between military and civilian stakeholders is labour-intensive. There is considerable secrecy about government and defence spectrum allocations or preferred bands, so “deconfliction” occurs behind-the-scenes. Because the telecom operators are also state-controlled, the type of legal disputes sometimes seen in the US are not replicated in China.

The report calls for “streamlining” of the US system while keeping auctions as a central element. This seems to imply better (and faster) alignment between FCC and NTIA and other stakeholders, in the hope of bringing spectrum to market more quickly.

However, this seems to overlook the much longer and larger legacy of existing spectrum incumbents in the US, including defence systems such as radars which have been in place for decades. China’s recent huge expansion of its military has given it an opportunity for a more “greenfield” approach to spectrum use. While the US system is more fragmented, this is not the only reason for extra complexity.

Share

5G leadership

The report partly attributes China’s early lead in 5G deployment and uptake to its speed in issuing spectrum licenses to national operators.

“In part as a result of its spectrum management decisions, the PRC has already become the most prominent market in the world in 5G deployment and consumer take-up”

Yet while its centralised model for spectrum allocation has assisted its own state-controlled carriers somewhat, this hugely underplays a range of other, non-spectrum-related levers for 5G.

Notably, China’s authorities focused much more on acceleration of coverage in terms of numbers of cell towers, rather than capacity through greater spectrum availability. It enabled simpler mechanisms for siting or upgrading cell towers, and also focused explicitly on extending wireless connectivity to indoor settings.

China’s government also encouraged many of its industrial and public-sector organisations (often State-owned or State-run) to adopt 5G networking solutions using domestic equipment and managed by State-controlled carriers like China Mobile. It also had an explicit push for later versions of 5G such as “Standalone” and “5G Advanced” which are more driven by the software running the network than the frequencies used.

Uncritically repeating the CTIA’s mid-band fallacies

Where the report falls down badly is its comparison of China’s 5G spectrum allocations to America’s. Sadly, it uncritically cites CTIA’s calculations and definitions, which systematically overcount China’s midband spectrum and undercount US allocations. These figures have been widely criticised, including by myself. (See detailed articles here and here).

Unsurprisingly, CTIA’s commentary on the report applauded this uncritical repetition of its own flawed data, which used an arbitrary and unique lower cutoff for midband at 3GHz, specifically designed to exclude T-Mobile’s large holdings around 2.5GHz, which the carrier itself describes as midband.

Meanwhile, the CTIA analysis included China’s shared low-power indoor 3.3GHz allocation, which is hardly a traditional full-power macro 5G framework. Correcting for this arbitrary cherry-picking, the US and China in fact have had broadly similar 5G midband resources.

The report also falsely asserts that China “is the first country in the world to open the entire 6GHz band for 5G and 6G applications”, when in fact the MIIT has only earmarked it for intended future allocation, and has not yet proceeded with that process since the announcement in mid-2023. An early auction of part of the band in Hong Kong yielded mediocre results, and there are also signs that the Chinese Wi-Fi industry is lobbying for the lower part of the band to be released on an unlicensed basis. Recently, in May 2026, it approved only the upper part of the band for early 6G trials, but not commercial 5G services.

The report does not highlight that China and US both had early 5G spectrum strategies, but diverged in approach. China chose to prioritise mid-band spectrum, while the US instead chose to pursue early releases of high-band (mmWave) spectrum. Despite that, the report indeed makes the factually-incorrect statement, “The FCC began auctioning licenses in the high band for 5G use in November 2018, but it did so in smaller allocations than China had”, when in reality, China issued no 5G mmWave spectrum at all during that period, and is only now starting to consider it.

The references to the FCC’s lapse in auction authority are also slightly strange – this occurred in March 2023, by which time the US had already conducted various releases of 5G-suitable spectrum. The C-band auction took place in 2020-21, and 3.45-3.55GHz band a year later. At the same time, operators such as T-Mobile were repurposing existing (and often underutilised) midband assets such as its 2.5GHz assets.

Subscribe now

Sharing and dynamic spectrum mechanisms

The USCC report discusses the importance of dynamic spectrum management, cognitive radio and future evolution of new spectrum sharing models. It gives an interesting review of China’s research work in this field, especially from the perspective of military and radar coexistence with other wireless systems, but wrongly concludes that “China’s research in cutting-edge approaches to increasing the efficiency and utility of electromagnetic spectrum use in general appears to be proceeding much more rapidly than that of the United States”.

That is incorrect.

The report fundamentally misses existing US efforts in this area – not just in research, but in terms of actual deployment and adoption for hybrid government / commercial use.

Its sole reference to CBRS is in a single footnote referring to the original FCC NPRM from 2012. Ignored entirely are the FCC’s successful 2020 PAL auction and the wide deployment of more than 430,000 base stations by thousands of CBRS network operators in the years since. It does not describe its sharing mechanisms (including sensor systems and the SAS databases), actual deployment since 2020, and the highly successful ecosystem it has helped create.

CBRS is a real-world example of shared military and commercial spectrum use via automated systems, with ongoing iterative evolution with CBRS 2.0 and likely future CBRS 3.0 generations. This represents a huge contrast with China’s current resource-heavy, manual coordination between defence users and public networks operators. It should be noted that Europe and UK have various sharing mechanisms and trials of their own, including the UK’s 2.3GHz military/commercial shared temporary licenses.

The report also omits any mention at all of US innovation in dynamic spectrum sharing, such as the AFC (automated frequency coordination) architecture for 6GHz, the core role of sharing in the NTIA National Spectrum Strategy, the NSF’s SpectrumX initiative or the work of the National Spectrum Consortium (NSC) on research and testbeds around advanced and “highly dynamic” sharing and coexistence techniques, including those based on AI.

Thus while the report’s recommendation for increased funding for research is welcome, it overstates China’s current position in this field and completely misses the distinct U.S. lead in this field. Recommended research funding is a matter of extending US dominance and widening commercial applications, rather than playing catch-up.

Leave a comment

Conclusions

The USCC report has some good insight, especially on the interaction of China’s civil and military spectrum users, and MIIT’s regulatory mechanisms and processes. This reflects the authors’ expertise in assessing geopolitical issues and especially the internal workings of the Chinese government.

However, the report does not do full justice to either US 5G / 6G spectrum allocation mechanisms and historic releases, nor the current state-of-the-art in dynamic sharing and cognitive spectrum technologies.

While the recommendation on streamlining US spectrum allocation processes is reasonable at a top level, it appears to misunderstand the real drivers for China’s 5G successes – and falls into the trap of repeating CTIA’s flawed assertions about an imaginary “spectrum gap” between China and US.

Importantly, while the report is right to call for more investment in spectrum-sharing and dynamic allocation mechanisms, it suggests that the US is currently lagging China. It does not discuss the multiple ways in which the opposite is true – notably the success of CBRS, the proliferation of Private 5G networks, and the use of AFC for 6GHz.

Overall, the USCC report is a welcome contribution to the spectrum debate, but its contents should be viewed in the appropriate context. A revision which covers the newly-published Chinese 15^th Five Year Plan (without which the recommendations are very rearward-looking), gives some broader global context, and updates its treatment of US 5G spectrum allocations and spectrum-sharing expertise would be welcome.

Subscribe now

Operators, vendors, regulators and investors should pay more attention to fixed uplink and what it shows about end-user and application behaviour, before overhyping the impact of physical AI, robots, drones or other future applications on 5G / 6G uplink needs, or spectrum requirements.

Uplink growth is real. But it needs to be seen in context

Uniquely, Italian regulator AGCOM publishes traffic stats split between upload and download volumes, and between fixed and mobile broadband. Other regulators do not do this, at least not publicly. They often publish total traffic, average data per user, 5G traffic share, or speed-test-style upload performance, but don’t break out uplink traffic volumes.

The Italian 2025 stats show that uplink was around 12% of fixed traffic volumes and 9% of mobile traffic. Both have been growing, albeit not that fast, and not in a way that implies massive stresses to networks, except maybe in localised places such as stadiums, concerts, airports, tourist centres and dense enterprise sites.

In 2025, mobile uplink traffic grew at a substantially faster rate than fixed, but in relative terms the gap is actually wider than it was 5 years ago.

That’s not a huge surprise to me - I’ve always expected many advanced in applications and also edge compute and AI to be more fixed-oriented, or at least fixed-first. But it might come as a shock to some in the cellular part of the telecoms business, who often suffer from “mobile exceptionalism” syndrome.

As a reference point, it’s also worth noting that overall fixed traffic (uplink + downlink) was about 3.2x mobile in 2025, compared with 4.8x in 2021, so there has been some narrowing of the difference, but overall most countries are still overwhelmingly fixed-broadband primary in terms of total “tonnage”. (Note: AGCOM warns that some figures are affected by accounting-method changes and methodological refinements, but I don’t think there’s an impact here).

Share

Why fixed uplink leads the way

Italian fixed broadband upload share moved from about 10.0% of total fixed traffic in 2021 to about 12.4% in 2025. That appears to be a real behavioural shift, even if it is gradual - and importantly, it seems to have pre-dated the emergence of AI as a mainstream phenomenon.

Based on assorted other sources of data and anecdotes, the likely drivers are not exotic. Think:

• Video calls and meetings

• Cloud upload, backup & sync of files, images, video and normal Office-type content

• Work-from-home data and productivity tools

• User-generated content uploaded to social media sites like YouTube and TikTok

• Security cameras and other home IoT

• Home-based streaming workflows (think Twitch)

• Gaming and interactive applications

• Probably some homes with vehicles uploading telemetry data overnight

These are mostly mundane applications. This is not realtime processing of uplinked robot video, edge processing of AR/VR rendering, or AI agents on smartphones, Mac Minis or home gateways chattering away 24x7.

One other factor should be borne in mind - fixed broadband typically includes a fair chunk of non-residential connections, while mobile is (mostly) consumer-centric. There is some evidence to suggest that business broadband use is much more symmetric (HT Jason Presement) - probably because most workers don’t spend 4 hours a day watching Netflix or other video streams.

In 2025, Italy’s fixed networks carried 8.33 EB of upload traffic, compared with 1.97 EB on mobile networks. In other words, fixed uplink volume - from these mostly normal and unsexy applications - was more than 4x mobile uplink volume.

This is a bit of a reality check for anyone claiming that mobile uplink is about to be transformed by AI and robots. If physical AI and distributed agents were going to change traffic symmetry, fixed networks would already be showing the leading signal. There isn’t any evidence of that yet.

When discussing residential broadband, it is also important to recognise the huge volumes of downloads driven by video streaming services, consumed on large (often 4K) TV screens. That makes the significant upload fraction even more notable.

One other interesting data point - US data company OpenVault LLC also measures fixed uplink/downlink traffic. Its latest report shows that uplink is 11% of the total on the FTTH networks it tracks, but only 8% on DOCSIS cable, which perhaps reflects the relatively poor uplink characteristics of older cable broadband, as the medium is shared between many users.

Share Disruptive Analysis by Dean Bubley

What can we expect from mobile uplink?

Mobile upload is indeed growing. AGCOM shows mobile upload up 35% in 2025, much faster than mobile download growth of 19%.

Ericsson has also shown evidence of faster growth in mobile uplink in a large number of other operators and national markets too - albeit from a low base.

That trend is definitely worth watching. But it’s still far too early to call out physical AI, drone video, streaming telemetry or edge-inferencing offload as the key determinants of the future shape of mobile traffic. And it’s also not clear whether mobile can overtake fixed broadband in trems of symmetry, or if it will continually be playing catch-up.

Again, I strongly suspect the explanation for recent mobile uplink growth is more ordinary:

• Social-media uploads, especially short video from smartphones

• TikTok, Instagram and similar workflows

• Live-streaming at sports and music events

• Cloud storage and backup - including backups of the images and video created above, plus (wastefully) a lot of the unused “takes” and badly-framed images that we keep on our phones & which doubtless get uploaded as well

• Video calls, including possibly more outdoor, macro-area video calling in some cultures and countries

• Adaptive codecs exploiting higher uplink capacities, by sending higher-resolution image/video

• Some IoT and vehicle telemetry or video uploads

Anecdotally, at least from my visits, Italy seems quite keen on outdoor video calling, perhaps more than some other countries. That is not a rigorous statistic, but it is a plausible contributor to the impressive 35% growth in 2025.

The physical AI / edge-offload argument sounds intuitive. Cameras upload video. Robots send sensor streams. Vehicles exchange telemetry, feeding into future model training runs. AI agents need cloud and edge inference. Wearables and smart glasses may send video, audio and positional data back into the network.

But while it might sound intuitive, it isn’t showing up so far.

There are a few reasons why not:

• Developers are selective about what gets uploaded, when and where. They may send filtered or compressed data, rather than raw streams over cellular. They may also wait until the device is on Wi-Fi.

• On-device processing (and AI inferencing) is getting better, removing the need to send everything to the network

• Mobile uplink performance and capacity is variable and unpredictable, especially indoors and incongested areas. That impacts user behaviour, developer preferences and also codec / compression usage.

• Uplink-oriented “differentiated quality” via Network APIs and QoD is still at very early stages, and only usable for very specific applications such as outdoor broadcast uploads, and on networks with 5G SA. There is limited evidence of broad willingness-to-pay, and plenty of questions about capacity and service-assurance in the real world.

Ericsson has leaned heavily into the uplink theme, as have other vendors such as Nokia and Qualcomm. At multiple events in the last few months, I’ve heard speakers argue that AI agents, cars, wearables and robots will require “a lot more uplink”.

But there are two problems here:

• It’s all very speculative, especially when considering the impact on total traffic. The data is patchy, cherry-picked and inconsistent, as my comparison vs. fixed upload shows.

• It’s likely to be very location / time-specific. Uplink is likely to be heavily localised in specific places where cloud access, social uploads, video calling and aspects of AI/robotics are concentrated.

If AI, cameras, robots and cloud uploads are going to change traffic symmetry, fixed networks are already showing the leading signal, with private cellular and enterprise / public Wi-Fi following closely. Public wide-area mobile still has a long way to catch up.

Beware the “uplink = more spectrum” narrative

The 5G / 6G discussion often jumps straight from “AI devices will generate more data” to “mobile networks need much more spectrum”. Uplink is now part of that story.

I’ve already spotted US mobile association CTIA shouting about AI-led upload growth, loudly trumpeting an Ericsson statistic that GenAI had a 26/74 up/down split at one operator... without also mentioning the inconvenient fact that this was just the breakdown for a 0.06% element of the overall traffic. The other 99.94% was still averaging 10/90.

There are at least four reasons to be cautious.

• First, national traffic averages do not show a sudden mobile-uplink break-out. Italy’s mobile upload share is still under 10%, and even with Ericsson’s prediction of 1.5x or even 2x faster growth, that means it’s still a long haul to 12%, 15% or eventually 20%

• Second, many AI, IoT and robotics uplink-heavy use cases are site-based and will use fixed, Wi-Fi, private cellular or local edge networks. Think about security / surveillance video cameras uplink at an airport, or inspection systems on a factory. Robots mostly live in warehouses or hospitals - they’re not wandering down the street, except a few delivery bots which are unlikely to be streaming HD video. For consumers, most uplink will be at entertainment or sports venues - or at home.

• Third, a lot of current mobile-network uplink spectrum is likely under-used. Many mobile networks have older “paired” FDD spectrum with a fixed uplink/downlink split, often below 2.5GHz. The downlink side has been the scarce resource. The uplink half of those paired bands may have more spare capacity, especially as the original reason for a 50/50 ratio (voice calls) has diminished in importance.

• Fourth, the bottleneck is almost certainly not national macro/outdoor spectrum supply. It will be local cell-edge performance, device transmit power and radio quality, indoor-to-outdoor propagation, antenna placement, interference and various other factors.

That does not mean uplink can be ignored. But it does mean that regulators should ask better questions - and audit existing spectrum use - before paying too much attention to uplink as an argument for future spectrum allocations.

They should also likely prioritise spectrum or policy that is location-specific or dynamic / event-based. Public mobile uplink will matter for some use cases: live video from phones, emergency services, vehicles, drones, temporary events, outdoor robotics and wide-area telemetry. But that is not the same as saying consumer mobile networks suddenly need to support ubiquitous Physical AI “just in case”.

Regulators and operators should collect & publish better statistics

The main takeout for policymakers and regulators is simple: collect and report more detailed uplink statistics like this, and collect and compare equivalent data for fixed and mobile networks. As far as I know, AGCOM is the first regulator to publish stats like this - but please let me know if you’ve seen similar figures elsewhere.

Indeed, the fact that the mobile industry already claims uplink as a driver of future growth (and perhaps spectrum and investment) suggests that (a) they must know the current uplink baseline, and (b) they cannot complain about the “burden” of providing relevant data. I can’t believe it’s an especially hard job for fixed ISPs to collect the relevant stats either.

In an ideal world, there would be a separate breakout of FWA traffic from both MBB and other “wired fixed”, so that we can understand the real-world traffic data, and also make sure to avoid either double-counting or leaving out important detail.

Regulators should also attempt to classify the applications and device types behind uplink growth. Is it video calls? Cloud storage? Gaming? Social video uploads? Home cameras? Business VPNs? FWA? Connected vehicles? Industrial cameras? Drones? AI inference?

And how localised is the demand? Is my assumption that much of the uplink is concentrated in specific buildings or sites correct, and does that imply that private / local wireless is going to be more important in the Physical AI / 6G era?

Without these details, comments about “uplink demand” become slogans, rather than a measurable and well-defined trend.This is especially important for spectrum policy. If a regulator is considering future IMT spectrum allocations partly because of projected uplink growth, it should first ask:

• Whether existing FDD uplink capacity is actually congested

• Whether the growth is public mobile or private/site-based

• Whether fixed/Wi-Fi/FWA trends are giving an earlier signal

Finally, regulators, policymakers - and also operators and standards bodies - should look into the scope for new approaches to “localising” uplink/downlink ratios. This is highly complex because of the need to synchronise radios between neighbouring cells to avoid interference, but there may be new sharing and coexistence approaches that could allow greater flexibility to meet future uplink/downlink demand trends or temporary spikes.

Leave a comment

Conclusion - uplink should be watched, but beware of hype and lobbying

The question is not if uplink will matter more. It almost certainly will. The better question is whether the data supports the idea of a sudden mobile-uplink inflection in the lare 5G / early 6G era.

So far, the answer looks like: not yet.

My current working view:

• Uplink will rise as a percentage of overall mobile traffic.

• The rise will be gradual, not sudden, as it comes from a low base.

• Physical AI is overhyped compared to more normal applications driving uplink traffic

• Consumer / macro uplink is about influencers and WFH, not robots and drones - at least in the medium term

• Specific locations and times will drive much greater symmetry, eg sports & entertainment events, or industrial sites with many cameras.

• Fixed broadband and Wi-Fi will be the leading indicators in many instances.

• Private wireless and FWA will complicate the categories.

• Public macro mobile will see some new uplink pressure, but not necessarily enough to justify spectrum-demand claims.

• The real evidence will come from regulators that publish upload/download traffic, not from vendor roadmaps alone.

• There needs to be more R&D into mechanisms for flexible uplink/downlink ratios

A key question is whether modern 5G networks finally allow mobile to catch up with fixed networks in terms of uplink performance and availability - or if they actually overtake it because of different application demands and traffic dynamics.

And one absolute indisputable output: everyone needs more stats like AGCOM’s.

00:30 - Wi-Fi trends, including Wi-Fi 8, AI + Wi-Fi

06:00 - US policy & regulatory about Wi-Fi, including ban on international router imports, thoughts on 6GHz spectrum

08:57 - Benoit's thoughts on the FTTH Council Europe event in London & the wider industry. What happens to the industry when fibre rollout is almost complete? What about Africa?

13:25 - Optical LAN, FTTX in-building, copper wiring/Ethernet replacement and Photonics

19:40 - Fibre uptake trends in Europe

24:00 - FutureNet and Unthinkable Lab outcomes. Trends in network automation, OSS / BSS, AI+telecoms trends, sovereignty and 5G standalone.

29:40 - What's next? Ben's partnership with Point Topic on FTTH takeup acceleration, Dean's May event schedule.

31:40 – Book Ben & Dean for your event! Get us on stage at your conference, or facilitating at your private meeting or executive offsite. Connect & message us on Linked.

Links:

Wi-Fi Now Global Congress: https://wifinowglobal.com/usa-2026/

FTTH Council Conference 2026, April 14-16: https://ftthconference.eu/

Dean's FutureNet commentary:

https://www.linkedin.com/posts/deanbubley_telcos-edge-sovereignty-activity-7452275477878132737-7aVy

https://www.linkedin.com/posts/deanbubley_5gsa-5g-neutralhost-activity-7452614197273849856-AODk

Unthinkable Lab on Telecoms Sovereignty, April 23: https://www.telecomtv.com/content/unthinkable-lab-sovereignty-in-the-unthinkable-age/

In May, Dean will be at Connect(X), Canadian Telecoms Summit, TelecomTV DSP Leaders  & Netaxis Inspiration Day. Click here for details and to arrange a meeting.

The latest DBTT Podcast by myself & fellow analyst Benoit Felten is out. In this episode we discuss FTTH, copper switch-off and the upcoming FTTH Council Europe conference in London, the recent BEREC Stakeholder Forum event in Brussels & EU DNA Digital Network Act updates, the importance of Wi-Fi and the upcoming Unthinkable Lab event on Telecoms Sovereignty.

Find it on your favourite podcast app (Apple , Spotify) or wait for the YouTube video version which should be out later today.

2.00 – Gigabit broadband, FTTH, VHCN, Cable, Copper Switch-Off and more European fixed broadband issues. What’s in the DNA, what is Body of European Regulators for Electronic Communications (BEREC) saying? Is there a genuine need to accelerate the end of copper, including in-building? What practical and legal issues are raised?

17:30 – European Targets for Gigabit Coverage: Can the FTTH gaps be filled by satellite broadband or 5G FWA? Are regulators helping or hindering this? Meet Benoit at the FTTH Council Europe Conference April 14-16 to discuss it

23:50 – Wi-Fi is the real Broadband Endpoint: The importance of Wi-Fi for the last 10 metres of connectivity, including 6GHz spectrum, shifts in regulatory stance on indoor wireless – and the FCC’s ban of foreign-made home Wi-Fi routers: sovereignty or protectionism?

29:50 – Telecoms Sovereignty: How it differs from cloud/AI sovereignty. What does it all mean, and what can telcos offer in terms of connectivity, solutions and the overall stack? Come to the Unthinkable Lab on April 23 to discuss

34:20 – Book Ben & Dean for your event. Get us on stage at your conference, or facilitating at your private meeting or executive offsite

Links:

FTTH Council Conference 2026, April 14-16: https://ftthconference.eu/

Unthinkable Lab on Telecoms Sovereignty, April 23: https://www.telecomtv.com/content/unthinkable-lab-sovereignty-in-the-unthinkable-age/

BEREC Stakeholder Forum: https://www.berec.europa.eu/en/events/14th-berec-stakeholder-forum

Benoit’s article on Copper Switch-Off: https://www.linkedin.com/pulse/conundrums-copper-switch-off-benoit-felten-7okpe

Dean’s review of the BEREC Forum: https://deanbubley.substack.com/p/event-report-berec-stakeholder-workshop

Dean’s article on Telecoms Sovereignty: https://www.linkedin.com/pulse/sovereignty-telcos-draft-structure-dean-bubley-sbuke

2.00 – Gigabit broadband, FTTH, VHCN, Cable, Copper Switch-Off and more European fixed broadband issues. What’s in the DNA, what is BEREC saying? Is there a genuine need to accelerate the end of copper, including in-building? What practical and legal issues are raised?

17:30 – European Targets for Gigabit Coverage: Can the FTTH gaps be filled by satellite broadband or 5G FWA? Are regulators helping or hindering this? Meet Benoit at the FTTH Council Europe Conference April 14-16 to discuss it

23:50 – Wi-Fi is the real Broadband Endpoint: The importance of Wi-Fi for the last 10 metres of connectivity, including 6GHz spectrum, shifts in regulatory stance on indoor wireless – and the FCC’s ban of foreign-made home Wi-Fi routers: sovereignty or protectionism?

29:50 – Telecoms Sovereignty: How it differs from cloud/AI sovereignty. What does it all mean, and what can telcos offer in terms of connectivity, solutions and the overall stack?  Come to the Unthinkable Lab on April 23 to discuss!

34:20 – Book Ben & Dean for your event! Get us on stage at your conference, or facilitating at your private meeting or executive offsite

Links:

FTTH Council Conference 2026, April 14-16: https://ftthconference.eu/

Unthinkable Lab on Telecoms Sovereignty, April 23: https://www.telecomtv.com/content/unthinkable-lab-sovereignty-in-the-unthinkable-age/

BEREC Stakeholder Forum: https://www.berec.europa.eu/en/events/14th-berec-stakeholder-forum

Benoit's article on Copper Switch-Off: https://www.linkedin.com/pulse/conundrums-copper-switch-off-benoit-felten-7okpe

Dean's review of the BEREC Forum: https://deanbubley.substack.com/p/event-report-berec-stakeholder-workshop

Dean's article on Telecoms Sovereignty: https://www.linkedin.com/pulse/sovereignty-telcos-draft-structure-dean-bubley-sbuke

The WG breakouts review each team’s activities over the last year, outline what they’re working on at the moment, and then have a sort of Q&A / group discussion which feeds into a possible work programme for next year. There’s also an ongoing consultation on that, which ends on April 15th.

However, because Brussels doesn’t seem to have many open, “set piece” conferences of this type, I get the sense that many attendees confuse policy vs. regulatory focus, and assume that the WG leads have broad domain knowledge and deep influence, rather than enacting regulations for other departments’ policies in specific areas. This means that some of the discussions go off in random directions.

Subscribe now

About BEREC’s role

As a background, Body of European Regulators for Electronic Communications (BEREC) is the collective group of EU national telecoms regulators (NRAs), which has to implement and monitor the various rules and regulations, as well as providing advice and input to EU policy suggestions and new legislative projects like the Digital Networks Act. The working teams are mostly affiliated with the national regulators – I guess on secondment.

BEREC also has (some) resources to pursue various projects of its own interest that it feels will contribute to ongoing regulatory effectiveness in European telecoms, including looking at new technologies, assessing business models and competition, and understanding end-user issues such as pricing and use of the (open) Internet.

Many of the speakers are very cautious and diplomatic with their language, as they have to tread a fine line between focusing on regulatory detail, keeping an eye on trends and helping “development and better functioning of the internal market for electronic communications networks and services”.

To give an idea of the tone of some of the sessions, compare a line I heard in the morning part was “we’re just bureaucrats, we don’t make policy” with what you’d hear from the current FCC in the US.

There is a clear self-censorship on occasion, with a limit on “speaking out of turn” or speculation, especially at an event like this, where everyone was trying to understand what is in the Commission’s mind, or read the tea-leaves about possible backroom discussions on taking the DNA forward.

This is all made even more awkward because the European Commission is currently trying to re-cast BEREC as the “Office for Digital Networks” or ODN. It also sometimes resents BEREC’s independence, while simultaneously “must take utmost account of any opinion, recommendation, guidelines, advice or regulatory best practice adopted by BEREC”.

In short, the EU wants BEREC to be more political, less technical and to “get with the programme”, while BEREC’s job is to criticise “the programme”. You see the tension.

This week, BEREC issued its preliminary feedback on the EU’s draft DNA. If you strip away all the pseudo-diplomatic (but obligatory) pleasantries, it features a heavy dose of “you what?!”, a main course of “well that sounds nice, but some definitions would be helpful” and a side-order of “if this is supposed to be simplification, I’d hate to see something more complex”.

(The actual wording of the press release was: “identifies important areas where further refinement is needed to ensure that the DNA strengthens, rather than complicates, Europe’s internal market for electronic communication“. There’s no way I could ever be that polite!)

Share Disruptive Analysis by Dean Bubley

Open Internet, Wireless and Future Trends

I couldn’t get along to all the sessions as they ran in parallel. I went to the WGs looking at:

• Open Internet

• Wireless Evolution

• Future Trends

Open Internet

The Open Internet group is one of the teams that probably causes most headaches for the Commission and also the classical telecom industry. It is the protector of Net Neutrality in Europe, and has been at the centre of pushbacks against the various attempts to undermine it.

It has repeatedly thrown scorn at politicians’ clumsy attempts to engineer a way for classical European telcos to somehow charge fees to Internet and cloud companies – “fair share”, “two-sided markets”, “cost recovery” and more recently IP interconnect “dispute resolution”. It pointed out that actually IP interconnect works remarkably well without regulation or any fake “resolution” attempts.

During the session the group leaders took questions and comments from people concerned with the treatment of network slicing, and also the “voluntary conciliation” process suggested in the DNA draft. The general sense was that more clarity is needed - a theme that cropped up multiple times during the event.

As they noted, it’s tough for BEREC to provide guidelines for dispute resolution processes, seeing as the few disputes seem to get resolved perfectly well today already. The phrase “is there a problem to be solved?” was mentioned.

There is a concern that this ends up being a “dispute creation mechanism”, and that a review of the DNA in subsequent years might harden the rules, in a sort of one-way ratchet.

The discussion also highlighted another thing I’ve been hearing about lately - the proposal to change the EU’s “general authorisation” rules for communications services providers to include some private and backbone network operators.

This sounds abstruse but is actually really important.

Essentially it could mean that CDN providers and others get treated as a type of telco, rather than part of the Internet market. That would subject them to a variety of extra rules, interconnection mandates and general regulatory hoopla - and would bring them into mandatory dispute resolutions which (depending on the NRA involved) could involve payments and various non-neutral measures.

This definitely needs to be clarified - and there’s a very strong argument that national regulators and others need to push the European Commission to avoid that type of miscategorisation. (It could also capture other “private networks” such as financial, healthcare and even utility companies).

One last thing - I asked if anyone had started looking at “token-based charging” and whether AI-related network services might be the next thing in line to attempt to circumvent neutrality concepts. Seems as if I’m the first to mention it.

Leave a comment

Wireless Network Evolution

From my point of view, probably the most useful and important WG meeting was the one called Wireless Network Evolution, which covers a broad portfolio of spectrum, mobile, satellite, Wi-Fi and many other areas. It aligns with a lot of my focus areas.

It’s been doing a lot of work in recent years on private 5G, network slicing, 2G/3G switch-off, satellite D2D – and is now also supporting / giving feedback on the Commission’s DNA work.

It was canvassing opinions on “what else should we next”, and after several years banging the drum, I was very pleased to hear that indoor wireless is on their list of possibles, including neutral host, Wi-Fi and more.

I argued strongly in favour of that (perhaps with an emergency call availability hook), as well as mentioning sensing and AI+wireless as other things that are worth covering, perhaps via some sort of broader horizon-scanning work. (See later for comments on sensing).

There is a tie-in here with the panel I spoke at with Safer Buildings Coalition at the IWCE conference in Las Vegas recently - indoor wireless is not just important for broadband access, but also for emergency calls in places like stairwells and fire-exits. The switch-off of 2G / 3G and forced use of VoLTE makes this even more tricky, as that has caused problems not just with coverage but also authentication. It’s also interesting to link to Wi-Fi, as Wireless Broadband Alliance Inc. is creating emergency-call profiles for OpenRoaming.

I’d urge readers to submit comments to BEREC advocating for the Indoor Wireless project.

The WG also pondered whether Open RAN is worth revisiting, and also on the timelines to start doing something about 6G. Others suggested resilience as a theme, and also how “use-it-or-lose-it” might work to drive a secondary market for spectrum.

Planning and Future Trends

This session primarily related to a recent draft report published by this WG about Virtual Worlds and Web 4.0. That seemed to be a bit of hold-over from the Metaverse era, albeit with some clear pivots in the draft towards more current hot topics such as Physical AI and physical/virtual coexistence in areas like Digital Twins. It considers ways telcos could get involved via APIs, traffic differentiation and more network capacity / investment.

My sense was that the room talked past each other a bit during this session.

The session leads were looking for feedback on the report and where to take it next. I suspect few people had read it (I’d scanned it & had AI summarise it for me, but not gone through in forensic detail), but other people in the room had just seen “planning and future trends” and assumed it was a general opportunity for asking about hot topics.

Lots of people asked about drones, which hasn’t been a major focus for the WG. Conversely, the leads were also anticipating lots of questions about 6G, which didn’t get much attention.

A few goods things for me about the report:

• Recognises the need for hybrid networks, especially for indoor and remote use (Wi-Fi, satellite as well as cellular)

• Mentions wireless sensing, including for Wi-Fi and cellular ISAC

• Some concerns about new “gatekeepers” such as device / AI vendors

• General belief that open Internet and spectrum policies are largely compatible with the vision

• Wariness about suggesting that telcos are natural winners in this space, or that it’s a huge potential boon for edge, APIs, slicing etc. Business cases are unproven, basically.

• Recognition of unknowns such as on-device processing vs. network traffic & in-network edge mythology

On the other hand there wasn’t much about a formal deep-dive into the impact of indoor users - which is important given the focus on AR/VR, healthcare, education etc. Perhaps this WG could work with the Wireless group on a combined indoor project?

I also think that it needs to consider network APIs at a top level - not just mobile / CAMARA but how these would work across multiple networks and domains.

Overall, I think this WG has a variety of future directions and possible projects - but personally, I’d like to see a much broader horizon-scanning remit, that doesn’t get trapped into writing reports on what was considered cool and important three years earlier, but has since faded. So perhaps “regulatory considerations for post-digital networks” or issues likely to emerge with 6G, or indeed AI Voice, which was mentioned by nobody at all.

Plenary sessions

After lunch, the main plenary part of the conference took place, with a couple of presentations and panels. A lot of this was (unsurprisingly) about the draft DNA proposals, what the drivers were behind it, BEREC’s initial feedback (ie this isn’t simplification), and assorted stakeholder positions.

It was notable that the Commission’s EVP Virkunnen, who is the architect of the DNA and who looks after a portfolio including sovereignty and technology, turned up right at the end of the conference with a short presentation entirely devoid of anything new or substantive, and then didn’t take any questions. I’m not sure if she was pushed for time, or didn’t want to get into a to- and fro- argument given BEREC’s criticisms. I was struck by the difference in tone and substance compared to her speech and fireside chat at last year’s Techritory, for instance.

Another Commission speaker went through in more detail what was trying to be achieved by linking the DNA to other initiatives around defence and sovereignty, as well as providing investors with more predictability via elements such as indefinite spectrum licenses.

There was the standard line about Europe’s ambitions for AI needing better connectivity, without explicit mentions that really it’s not about access connectivity. There’s an elephant in the room here, in my opinion. At least 80% of the DNA and policymakers’ focus is about access networks (gigabit broadband, 5G & 6G, copper switch-off etc), but the AI era is far more linked to backbone and transport networks and interconnect. Ironically the EU wants to regulate these more tightly and give more control to telcos, rather than incentivise huge investment and diversity from multiple sources.

The best presentation was from the 6G Smart Networks and Services Joint Undertaking (SNS JU) which was the first I’ve seen to look at the regulatory implications of some key 6G elements. I’d spent the morning asking every WG about ISAC Sensing, so I was very happy to see a slide asking almost exactly the same questions. Privacy, data ownership, liability and so on. I’d also add competition and wholesale - and I think that regulators everywhere should be thinking about network sensing as a whole as it also crops up in Wi-Fi, subsea cables, FTTH networks and more. (Please get in touch if you’d like me to facilitate a workshop, or run a study on this).

Another good set of questions were around policy issues relating to network AI and terrestrial / satellite hybrids.

The final panel followed on from this, featuring a few speakers talking about the future of networks. The GSMA speaker liked the spectrum aspects of the DNA but was irked by net neutrality and a lack of clarity about slicing and specialised services. Ericsson’s panellist talked up the need for more uplink, low latency and distributed compute for physical AI, but swerved the contentious topic of edge computing. He also pointed out the need to drive adoption and deployment of 5G SA before focusing on 6G. The Amazon spokesperson was primarily focused on its Leo satellite proposition and NTN integration, but also referenced the risks I mentioned above about the DNA capturing private networks and CDNs by equating them to telcos.

But this session also brought the worst cringe of the day. This was a panellist from an organisation called OpenNebula, who was still fully bought-in to the 2018-era view that telcos can displace hyperscalers via the deployment of lots of edge-compute nodes. Obviously this time the stale edgewash narrative has been reheated with a dose of AI inferencing and GPUs, and the patently-wrong cliché of a network/cloud “continuum”.

He appeared to exhort telcos to unwind their hyperscaler relationships, deploy network edge nodes, and ideally convince policymakers to pay to build and deploy a European cloud stack. Apparently “it’s commonsense!”. I’m still trying to unroll my eyeballs from my cranium, a day later.

I asked a question about this, which was half-addressed by Ericsson who made a reasonable point about a mix of on-device, edge(s) and hyperscale compute. From the Eurostack zealot, substantive answer came there none.

Summary

Overall, this was definitely a useful day in Brussels. As well as the sessions themselves, I caught up with a lot of existing contacts, met some new people, and “took the temperature” on the DNA and other issues like 6G, spectrum, Wi-Fi, IP interconnect, sovereignty and more.

A key takeout - regulatory change is not going to happen overnight, and the Commission is getting a lot of push-back on the DNA draft, especially on the notion that it’s somehow a simplification of the existing rules.

I’d also say that I was impressed by some of the realism in the WGs. There’s a recognition of practical challenges in areas like indoor wireless. And there’s some early recognition that sensing, AI-powered networks and other trends pose new regulatory touch-points.

On the other hand, there are still some flawed ideas doing the rounds about the imaginary “cloud-network continuum”, and the view that the EU can have a complete sovereign tech/telecom stack rather than selective control points. (Come to my workshop on April 23 for a realistic view on Telecom Sovereignty).

Meanwhile, BEREC is going to carry on doing its existing work, and I have a number of ideas for projects that I’ll try to suggest it focuses on.

Lastly, there’s another whole sphere of fixed networks, competition, copper switch-off and more. If you’re interested in that - and the read-across to mobile, Wi-Fi and what’s happening in the US and with geopolitics - you should listen (or watch) the upcoming new DBTT podcast edition, which I recorded with Benoit Felten this morning. Watch out for the links, or subscribe to audio or YouTube editions!

This article introduces and explores another of the “error classes” that I am tracking in technology and telecoms sectors: I call it Assumption of Uniqueness.

It follows on from a recent post of mine on “Scale Extrapolation” errors.

Assumption of Uniqueness occurs when a problem is framed so that one favoured solution appears uniquely necessary. Substitutes, workarounds, hybrids and orthogonal alternatives disappear from view. It is not about novelty. It is about substitution blindness.

In telecoms and technology, assumption of uniqueness often appears when a network-, standards-, or infrastructure-based solution is treated as indispensable. This uniqueness is often used to influence investment and regulation.

Similar errors occur in regulation and governments, where a specific policy is viewed as the only way to attain a objective.

It is basically hubris and confirmation bias, scaled up to industry level.

Subscribe now

A topical example: Physical AI & 6G

My recent article on 6G Physical AI highlighted a salient example of this error class.

Parts of the mobile industry are starting to imply that robots, drones and other embodied AI systems will naturally depend on 6G — and that 6G’s own capabilities, from sensing to uplink to edge compute, will effectively define the category.

They then suggest that this audacious vision therefore justifies fresh spectrum, new standards priorities, extra capex and another round of “AI-native network” rhetoric.

I wrote recently that this is exactly backwards. It is Assumption of Uniqueness.

Physical AI is real and likely to grow. That does not make public 6G its unique gateway. When 6G arrives, it will be a useful part of the connectivity landscape, but initially at least, it will not be that important. Sectors like robotics exist already and are not obviously waiting impatiently for the next generation of cellular to arrive with almost-magical capabilities.

Maybe it will become so successful in some instances that it will help redefine some elements - time will tell. But it will need to compete (directly or indirectly) with many other technologies, architectures and workaround.

This is a general, tech industry-wide problem

This issue and error class is not confined to the mobile industry and its advocates - although I regularly accuse them of arrogance and self-importance. It’s common in many other areas of tech as well, with satellite D2D probably the other biggest culprit at the moment.

If you go back 5-8 years, the supposed “uniqueness” of blockchains to solve multiple problems was another stereotypical example.

The recipe is simple: an industry group starts to believe (or at least claim) that its favoured technology, architecture, policy tool, business model or industry sector is the only credible solution to a given problem, while alternatives, workarounds and hybrids are ignored or dismissed.

“If you want outcome X, then solution Y is the only serious way to get it.”

It is one of the main reasons why smart people and capable companies end up with bad forecasts, brittle strategies and overconfident regulation.

Now to be fair, sometimes the hype comes with elements of self-awareness. There might be a “fake it until you make it” attitude, or perhaps more generously, the desire to create enough momentum for a self-fulfilling prophecy. But even there, social media and especially investors tend to amplify it out of faith rather than rigour.

Share

How are Assumption of Uniqueness errors structured?

In my earlier post about scale extrapolation, I looked at a neighbouring error: assuming that because a technology works at one scale, it will scale smoothly and replace something larger, smaller or older.

The uniqueness error often comes even earlier in the chain of reasoning. Before people ask whether their favoured solution can scale, they have already decided it is the only one that matters.

Telecoms / cloud is especially prone to this, because every part of the stack wants to cast itself as indispensable:

• Mobile people think “the answer” stems from spectrum and radio (& SIM)

• Cloud people think “the answer” is massive hyperscale or distributed regional compute and accelerators.

• Application people think the network is a commodity and that “software eats the world”

• Fibre and fixed-network people seem a bit more humble, but in reality they’re also quietly confident and are always looking for another zero or two in speed

• OSS / BSS and identity platform providers think that monetisation comes first, and that complex charging approaches define the architecture.

Regulators and policymakers then often inherit whichever framing is loudest, best-funded or dressed up most convincingly as industrial policy.

So the debate gets rigged before it starts. The underlying problem gets defined in a way that already presupposes the preferred approach. “The solution is mobile. What’s the question?”

Inference or control becomes a latency problem. Latency becomes an edge problem. Edge becomes a location problem. Location becomes a network topology problem.

Competition becomes an “interoperability” problem. Sovereignty becomes a “where is the hardware?” problem. Capacity becomes a “give me more spectrum” problem. Monetisation becomes a “charge the application provider” problem.

In each case, a real requirement is quietly collapsed onto one chosen mechanism.

What causes an Assumption of Uniqueness error?

While I don’t claim to be a corporate psychologist, my observations and experience over the last couple of decades have shown that a few industrial pathologies recur again and again:

• Single-layer thinking. People assume the problem has to be solved in one layer of the stack, rather than across device, application, network, cloud, operations, identity, policy and user workflow.

• Incumbent self-importance. Every sector wants a larger role in the next big thing than it probably deserves. They do not just benefit from inertia; they mistake it for inevitability.

• Narrative economics. “We are one useful part of a messy hybrid solution” is not a good keynote slide. “Only we can enable this future” is much better for lobbying, funding and standards battles.

• Policy simplification. Governments and regulators like clean mappings between goals and technologies. Reality is generally more awkward.

• Path dependency. Existing assets, products and customer base shape imagination. If you own fibre, towers, GPUs, cloud regions or public rights-of-way, you tend to see those as the natural place where the problem should be solved. And while path dependencies are important, they often conflict with other paths which are equally well-trodden.

• Invisible substitutes. The alternative is often not one glamorous rival technology. It is a mixture of software changes, local optimisation, partial redundancy, workflow redesign and hybrid deployment.

The last point on substitutes is subtle but very important. I’ve often come across solutions that are non-obvious substitutes for network-based solutions. I wrote an article on this a few years ago. My favourite is that of improving wireless performance for train passengers, by laser-etching the windows to make the propagation better, rather than adding trackside radio capacity, or on-train repeaters.

Why Uniqueness often feels reasonable

This error is seductive because the favoured solution usually does have a real advantage in some circumstances, for certain use-cases.

• 6G may indeed help certain types of Physical AI deployment and use

• Network QoS may improve performance for some applications, in some places

• Network edge compute may lower latency in some architectures

• Domestic “sovereign” infrastructure may improve some aspects of control or resilience

• Interoperability may strengthen competition in some markets

• More spectrum may increase capacity or improve coverage in some places

• Charging application providers may create some bargaining leverage for some operators, and provide desired capabilities for the developers

The problem is not that these ideas are always wrong. It is that they get overstated into exclusivity and ubiquitous relevance.

Confusing “important” with “indispensable” is bad strategy.

Examples of Assumption of Uniqueness

I’ve already discussed the Physical AI / 6G example. But there are plenty of others. I expand on a few common ones below.

Network QoS guarantees are required for good application outcomes

This is one of the cleanest versions of this error class. The real problem is not QoS - it is user experience and application outcomes. Telecom discussions often jump from “some applications are sensitive to delay, jitter or packet loss” straight to “therefore the network therefore needs explicit guarantees or slicing”.

And maybe sometimes it does.

But often the application can absorb or route around the problem using:

• Adaptive bitrate mechanisms

• Retries, caching and buffering

• Codec and compression innovation

• Offline, asynchronous or local decision-making

• Redesigning the user journey around graceful degradation and alerts

• Network bonding, multipath and fallback techniques

• Interpolation and AI techniques to “hide” glitches

Network edge compute is the only way to reduce latency

Many applications need low or deterministic latency. The assumption is that this inherently means network-based edge compute.

But low latency can also come from:

• On-device inference & local controllers

• Prediction

• Pre-fetching & caching

• Smaller models

• Simplified control loops

• User-interface tolerance for delay

• Changing the workflow so the time-critical part is local

…. we saw this with MEC’s failure, and it’s being repeated for AI inferencing. My general view is that edge-interconnect is more relevant than edge-compute - but uniquely (ahem) most mobile networks have few breakout points to the cloud / Internet / IXPs

More spectrum is the only way to increase mobile capacity

The standard response from many (especially in operator policy & vendor R&D roles) is to frame 5G or future 6G capacity as if it were mainly or uniquely solvable through access to new spectrum bands.

Again, more spectrum can matter. But it is far from the only lever.

Alternatives and complements include:

• Densification and sectorisation

• Better radios and antennas

• Spectral efficiency gains from MIMO, AI and other approaches

• Wi-Fi offload, or indoor / neutral host systems reducing pressure on macro RANs

• Caching and content localisation

• Price signals or other incentives to modify user or application behaviour

There is often a second mistake layered on top of this one as well: treating “more spectrum” as though all spectrum bands were equivalent. They are not. Propagation, indoor penetration, device ecosystem and deployment economics vary enormously. So a lazy “we just need more spectrum” argument often hides two errors at once: one about uniqueness, and one about false equivalence.

National tech sovereignty requires domestic infrastructure

This is the policy version of a uniqueness error, and it is becoming more important. (We’ll cover it in the April 23rd Unthinkable Lab event on sovereignty in London)

The real, underlying, problems for “sovereignty” in telecoms and computing to solve include control, resilience, strategic autonomy, legal & governance tools, continuity / disaster recovery and reduced exposure to external coercion.

But the debate often gets reduced to one crude proxy: physical ownership, country-of-origin or national placement of infrastructure. That is sometimes relevant. But it is not the only route to sovereignty - and indeed can cause problems of its own. (For instance, Ukraine relies on extra-territorial cloud resources, as they are less at risk from Russian attacks).

Other sovereignty-enabling tools include:

• Tight contractual control

• Solution diversification and redundancy, including open-source

• Data escrow

• Source-code reviews for software / cloud elements

• Localised encryption and key management mechanisms

• Backup suppliers / supply chain resilience

The risk is that “sovereignty uniqueness” rhetoric can easily become overly simplistic. It is more politically satisfying to point at a datacentre within national borders, than to talk about exit rights, governance structures or operational resilience. But the latter may matter just as much, or more.

Domestic infrastructure may be part of the answer. It is not the only one.

Other examples

• Interoperability and standards are assumed to be the main path to more vigorous competition between vendors. Regulators sometimes try to over-enforce interop, such as the EU’s poor decision about messaging apps in the DMA. But often, proprietary solutions can address niches or get to market faster, and iterate more quickly. There are also other solutions here, such as switching, multi-homing, gateways, aggregation & competition across layers

• To monetise networks, telcos must somehow charge application providers directly. This is at the route of fallacies such as “fair share” or attempts to reinvent IP interconnect business models. Instead, wholesale, API exposure, bundled vertical solutions, enhanced security and other network monetisation models are more viable, and do not disrupt existing Internet ecosystems.

• Network APIs such as SIM-swap or silent authentication are essential for mobile apps’ fraud management. But alternatives such as passkeys, device attestation, behavioural analytics and assorted other approaches can yield similar results.

One last small, but revealing, Assumption of Uniqueness example comes from city infrastructure. Telecom planners often assume that metropolitan authorities will “obviously” want to rent out lamp-posts or other street furniture for small cells, as though that were the only natural and uniquely valuable use of the asset.

But cities have many other competing priorities beyond wireless.

They might prefer to use limited pole capacity (power, space, wind-rating etc) for many other uses: lighting upgrades, sensors, signage, CCTV, advertising - or even just flowerpots and seasonal lights and decorations.

The uniqueness error here is institutional rather than technical. It assumes that because a network planner sees a lamp post as a small-cell host, the city should too.

Telecoms is not unique here. It sometimes competes with florists and Santa Claus.

Share

The same mistake exists outside telecoms

Telecoms is not special here. I’m not going to fall into the obvious trap of saying this error is unique to tech, cloud and telecoms. While I know them best, I’m also aware of similar assumptions elsewhere.

Energy policy debates often assume that grid stability uniquely requires one category of power-generation asset, when flexibility can also come from storage, interconnection, demand response, smarter tariffs and flexible loads.

Healthcare does something similar when capacity is framed narrowly as just involving hospital-centric care. But home healthcare and virtual ward models are increasingly treated as viable substitutes for selected patients and conditions as well.

Same pattern. Different sectors. I’m sure there are many others as well.

How to spot and avoid this error

The easiest way to catch Assumption of Uniqueness is to force the conversation back to first principles. Industry participants and observers should ask:

• What is the underlying problem that needs to be solved?

• Is everyone describing an outcome, or smuggling in a preferred mechanism?

• What other layers in the stack could satisfy or absorb this requirement?

• What substitutes or hybrids exist already? (They may be non-obvious)

• What would the fallback be if the favoured solution didn’t exist?

• Is the “must-have” claim universal, or does it only apply to a corner-case?

• Who benefits from presenting this as an exclusive answer?

Conclusion

Assumption of Uniqueness is one of the most common strategic errors in technology because it flatters everyone involved. This ties in with the “tyranny of consensus” I identify around Physical AI and 6G. Many stakeholders agree to the same fallacy, but for different reasons.

• Vendors want to feel central.

• Operators want their assets (& installed base of users) to be indispensable.

• Policymakers want the world to look neat, governable and soundbite-friendly.

• Commentators and analysts prefer simple narratives to messy systems.

But real systems are rarely solved by one layer, one actor, one standard, one policy tool or one commercial mechanism. Usually the truth is less theatrical and more hybrid.

Once you start looking for Assumption of Uniqueness, you see the same move everywhere: a real problem, a favoured solution, and a suspicious insistence that nothing else really counts.

That is usually the moment to stop, strip away the jargon, and ask perhaps the most useful question in technology strategy:

What are the substitutes?

6G is now falling into the same trap, driven by vendor hype, MNO wishful thinking and the quest for a justification narrative.

I’ve lost count of the times that people claimed that 4G LTE “enabled” AirBnB or Netflix, or that 5G was essential for autonomous vehicles or AR/VR or remote surgery. Apparently 5G should also singlehandedly increase GDP, improve productivity, and decrease CO2 emissions.

These are examples of two error classes that I track - the familiar “category error”, and another I call “assumption of uniqueness”, the belief that one industry, technology or architecture is the only credible solution for a given problem, and that there are no alternative approaches or workarounds.

And now, based on what I’ve been reading and also hearing in Barcelona last week, a new message is starting to coalesce among parts of the mobile industry:

… that “physical AI” will be dependent on 6G connectivity, and as the “killer app”, it justifies extra capex, and will steer architectural and standards priorities.

You know the drill. We’ve been here before with 5G and URLLC. Many of original 5G design decisions - and the ITU’s IMT2020 requirements - were justified by visions of ultra-reliable low latency communications. Yet even today, you cannot buy millisecond latency 5G, and the use-cases suggested for it either don’t exist, or use alternative mechanisms.

But 6G + physical AI is broader than a reheated low latency narrative. And that’s why it is such a problem.

Subscribe now

It looks like a clunky attempt to justify and rehabilitate all the worst and most irrelevant elements of ITU’s infamous IMT2030 “Wheel” diagram, which I critiqued via the 6G Reset initative in Oct’24.

What is Physical AI?

There is no single formal definition of Physical AI - or the other term I occasionally see, "Embodied AI”, which comes from an older and more academic background about robotics.

But roughly speaking, it means various IoT systems that use AI to perceive, reason and act in the physical world. This is often in real time, mediated through sensors, motion, machine vision, ambient voice, environmental interaction or control loops.

The main examples include:

• robots and humanoids

• drones

• autonomous or semi-autonomous vehicles

• AGVs and industrial mobile systems

• machine-vision and sensing systems tied to real-world action (eg drone detection)

• industrial automation with AI in the loop

There is also a second, somewhat fuzzier tier of adjacent categories, such as:

• digital twins linked to real-world sensors and actuators

• smart cameras or infrastructure with closed-loop automation

• ambient voice systems I discussed recently, embedded into homes, buildings or workplaces

• some wearables and XR systems, if they are continuously context-aware and interactive

You could even argue that some parts of telecom networks, such as antennas with automated tilt mechanisms, fit here.

All of these are likely to have some form of AI “in the loop” in their interactions with the real (physical) world, rather than just an as analytic layer looking at historical data and telemetry.

As a broad category, this is certainly interesting and growing - it’s essentially the AI evolution of conventional IoT and some industrial systems.

But that doesn’t mean that any of it needs 6G, or that 6G’s architecture needs to be specifically designed with these types of devices and applications in mind. In particular, they don’t really align with the main concepts for “AI native networks”.

The Physical AI argument for 6G

The central arguments linking Physical AI and 6G are one or more of the following

• Physical AI will require a new set of network capabilities that only 6G can provide

• 6G can be designed with additional novel features that Physical AI systems will consume at scale and pay for profitably

• All these cool capabilities and features (and inevitable revenue) help justify a complete new AI-native architecture and should be enshrined in 6G standards

Taken together, this becomes an argument for more network capex, more spectrum, more cloud and AI integrated into operator infrastructure, and - inevitably - more pressure on policymakers to “support innovation” by smoothing away inconvenient objections.

In essence this is a way to combine multiple stakeholders’ arguments for:

• Growing uplink capacity and traffic. The claim is that many Physical AI systems will generate large volumes of upstream video, telemetry or sensing data. There is some evidence to believe this, given the asymmetry seen in some enterprise and private network settings. But it does not follow that public 6G becomes the natural or exclusive answer - and also does not account for compression improvements.

• Hyper-low latency: Apparently 6G will offer 0.1 millisecond latency, in much the same way that 5G supposedly offered 1 millisecond latency (It didn’t). I remain doubtful. More importantly, many AI systems have inference or application-layer delays in the tens or hundreds of milliseconds anyway, especially where LLM-style interaction or MCP-based interfaces are involved. That does not make network latency irrelevant, but it does make simplistic “sub-millisecond therefore Physical AI” claims look absurd.

• In-network compute, and especially AI-inferencing edge capability: I’ve written before about the hopes that AI will drive demand for “edge inferencing” and compute offload. Some of that may happen, especially at aggregation points, metro sites or enterprise locations. The leap from that to GPUs everywhere, let alone cell-site or street-cabinet inferencing as a mass business model, is much less credible.

• ISAC (integrated sensing and communications): While I’m definitely interested in adding sensing functions to 6G and other networks, I’m highly doubtful that they will be ubiquitous, or that there are general-purpose business models. They might be good for drone-detection near airports, but I’m not sure that cell-tower radar is much use to stop a car reversing into a concrete bollard, or alerting you to a trip hazard as you walk down the street wearing VR glasses.

• Dynamic network slicing: My thoughts on slicing are well-known. It’s the largest strategic technology error in mobile since IMS. It’s workable for narrow use-cases in specific contexts, with a lot of heavyweight engineering and not much revenue compared to the costs involved. The idea that Physical AI devices will suddenly make 6G-era slicing economically compelling feels like another attempt to revive a weak concept with new branding.

• “Kinetic tokens” and flexible (cough non-neutral cough) rules to allow charging for them. This is the latest pitch from some in the mobile industry. It’s not really about AI tokens in an LLM context. Instead, it appears to be an attempt to define a more commercially-attractive abstraction for contextual, action-linked communications traffic. Basically it is prioritised, interaction-based URLLC rebranded in a way to make it sound billable.

• Timing and positioning: Fine-grained positioning and timing sychronisation are potentially valuable capabilities of 5G Advanced and 6G, but seem more appropriate for localised private networks than wide-area MNO coverage.

• Government funding / subsidy / support to help with “physical AI sovereignty”. Everyone is hoping that governments incentivise 6G with money, favourable regulatory policy, extra spectrum and other sweeteners. Adding in another current and important buzzword to the mix - sovereignty - is appealing, especially if the Physical AI systems are also somehow critical to national strategy or security.

Now clearly, most of these are interesting, and have some bearing on how 6G evolves. Uplink efficiency and ISAC are definitely worth investigation and probable investment. But I’m not sure that they can all be bundled together as a cohesive vision, linked to the equally-nebulous idea of Physical AI.

The Tyranny of Consensus

The really interesting thing about these Physical AI claims is that it creates a rare and unlikely coalition. It manages to unify multiple separate - and sometimes competing - narratives about 5G and 6G - even if all the stakeholders all mean something diferent:

• The AI-Native and AI-RAN “true believers” can point to Physical AI driving the need for compute everywhere, or for multi-tenant device/inferencing workloads on RAN infrastructure, especially with GPUs

• RAN vendors and spectrum advocates can point to uplink demand, lower latency - and RAN-integrated sensing ISAC, as reasons for extra spectrum, new radios and infrastructure capex

• 5G SA Core upgrade proponents can cast Physical AI as the reason to accelerate slicing, positioning, RedCap and other advanced-IoT “bridge” capabilities on the eventual path to 6G

• Operators seeking new monetisation, fed up with flat-rate smartphone data plans and value flowing to app/cloud providers, can hope to better monetise new device classes, especially if they can invent some new basis for pricing, like “kinetic tokens”

• Device vendors get much more access to operators, publicity, tests and trials, event speaking slots and attention from investors. They can quietly ignore whether their products actually need cellular connectivity.

• AI companies get to talk about use-cases that aren’t chatbots, or agent and coding examples that few people outside the industry really understand. Robots and drones are “real”, and the connectivity part can be mostly ignored as someone else’s problem.

• The wider industry (including event organisers and analysts) get a fresh narrative that makes mobile access networks relevant again, in an era where AI and satellite are sucking away attention, capital and people from “normal terrestrial mobile” and even APIs.

• Politicians can put out statements about a “race to 6G” referencing things like “AI, robotics and implantable technologies”.

In other words, there is something of a “tyranny of consensus” here. It’s in almost everyone’s interest to talk up the Physical AI + 6G story. It’s pretty cool and sci-fi, which is necessary to grasp attention, when everyone else is talking about huge rockets to the Moon, or datacentres powered by nuclear SMRs.

Otherwise nobody would be interested in “Yet Another G”.

Subscribe now

Connecting Physical AI systems.. but where and when?

So how does 6G come into the picture, beyond the rhetoric?

Do the technical vectors like latency and uplink and ISAC map to the realities of the new physical AI solutions and markets, exclusively or primarily with 6G as the glue?

The core problem is not just that Physical AI is broad. It is that different classes of system operate in very different connectivity environments:

• fixed industrial cells with hard real-time requirements

• indoor local-mobility environments such as hospitals, warehouses and retail

• campus-scale industrial or logistics estates

• wide-area outdoor fleets in urban or rural settings

• remote or low-density environments such as farms, mines, ports and offshore sites

These do not point to one network architecture. They point to a mixture of wired, Wi-Fi, private cellular, public cellular, on-device compute and local edge.

Most obviously, many of these devices and applications already exist. Robots, drones, XR glasses, even robotaxis are here and evolving far faster than mobile networks. They’re not impatiently waiting around for the 2030s when 6G arrives as a saviour. At best, 6G may later become an optional enhancement, fitting into solutions that are already architecturally hybrid.

Also importantly, many of these Physical AI systems operate in locations which don’t even have 5G, and are highly unlikely to be at the top of the list of priorities for 6G deployment. As you may imagine, the word “indoors” is at the top of my mind.

For every robot tourist guide or police officer outdoors, there will probably be ten cleaning houses or working in hospitals. Same with industrial automation systems, most XR usage and many other products used in factories, warehouses, malls and so on. Obviously drones and vehicles are outdoor-centric, albeit not 100% of the time (underground carparks?).

Many devices will operate in areas with low population and limited scope for MNO coverage, such as farms, ports, offshore, minesm, especially with full features like wide-area ISAC. Will there be millions of Private 6G sites deployed early in the 2030s? I doubt it.

In other words, 6G’s theoretically amazing uplink, low-latency, ISAC, security, slicing, sovereignty and other capabilities won’t be accessible to a large part of the Physical AI user base, a large part of the time.

The fundamental problem is the same for 6G as 5G. It won’t be everywhere (and affordable) for a long time, if ever. Which means that even if you believe in all the wonderful 6G capabilities, developers and customers will need fallbacks and alternatives. And companies will start with the common denominator in most cases, which will then gain scale economies.

Alternatives exist

This comes back to the “assumption of uniqueness” I referenced at the beginning of the article.

Technology engineers are clever. They are rarely limited to a single solution for a given problem. There are always multiple options - and I suspect that engineers using AI will find even more choices. And the other engineers working on the supplier side are equally smart.

Almost all of the 6G capabilities have alternatives, which are sometimes better, cheaper or already available. There are also path-dependencies and architectural lock-ins, either in terms of technology, skills or regulation.

So for example, we will have trade-offs such as:

• 6G vs. 5G (public & private) vs. Wi-Fi 8 vs. fibre or industrial ethernet

• Telco edge compute vs. on-device vs. on-premise vs. neutral edge-cloud

• ISAC vs. cameras vs. LiDAR vs. UWB vs. various other sensor types

• Slicing vs. adaptive applications vs. network bonding

• Kinetic tokens vs. direct device coordination and local control

• Uplink volumes vs. semantic compression vs. AI-enhanced filtering

You get the point. There will almost always be options, unless there’s a legal or regulatory constraint, or a genuinely extreme corner-case.

There is very little that can be expected to be 6G-only or even 6G-primary, which means that 6G will have to compete with, substitute for, coexist with, and sometimes be subordinate to, other systems.

Conclusions

When I look at the rhetoric around 6G & Physical AI, none of the “alternatives” appears to be considered. There’s a simplistic narrative of Physical 6G needs AI, and therefore AI needs to be AI-native and use AI-RAN. But those are non-sequiturs.

The mobile industry has to combat “G Fatigue”.

While 6G technologies, standards and enablers are evolving quite fast, there is a growing split between those who want AI-powered everything, those wanting slow incrementalism from 5G, and those who focus on specific narrow capabilities like RAN performance, quantum-proof security or low energy.

Overall, 6G already seems to be turning into “Yet Another G”, even while 5G has yet to demonstrate economic returns. There are some useful-sounding features like ISAC, better uplink and also some warmed-over existing 5G concepts like low latency, edge compute and slicing that went nowhere, but which people seem to want to try again.

Meanwhile, the satellite and datacentre/AI industries are soaking up capital and attention. There’s a 5G/6G angle to satellite, but that’s dominated by the space players.

Net result: 6G seems to need an injection of cool of its own.

So it has latched onto Physical AI, which is real and growing, albeit as a very broad and heterogeneous super-category of solutions, devices and applications. It’s basically IoT + AI. There are some loud advocates, especially around humanoid robots, autonomous vehicles and drones. And to be fair it’s a much better name than the cringey “cyberphysical systems” of early 6G presentations.

But the problem is that Physical AI doesn’t actually need 6G. It might make use of it, in some places, for some applications and devices, where it’s available and makes technical and economic sense. The rest of the time it will use other networks (especially indoors), rely on onboard compute and AI, use cameras or LiDAR for sensing, deal with variable latency - and seek application-layer workarounds.

Share

But we’re already getting the delusional rhetoric that robots and drones and all the rest will absolutely mandate the use of 6G, and ideally from an MNO network rather than a local private 6G implementation.

It’s cropping up in spectrum discussions, being used as justification for some of the more fanciful AI-RAN concepts, and is trying to be crowbarred into 3GPP standards. There are the usual consultancy reports with too many zeroes in their forecasts.

I understand the desire to create self-fulfilling prophesies. But the problem is that this bias towards “6G-Primary Physical AI” means too little focus on solving real problems such as 6G-Optional or 6G-Hybrid Physical AI, as well as making sure that 6G actually works in the places it is needed for these systems, and delivered by the business models (eg private cellular) that make sense.

None of this means operators have no role. They do. But the realistic role is narrower than the rhetoric suggests: hybrid connectivity, selected private-network deployments, wide-area fleet coverage, trust and identity, some local edge breakout/interconnect, and APIs or orchestration across heterogeneous infrastructures.

That is a useful and potentially profitable set of positions. It is just not the same as claiming that Physical AI is the killer app that makes public 6G indispensable.

Physical AI is a potential 6G use-case in some environments. But it’s not the killer app, and 6G will neither define it, nor gate its capabilities.

6G needs to actually prove itself. It cannot just declare itself the natural champion of robot, vehicle, drone and sensor connectivity.

NGMN, which is a mobile operators’ association focused on network evolution and input into standards, has put out a paper on its views on AI & what that means for 6G, which is a key area of its current focus.

The paper is only 20 pages but covers a fair amount of ground split between:

• Impact of AI-related traffic on networks, especially uplink

• Designing mobile networks for supporting AI services, and in particular any extra features or capabilitgies that might be needed

• AI for optimising / running mobile networks, described as “AI for network and implications for 6G architecture evolution”

(It doesn’t mention MNOs using AI to write code and applications, which I think is a serious omission as in many ways its Ground Zero for telcos+AI. Also completely missing are “GPU” and “spectrum”).

Subscribe now

AI impact on network traffic

The network traffic section is pretty reasonable. It acknowledges the unknowns about where AI models and agents will run and what that means for traffic to and from them.

It notes that models might run on-device, that agents might run on-cloud, that we don’t know what applications will be popular, or if new AI architectures shift behaviours rapidly or locally. It correctly identifies that a lot of traffic will be substitutive not incremental as well.

In particular it argues for flexibility, especially in configuring uplink / downlink ratios. That makes a lot of sense to me, in terms of both a 6G architectural priority & also knock-on implications for things like spectrum policy.

“explore mechanisms that allow semi-permanent adjustments in uplink/downlink ratio without requiring major standard revisions, as well as solutions to enhance the uplink coverage”

“predicting the scale of impact remains highly speculative”

“While some scenarios present significant potential for increased demand, they must be weighed against their likelihood”

There’s also a good point that a lot of the uplink rhetoric (eg realtime streaming from AR glasses) has a huge legal-shaped problem called “privacy”. Good luck getting consent from all the relevant people…

One option I’m wondering about is the potential to either change existing FDD bands to TDD, or just improve the uplink efficiency of the upstream FDD spectrum. It’s not obvious to me why FDD usually has the same bandwidth for uplink & downlink, despite the traffic imbalance. Could be because upstream from devices can’t use MIMO or other fancier techniques today - but maybe “two sided” network AI including device-side elements could help.

Definitely an area for more work, I feel.

Share

Network for AI

This is the weakest part of the report. Rather unsurprisingly, NGMN positions 6G as pretty much central to AI’s future, and doubles down on the 20-year old “QoS myth” claiming that 6G can offer network quality guarantees for AI applications.

In my view, 6G needs AI, but AI definitely doesn’t need 6G.

This follows on from multiple G’s worth of slicing, QoS and similar wishful thinking liked to VoIP, then video, then AR/VR, and occasional fantasy diversions towards autonomous vehicles and medical robots.

The paper tries to align with the worst part of the ITU 6G “hexagon”, which pretends that compute and (mobile) networks will converge, and that MNOs will become de-facto cloud providers. That might be more true in China, but in most NGMN members’ countries it’s just reiterating the Brussels-style myth of a “continuum”.

Distributed and edge-computing architecture isn’t going to be driven or designed by the telecoms industry, let alone waiting 5 years for 6G to start or 10 years for it to be widespread.

This section also seems heavily oriented to fanciful business models.

Token-based charging for AI+networks seems to be a popular (but very vague) idea & we will probably hear more about in coming weeks. T-Mobile was talking about “kinetic tokens” last week, for instance, with the idea of moving AI workloads from site to site (eg for vehicles moving along a road).

However, the phrasing in the NGMN report is much more about “resource consumption” which makes it seem as if they’re just trying to set the scene for the next tedious round of anti-net neutrality and fair-share lobbying nonsense. “Oh, we need to charge differentially for AI-related applications and tasks. So therefore we need to be in the AI workflow. I’m sure all the AI companies will willingly pay for QoS, and we won’t just be trying to gatekeep. Again.”

Another problem here is that the number of tokens is very dependent on the type of model for instance LLMs working with text vs images vs voice vs motion and so on. And various AI approaches aren’t token-based at all.

Another big gap in this section is that there is no mention of network coexistence or alignment. While clearly this is a 6G document, it ought to mention that the real world will involve lots of other networks, and that 6G will need to fit into that landscape. It isn’t “network for AI”. It’s “networks”.

AI for network

This section ought to have been first, not last. This is what “AI native” actually means. Although perhaps this is a deliberate choice. There’s a pivotal sentence in the Executive Summary that highlights divisions in the mobile industry: “NGMN expects that 6G will be AI-ready”.

Not “native”. “Ready”.

Some other telltale phrases:

• “avoids unnecessary impact on the existing architecture”

• “5G Service-Based Architecture (SBA) will be considered as the starting point towards 6G”

• “align with cost and sustainability goals”

• “Continued support for non-AI alternatives”

• “validate the necessity of any network enhancements from a business value perspective”

• “not all scenarios require extreme performance gains”

• “functions that already operate close to the optimal limits …. are not expected to see substantial performance gains from AI-based algorithms.”

However, this line is, in my view, absolutely right and also absolutely what should be standardised in AI+6G":

“6G should introduce an end-to-end data framework to enable efficient and flexible data, model, and inference sharing, management, processing, and storage across UE, RAN, Core Network functions (NFs) and application functions (AFs)”

Data management and model management are in many ways more important than specific use-cases.

But there’s also a really telling omission. Although it says that “6G should support distributed edge computing to enable real-time processing, collaborative intelligence among agents” there is absolutely no mention of GPUs or GPUaaS.

On the other hand, there is also a surprising focus on two-sided operation and on-device compute as part of the overall 6G AI landscape. It’s notable that the AI-RAN view of the world rarely mentions that devices might participate, which might be partly reflective of the lack of Nvidia footprint in smartphones and modem chipsets.

And while there is discussion of uplink or downlink and the potential for full-duplex operation, the word “spectrum”, or how AI might help with spectrum-sharing, spectral efficiency or spectrum-management are all absent.

Coming back to the previous discussion point, the report doubles down on the inspection / analytics function: “6G should support fine-grained traffic analytics to distinguish between model updates, inference requests, and agent communications, enabling optimised management”. Really? Why? And what makes you think the wireless network is in any way special here? Do we see something similar in the fixed and fibre networks that make up the vast bulk of AI connectivity? It sounds like a great way for bad-actors to target specific AI workloads, though.

One thing that is important, especially in Europe, is the explicit mention of the trade-off for intelligence vs. energy. “Network evolution should focus also on aspects such as network simplification and energy efficiency. These two aspects may contradict AI requirements”.

Also good is the discussion about how to evolve the exposure layer with things like MCP.

Conclusions

This is a pretty good document, that I largely agree with, with apart from some of the more eyebrow-raising telcowash bits about token-based charging for “resources consumed”.

There’s more realism about AI-related traffic growth (and unknowns) than many vendor reports. The emphasis is on flexibility to respond to the unpredictable, rather than heroic assumptions on renewed hockey-stick growth in data from AI+AR or AI+robots or whatever else is being pitched.

It’s far more realistic about the opportunities and challenges of AI and “AI native” 6G than much of the GPU-fundamentalist rhetoric that I hear and read about. Instead it’s more aligned with the cautious incrementalism that NGMN has been pitching for a while, anchoring it in the reality of patchy 5GSA rollout, a desire for targeted upgrades rather than huge heavy lifting, CAPEX-centric visions.

That said, there’s some missing parts. The key one for me is a lack of recognition that 6G will have to live in a “multi-everything world”, as I wrote about recently. We will have multiple networks, multiple cloud, multiple models and multiple ways to solve any problem. 6G won’t get to define the way in which networks are architected for AI applications, as that is a continuous process that all the other network technologies have a head start on.

6G will need to be a good telecoms citizen first. Over time it may prove itself to be unique enough to start pitching “6G-centric”, “6G-primary” and even “6G-only” AI use-cases, but that won’t be happening on Day 1 in 2029 or 2030. In the meantime, it will need to play nicely with the incumbents, from 5G to Wi-Fi to fibre and ethernet.

There’s also no reference to GPUs, which I suspect is deliberate, nor to spectrum or the use of AI for internal software development at telcos. Maybe NGMN will come back to those at a later date…

In this episode, Ben & Dean discuss the recent Connectivity Summit event in Brussels and what it told us about the EU DNA Digital Network Act proposals. Dean gives a preview of what to expect in Barcelona at MWC, and also a quick update on the state of Private 5G markets. Benoit highlights an upcoming BEREC event about FTTH and copper switch-off.

00:33 - Last week, Dean and Ben met in Brussels at the Forum Europe event on Connectivity, which brought together regulators, policymakers, telcos, industry associations, vendors and a slew of consultants to discuss the DNA proposals. 

It covered a lot of ground, from satellite and spectrum, to overall telecoms consolidation and investment plans. But while it had a lot to say about 6G and sovereignty, the fixed broadband world and contentious issue of "voluntary conciliation" for IP interconnect and other disputes got much less airtime. We discuss the good & bad outcomes. (Dean's LinkedIn write-up of the Brussels event is here).

16:10 - Dean kicks off the segment what to expect as themes at MWC in Barcelona, starting with Sovereignty, Satellite & Automation. Yes, there will also be a bit about spectrum, 6G and network APIs as well, but they probably won't be the stars of the show in terms of announcements and on-stage discussion. We also dive into the difference of sovereignty for telcos vs cloud. (Dean's next Unthinkable Lab workshop will cover telecom & tech sovereignty & resilience issues on April 23 - Link)

25:25 -  A quick segment on growth of private cellular networks, and the role of telcos in P5G, plus a comparison with Wi-Fi for enterprise wireless. (Dean's write-up of the recent Uptime Private 5G event is here).

30:01 - Upcoming events:

- See Benoit at BEREC's workshop on European Copper Switch-Off (link)

- See Dean at the IWCE event on Critical Comms, courtesy of the Safer Buildings Coalition, on a panel about in-building emergency comms (link) and at vCon Spring (link) to discuss vCons and AI+voice for telcos.