Investment conclusions

BofA thematic strategists define the onset of generative artificial intelligence models such as ChatGPT as a pivotal moment similar to the launch of the iPhone in 2007 (see Me, Myself and AI - Artificial Intelligence Primer).

For the beleaguered telecoms industry, the evolution and integration of AI provides a 'second chance' of sorts to evolve and grow profitability after poor monetisation of the initial growth of the internet and mobile connectivity. Considering multiple moving parts:

• Direct implications - Accelerating demand for data (needs networks): Managing the relentless acceleration of data growth requires a combination of advanced telecoms networks to carry traffic and the application of artificial intelligent to analyse and interpret results. For telecoms operators that means investing in fibre networks capable of 50Gbps upstream speeds to replace legacy copper and cable networks, and 5G mobile networks delivering multiple times better performance than 4G with speeds >500Mbps. Progress is well underway and ahead of the demand curve, and capex looks to have peaked. Beyond traditional infrastructure we regard satellite services as additive to coverage but not a substitute. And nor do we think fixed wireless access is a volume-based product given its poor returns on spectrum investment.

• Indirect implications - Accelerating regulatory change (rebuilds barriers): The need for well invested large-scale networks is mission critical for Governments to ensure economic growth and national security, both materially impacted by the onset of AI. This brings regulation into focus and will surely compel a shift to support investment returns - potentially via consolidation - or smaller scale operators could struggle to survive (with cracks already appearing). We show that fixed network economics are optimum in a 2-player market, but unsustainable in a 3-player market. Similarly, 5G networks are providing a better service with bigger spectrum blocks, 4 players per market looks increasingly unsustainable. An ultimate incarnation of this shift is the move by governments to break telcos up to provide more national influence over the 'NetCo' to influence investment and security implications, but this could realise value for telcos too. Meanwhile 'Big-tech' taxation, currently under consultation by the EC, feels like an unlikely source of support to us.

• Application - Accelerating digitalisation (boosts ROCE): Telco's digitalisation has proven sluggish with digital sales channels still around 25% despite the 'commodity' nature of data. AI is a means to expedite strategies through more automated and relevant customer interaction, (cutting call centre costs), and enhanced on-line sales channels (cutting expensive high street stores and commissions). But also, AI can be a powerful tool to further boost returns through more automated network planning and installation (saving capex), being able to predict and resolve faults in advance and optimise energy consumption (opex). Our analysis suggests Telco could boost ROCE 2ppts from current c6% levels in the event of deep AI integration, this is a key support to surpass cost of capital 'hurdle'.

Operator case studies

Telecoms operators have been vocal on their digitalisation plans over time. All are communicating their intention to integrate AI over time, however we observe increased disclosure and perhaps some advanced integration at Deutsche Telecom, Telefonica and Elisa.

We provide more detailed case study disclosure across all operators later in this report, but noting some commentary from these three:

• Deutsche Telekom: DT is one of the more advanced operators, with an all-IP network supporting its 'Telco as a platform' concept with software layers supporting the integration of AI into the digital journey from network planning through implementation, customer installation and management. Examples of efficiencies include the 10x acceleration of network planning from up to 4months to less than two weeks using AI to scan and plan potential rollout possibilities and further manage the engineering process. Also using AI within network maintenance has reduced downtime by 27% with 31% customers less impacted.
• Telefonica: TEF has established a global AI lab to support the integration of AI into day-to-day operations and is vocal on the use of technology in network planning, maintenance and deployment. And at the group level TEF has >30% of sales now via digital channels with ranges of 30-50% across lines of business.
• Elisa: Elisa is also one of the more advanced telecom operators, with a third of customer processes already automated (flagging a significant cost reduction) & efficient network management helping to keep capex/sales at a maximum of 12%. In addition to progress on own processes, Elisa is somewhat special as its international digital services segment is offering a mix of software services to other telecom/manufacturing operators, sharing their expertise & infrastructure. Accounting for <4% of group revenues, but growing by double-digits, the segment can potentially benefit a lot from advancements in AI. A key requirement remains reception by competitors for those services (we have heard positive feedback), while the scalability of the business is important to improve profitability: Despite past acquisitions of profitable businesses, EBITDA is still negative as investment focus now shifts from research & development more towards sales.

 Direct implications

Franklin wrote that "...nothing is as certain except death and taxes". To that we can almost certainly add data growth as the big data paradigm gathers pace, perhaps even faster than previously expected.

As the volume of data doubles every 2 to 3 years, our vocabulary on data size will need to expand from gigabyte (GB) to terabyte (TB), petabyte (PB), exabyte (EB), zettabyte (ZB), yottabyte (YB), brontobyte (BB), and geopbyte. 

Managing this requires a combination of advanced telecoms networks to carry traffic and the application of artificial intelligence to analyse and interpret results. For telecoms operators that means fibre and 5G technology investment and the decommissioning of legacy copper and cable networks. Satellite services are additive to coverage but not a substitute. And nor do we think fixed wireless access is a volume-based product given its poor returns on spectrum investment & remains focused on less-dense areas.

  Welcome to the age of Yottabytes

The digital universe has reached the level of the yottabyte, with 90% of the world's data having been created in the past two years (source: IBM).

According to IDC, the amount of data created is projected to double every 2-3 years and, according to Statista, in 2020 we created c.64ZB of data, projected to grow to 181ZB by 2025E. Nevertheless, new applications are being developed constantly that require more data. Holograms, metaverse, brain computer interfaces and EVTOL are just some of the technologies that will be very data-heavy and are yet to be launched. Not to mention quantum computing, which could leapfrog the total data creation once commercially available. In short, data could grow even faster than expected in the coming years.

Networks & AI are essential partners - case study connected cars

Telco networks are an obvious factor required to support data growth. But it is not just about data generation, or traffic. It is about data complexity. AI can handle big data complexity problems, for example, connectivity of autonomous vehicles (AV). Intel believes one autonomous vehicle (L-5 level) will generate 4TB of data every day. One connected AV will generate the same amount of data as 3,000 internet users (Intel), However, two cars will generate data closer to that of 8,000-9,000 users, as not only will they generate data, but they will also need to communicate between themselves, and so on. The growth of data just from AVs will be exponential. On this calculation 1,000 cars will generate more data than the entire planet.

In fact, connected car data volumes are already staggering. Wejo estimated 95 petabytes of data generated globally in 2019 from connected cars. And this is the tip of the iceberg. Petrolus estimates that by 2030, 88% of cars sold worldwide will be pre-connected to networks via embedded devices and almost 100% of vehicles in the EU and US will be connected. According to a McKinsey study, >60% of cars will also have low-level autonomy by 2025.

And this is just one small example of the complex data we are creating. Smart cities with populations of 1mn people will generate 200EB of data every day (200mn GB), which will need to be connected, transmitted via telco networks and analysed with different agents. The issue is not just how much data, but the complexity of this data that needs to be processed and transmitted in real time. Thus, smart algorithms collaborating on a large scale will generate data complexity problems that current technologies will eventually not be able to cope with. And this will affect every aspect of our lives, as the number of connectable devices, sensors and machines is expected to grow to >200bn by 2025.

 Network technologies evolving to support growth

Telecom networks are undergoing significant change to meet growing demand for data. In mobile this is more of an evolution as 5G is introduced to support not just bandwidth but lower latency services with more connected devices, into the IoT paradigm. In fixed line changes are more revolutionary as copper networks are decommissioned and fibre networks introduced. We consider each, looking at the improvements versus previous incarnations and how technologies could further evolve over time. We also look into 'alternative' broadband technologies including satellite services and fixed wireless access (FWA, replacing fixed line broadband with higher bandwidth mobile).

Fixed line: from copper to fibre, at the speed of light

Traditional telecoms landline networks were founded on copper infrastructure, connected through the PSTN (public switched telephone network). Initially used to provide voice services, the networks evolved to provide internet access using dial-up broadband, and then through DSL (digital subscriber line) technologies using higher frequency channels to transmit data. At the same time, traditional cable-TV networks also evolved to provide broadband access through DOCSIS (Data Over Cable Service Interface Specifications).

However, an alternative technology emerged in the form of optical fibre, with wire-thin strands of glass used to reflect pulses of light (initially from a laser), representing a digital signal.

How fibre works

Versus cable coax and copper networks that transmit data through electrical signals (electrons), fibre uses optical waves (photons), In short:

• The fibre cable has a transmitter and receiver at each end.
• The transmitter receives data in the form of electrical signals and converts these into an optical wave using laser or (lower cost) LED.
• This beam is transmitted through the fibre - constructed of high-quality glass to avoid absorption, using the principles of reflection.
• At the receiver the wave is then converted back to an electrical signal.

Fibre versus copper

There are two main advantages of fibre vs copper/COAX:

• Speed and latency: transmitting data as photons of light is multiple times faster than vibrating electrons, in fact data moves around 2/3 the speed of light across optical fibre (with some impairment due to reflection). This is significantly faster than what can be achieved through copper. The table below shows that fibre networks - notably the commercial GPON based architecture - are capable of 2.5Gbps transfers and that this is a material shift from previous copper derivations. Latest DOCSIS 4.0 cable speeds are more comparable but new 'Higher-speed PON' fibre standards are now being introduced to support 50Gbps speeds with no significant change to architecture.

Note too that fibre latency (the time required to send data across fibre networks) is lower at c17 milliseconds vs 100ms for cable (c5-6x difference). This should not impact current commercial services such as browsing (low latency) and streaming (higher latency but with buffering). However, gaming would require higher latency as would applications that are time critical such as connected cars.

• Cost: Per the simplified illustration above, fibre is a more direct connection due to its potential to transmit data over longer distances with no degradation of signal. GPON architectures can transmit data cross 10-20km distances. In contrast, copper typically passes through some element of local exchange or - in the case of cable coax - amplifiers to boost the signal that degrades over 100m distances. This adds cost through power (that could be orders of magnitude 5x higher) but also risk of failure that requires truck rolls to fix and replace equipment over time.

• On this subject we also observe the general fault rate in fibre could be 50% lower as it is also less exposed to factors such as adverse weather.

Current rollout, leaders and laggards

The chart below illustrates fibre coverage across Europe at the end of 2021. It should be noted that some of the laggards, notably Germany and the UK, are now seeing rapid expansion with BT looking to pass up to 4m new homes per year, DT up to 3m (boosted by AI network planning, see later).

We illustrate the current stated ambitions of the telecom operators versus more up to date coverage per recent reports. And alongside existing incumbent fixed line build we note the significant input of alternative fibre operators, notably in the UK and Germany:

Mobile: from 'brick phones' to 5G, supporting the IoT paradigm

Cellular mobile technologies have evolved over time, in summary (and with approximate timelines representing commercial availability):

• 1G (1980s): initial voice only communication via analogue signal.
• 2G (early 1990s): adding digital voice services through GSM (Global System for Mobile) and CDMA (Code Division Mobile Access), supporting SMS (short messaging service) and MMS (multimedia messaging service) capabilities between devices.
• 3G (early 2000s): adding mobile data services through CDMA2000 and UMTS to support video calling and internet access.
• 4G (2010s): adding more advance packet switching technology to support higher speed data transfer and, under its LTE guise (Long Term Evolution), provides voice calling over higher frequencies in digital form.

More recent 5G standards, now in commercial deployment from c2019, have evolved to provide a more unified platform to facilitate the Internet of Thigs (IoT) paradigm, comparing to 4G as follows:

• Spectrum: 5G can be applied across a wider range of spectrum including 'milimeter wave' to support super high capacity short-distance services.
• Speed and capacity: 5G should be c500% faster than LTE with a 100x increase in capacity and efficiency.
• Latency: a c10x reduction in latency can support more mission critical applications.
• Unified platform: deployment can expand across 5G more easily as hotspots add to traditional macro networks and devices communicate directly.

Real-world application

Data from network analysts at OpenSignal shows that real world average 5G speeds are proving 4-8x faster than 4G with a median observation around 180Mbps.

However peak speeds, 2-3x faster than 4G, indicate the potential to deliver >500Mbps as a median observation, with a maximum in South Korea approaching 1Gbps.

Telco capex is past its peak, ahead of the demand curve

Telco capex has continued to grow as a percentage of sales with ongoing capacity investment and the evolution of new technologies.

However, the past 2-3 years has arguably seen a 'perfect storm' of capex with a new mobile technology (5G), new fixed technology (Fibre) and digital investment to bridge into external providers for cloud and cybersecurity services, for example. It is not unrealistic to expect mobile to continue its evolution into new technologies (extending 5G and into 6G), but fibre is perhaps the first major iteration of fixed network architecture with a full rip and replace of previous copper architecture. That could be considered a 30-year replacement cycle.

Once built, and with around 2/3 of fibre capex absorbed by civil engineering (digging roads), fibre technologies should evolve much more as a software product, i.e. at significantly lower cost.

 Thus, and with previous charts illustrating a relatively evolved fibre build across telco operators and with 5G also relatively advanced, we expect capex levels to peak in 2021/22 and very gradually fall during 2023/24 before more material fibre completion into FY25/26, e.g. BT guides to £1bn less capex post build at the end of 2026 from its current £5bn expenditure. There are other examples of guidance for capex declines, including Telenor (NOK2bn lower in the Nordics by 2025 vs 2022) & Telia (>SEK2bn lower already in 2023 vs 2022). And both capex/opex should benefit from the decommissioning of overlapping copper networks.

But so too we argue that these technologies are already running ahead of the demand curve, with potential >1GBps fibre speeds more than 10x the average across global households today.

New services are expected to expedite demand for speed and latency in the home with AR and VR-based gaming a major driver. Nevertheless, telco networks look to be ahead of the demand curve for now.

 Alternative cellular technologies are niche: satellite / fixed wireless access (FWA) / Wi-Fi

Much has been made of the potential to substitute fixed line broadband with higher speed 5G mobile services, while the recent launch of satellite connectivity in the iPhone draws questions around the potential substitution of ground-based networks.

However, we take a view that these are not likely to be substitutional technologies en-masse as the volume of data growth in the coming years and need for latency means that neither physics nor economics work. Nevertheless, with topographical challenges, they could have a more niche role to play.

Fixed Wireless Access - diluting ROCE

FWA has been in the headlines in recent quarters following the momentum in US market leader T-Mobile with around 500k adds per quarter. Our view is that this is part related to the shift in US cable as cord-cutting exposes high priced cable operators to commodity pricing of broadband, in which case a home internet product via 5G could provide an alternative.

However, we note the following points relevant to data needs:

• Latency; 5G latency is around 50ms vs. fibre 20ms, both are arguably fast enough for more applications, but less so into gaming/VR.
• Speeds: TMUS points to potential download speeds 33-180Mbps, upload 6-23Mbps. This is clearly uncompetitive versus any COAX or fibre speeds into the Gbps.

But perhaps most relevant above we note the yield per Gb that feeds directly into ROCE, with an incremental $30 home internet service yielding just $12c/Gb at 250Gb usage versus the Magenta Max mobile plan at $80 with average 20Gb usage yielding $4/Gb … 97% lower. Note that a standalone $50 plan versus Magenta essentials at $60 would still compare as $20c and $3/Gb = 93% lower. And even with a 4 SIM offer Magenta max offer the discount would be 77%.

With this in mind our view is that incremental data growth that adds to network capacity is much better monetised through pure-play mobile plans than a fixed alternative, and as such we should expect this product to be more niche - perhaps where fixed broadband is less available. But as a larger scale product it is not economical versus the alternative.

Satellite connectivity - latency is compromised

The recent launch of the iPhone 14 added a feature called 'Emergency SOS via satellite', enabling users to reach out in the event of no cellular coverage through partnership with Globalstar's 24 low-orbit satellites. But clearly this is not even intended to substitute traditional mobile services.

More advanced from a functionality perspective is the recent partnership between T-Mobile and SpaceX using the Starlink satellites to boost coverage in areas out of cellular reach using existing PCS spectrum. But even then, this should be limited to messaging services and potentially images in its initial guise. Potential bandwidth associated with the technology has been reported between 2-4Mbps.

The biggest challenge to satellite phones is latency due to the simple challenge of transmitting a signal to Geostationary satellites orbiting at 36km above earth or even Low Earth satellites at 640-1.1km. This could be up to 550Ms for a Geo round trip, this compares to <50ms for 5G services and <20ms for fibre. Into more time critical applications this is not a feasible technology.

Wi-Fi

Wi-Fi is a more complex alternative as is often not a replacement for wireless but is a 'hand-over' technology at home from the fixed line product. More public Wi-Fi is underperforming 5G already however and especially examples with US millimetre.

Form a more commercial perspective there are multiple potential examples where private 5G networks could substitute Wi-Fi with some countries reserving spectrum for industries. In theory there are speed, latency and security advantages from 5G but so too Wi-Fi standards are evolving to become a more competitive technology.

 Indirect implications

AI could add $16tn to global GDP by 2030, supporting the need for well invested large-scale networks that re increasingly mission critical for Governments to ensure economic growth and national security. This brings regulation into focus and will surely compel a shift to support investment and consolidation as we prove that smaller scale operators will struggle to survive:

• We show that alternative fixed operators need close to 40% penetration of marketable homes passed to deliver a credible IRR; mathematically this limits any region to just two sustainable, overlapping networks.
• Mobile 5G is proving more efficient with bigger blocks of spectrum, not the limited spectrum required to service 4 players. With some stress already showing, markets could see an evolution to 2-3 scale players.
• Spectrum auctions are also likely to shift more to coverage-based obligations than previous financial auctions.

Evidence shows these dynamics are already playing out, potentially rebuilding Telco barriers to entry through scale investment. And we consider the potentially accretive 'end-game' of Governments seeking more network ownership and thus a breakup of the incumbent model to the NetCo/ServiceCo paradigm.

 The incentive: AI +$15.7tn to global GDP by 2030

Most studies conclude that AI will have a significant positive economic impact.

According to Accenture, AI could double annual global economic growth rates by 2035. AI is likely to drive this in three different ways: firstly, AI will lead to a strong increase in labour productivity (by up to 40%) due to automation. Secondly, AI will be capable of solving problems and self-learning. Thirdly, the economy will benefit from the diffusion of innovation.

A study by PwC estimates that global GDP may increase by up to 14% by 2030 (or US$15.7tn) due to AI adoption via productivity gains in the manufacturing and transportation sectors, and due to businesses complementing their workforce with AI technology. This would enable the workforce to perform tasks better.

McKinsey estimates that AI may deliver an additional economic output of c.US$13tn by 2030, increasing global GDP by 1.2% per year. This is to come from the substitution of labour by automation and increased innovation in products and services. AI is likely to create a negative externality in the labour market whereby there is a loss of domestic consumption due to unemployment.

 Channels of AI impact

We discuss ways (ITU, 2018; PwC, 2018) in which AI could impact the economy via production, externalities, and demand.

Production channels:

• Augmentation: AI may change the Future of Work, reshape existing jobs and augment human capabilities enabling workers to be more productive.
• Substitution: AI could substitute factors of production e.g., labour since repetitive tasks can be automated.
• Innovation: Investment in AI can produce economic output through developing new products and services.

Externality channels:

• Economic gains from increased global flows: digital data is a large proportion of international flows in the form of knowledge and information exchange. AI can facilitate more efficient cross-border commerce. Gartner (2017) estimates that AI-based recommendation engines can contribute c.30-40% of sales in leading e-commerce players.
• Wealth creation and reinvestment: increased output from efficiency gains and innovation can be passed on to workers in the form of higher wages and to firms in the form of profits. Wealth generation could create spill-over effects that increase economic growth.
• Transition and implementation costs: AI adoption may incur costs e.g., severance pay, integration costs, hiring new workers to operate AI.
• Negative externalities: AI could displace workers; hence the increased economic activity may cause decline in labour share which puts pressure on employment and wages and could decrease consumption. Furthermore, government support for affected workers may be required.

Of these seven channels of impact, according to ITU (2018) three stand out. The substitution of labour could add c.11% or c.US$9tn to global GDP by 2030. Innovation in products and services could deliver c.7% or c.US$6tn of potential GDP by 2030. Negative externalities could reduce the gross GDP impact by c.9ppts or c.US$7tn.

Demand channels:

• Consumption: increased consumer demand resulting from the availability of higher quality AI-enhanced products and services. PwC (2018) finds that c.58% of the 2030 GDP impact (or US$9.1tn) is likely to come from the consumption impact.
• Product personalisation: AI can enable more efficient discovery of consumer preferences by gathering more data and analysing it. Increased product personalisation can increase the marginal utility of consumption for a given product and increase the variety of products available for consumers.
• Product quality: In facilitating better product personalisation, the value of the product can increase too.
• Time: AI and AI-enabled products could save consumers time and lead to greater consumption. Gartner predicts that in 2018, half a billion users will save two hours a day due to AI.

Global GDP could by up to 14% higher in 2030 (+US$15.7tn) vs 2018 due to AI (source: PwC Analysis). Looking at a breakdown of this increase, demand-side effects are more delayed but increase significantly over time (Exhibit 28). This is due to the longer transmission mechanism from product enhancements to consumption. 58% of all GDP gains in 2030 will come from consumption-side impacts.

 Well invested telco networks a pre-requisite

It is intuitive that a country with well invested broadband infrastructure to support the growth opportunity from data and AI should economically outperform a country without.

Current regulatory stance

On February 23th 2023, the European Commission presented 3 new initiatives, that essentially build on previous efforts made for The Gigabit Society vison of 2025 (at least 100Mbps connectivity for all households) and in line with Europe's Digital Decade of 2030 (at least 1Gbps connectivity for all households).

• Minimizing Capex: A 'Gigabit Infrastructure Act' which put forward new rules to enable faster, cheaper and a more effective roll-out, reducing 'red tape' and enhancing coordination of civil works. With the latter representing up to 70% of the cost of network deployment, significant savings are expected for telecom operators.
• Wholesale fair access: New recommendations to national regulators to allow undiscriminated access for alternative providers to scaled networks, with a bottom-up long-run incremental cost plus fee methodology to set regulated prices for both wholesale access and access to civil-engineering infrastructure. It also incentivises the switch-off of legacy technologies without undue delay (i.e. within 2 to 3 years) which could also accelerate Opex and Capex savings for telecom operators, given the significant hurdle of running and maintaining two parallel networks (copper and fibre).
• Fair Capex contribution: Thierry Breton, Commissioner for Internal Market, has been vocal of his intention to find a way to make global tech platforms contribute fairly to network roll-out costs. Whist this is still at an early stage and lacks a clear perimeter, it should continue to grab headlines given its 'big tech' angle. The potential windfall for telecom operators could be significant given that most of the capacity deployed today is to sustain 'big tech' services volumes.

 Scale could compel consolidation

European telecoms has seen the proliferation of mobile network operators (MNOs) and alternative fixed line operators (altnets) in recent years. Across Europe we now typically observe 4 MNOs in each market, while in fixed there have been multiple examples of altnets emerging to build fibre networks alongside the incumbent networks and cable operators.

However our analysis suggests that scale is key for networks to perform efficiently while delivering sustainable investment returns.

 Mobile: 5G needs more spectrum = less players

Recent research by OpenSignal indicates that 5G networks deliver better performance with bigger blocks of spectrum, or at least bigger than those historically auctioned by governments. This brings into question whether four or even three mobile networks - as is the shape of most European markets - is a realistic mid-term proposition given that spectrum is finite.

Pressure already showing across European mobile operators

Darren Purkis, Three's chief financial officer, added: "Very simply, we're spending more than we're earning. It's unsustainable in the long term."

In fact the pressure on smaller mobile operators is already manifesting, we consider three case studies indicating some stress:

• 1&1: 1&1 announced its intention to acquire spectrum & build the fourth German network in 2019, winning frequencies in the 2GHz/3.6GHz bands during the same year. The operator is building up an Open RAN network (using technology from Japanese telco Rakuten) & planned to lease most of the passive mobile infrastructure, while having national roaming on Telefonica Deutschland's network. A main reason for the planned network build was better visibility of future economics given 1&1's reliance on its MBA-MVNO contract with Telefonica Deutschland before, which was a remedy of the O2D-E-Plus merger in 2014 & resulted in a number of legal disputes. When announcing the plans, 1&1 already had an established mobile subscriber base, which now reached 12m customers. However, despite having all relevant contracts in place the network roll-out is taking much longer than initially expected: By the end of 2022, 1&1 still had only 5 sites in operation vs a regulatory target of 1,000. It blames the delay partly its contractor Vantage Towers, that is supplying some of the passive infrastructure. While a marketing of the new network is still planned for 2H23 & migration of existing customers thereafter, the successive delay & significant network costs have clearly weighed on sentiment: Since the announcement of network plans, shares have fallen by >74% & uncertainty remains as management continued to push forward its presentation of more financial details around the network plan. With network costs already weighing on EBITDA ahead of any migration (E120m in 2023) & capex likely elevated for the next couple of years, 1&1 would see a multi-year transition phase over the next years, if it continues its network build, with a temporary decline in OpFCF and EBIT as network cost benefits take time to come through. Meanwhile, strategic uncertainties remain (1) in the short term about low-band spectrum availability from 2026, which is crucial to the financial targets, & (2) in the longer term whether 1&1 will pursue 100% network coverage or continue to rely on national roaming from Telefonica Deutschland.
• Italy: Iliad was the remedy taker from the merger of Wind and H3G (Hutchinson) in 2016, launching operations in 2018 with aggressive pricing around E6/month. Iliad's strategy was to build wholesale market share and migrate to network economics through its own network over time. However, the cost of doing so has been above initial expectations, Iliad initially guiding to EBITDA break-even in late 2018 but took until Q1 2021. Capital investment was also not helped by competitive spectrum bidding and then some delays to the towers made available by the merger of Vodafone and Telecom Italia towers under operator Inwit. Supply chain issues have further compounded the challenge. In response, we observed Iliad make an audacious attempt to acquire Vodafone Italy in February 2022 that was rebuffed, but Iliad has since committed to being involved in potential consolidation moving forward. Meanwhile WindTre has recently committed to a potential sale of infrastructure assets to support required investment.
• Tele2 Netherlands: Tele2 acquired Dutch spectrum in 2012 and invested to build a 4G network while using T-Mobile for wholesale roaming. However, it took until Q3 2018 to break even at EBITDA. Tele2 and T-Mobile merged in 2017, with the approval by the EC commission referencing the potential failing firm risk at Tele2.

Looking at actual mobile performance

Industry data actually shows on average better mobile performance of three-player markets already. The latest Open Signal data shows that average mobile download speeds in three-player markets stood at 54.0Mbps compared to four-player markets at 42.3Mbps - with the UK, Spain and Italy forming the bottom three.

Looking at coverage, three- and four-player markets actually perform quite similar with no difference in the average, when correcting for outlier Germany.

Considering a move away from high-cost spectrum auctions

An increased focus of governments/regulators on network coverage & performance could also lead to a more favourable position on mobile spectrum auctions. Historically, spectrum awards (e.g. for 3G/4G/5G) have taken place through an auction procedure with billions spent on spectrum frequencies over the last decade. The industry often argues that this money is missing when operators undertake their network investments & that regulators should focus more on ensuring the best infrastructure, rather than generating proceeds. Over the last couple of years, we have already seen some regulators enhancing their auction procedures to incentivise more build-out. While for example coverage requirements have been an element of spectrum auctions for some time (as an underlying requirement), incentives could be built more efficiently while telecom operators could feel some financial relief. Recent examples include:

• France 2020: As part of the 5G auction in the 3.5GHz band, France conducted a mixed award process. The first 200MHz were allocated directly to the four telecom operators, contingent on a number of commitments. These frequencies were allocated at a fixed price of E70m per 10MHz, before the remaining 110MHz were then auctioned in a competitive bidding - which resulted in a much higher price of E126m for each block if 10MHz.
• Austria 2020: Austria allocated frequencies in 700MHz, 1.5GHz & 2.1GHz in 2020 and went one step further in creating a competitive bidding for commitments. In the first part of the auction, operators were bidding as usual for the spectrum allocation, while the coverage of 900 underserved communities was an additional requirement for the winners of the auction. However, in a second step operators were able to 'bid' for the coverage of further underserved areas with successful bids reducing the costs incurred in bidding phase 1. As a result, the number of covered communities in the end increased to 1,702.
• Germany 2024? Frequencies in the 800MHz, 1.8GHz & 2.6GHz bands will come up for renewal from 2026 in the Germany market, with an allocation expected from 2024. The latest position paper by the German regulator shows that focus is shifting towards more roll-out incentives rather than just maximising proceeds. The regulator considers a number of options (the final auction process is not yet decided): These include (1) using coverage commitments as bids, (2) using coverage commitments to determine rebates, (3) connecting investment commitments to bids (i.e. the bid needs to flow into network investments) or (4) negative auctions based on the lowest costs for covering specific underserved areas.

Extensive bidding on spectrum frequencies in the past has weighed heavily on telco sector ROCE, while operators still had to pay capex to invest in mobile infrastructure. Combining these two elements could lead to an improved investment environment in our view.

And/or a potential move to fewer bigger blocks - case study Singapore

In 2020 Singapore awarded just two 100MHz blocks of 3.5GHz 5G spectrum in a 4-player market. Singtel was a winner but Starhub and M1 chose to set up a joint venture to acquire the second block, while TPOG did not acquire a license. The two winners were obliged to cover at least 50% of Singapore by end 2022, and full coverage by 2025.

There are also some examples for sharing mobile spectrum in Europe, particularly in the Nordics: In both, Sweden (Telenor-Tele2/Telia-Tele2) and Denmark (Telenor-Telia), operators have increasingly moved towards sharing not only mobile infrastructure, but also spectrum & are bidding jointly when frequencies are awarded.

 Fixed: 2 networks is the optimum footprint

With the onset of fibre technology multiple alternative operators (altnets) to the traditional incumbent fixed and cable operators have committed to smaller scale regional rollouts. We highlight the UK and Germany more recently as countries where the aggregation of multiple altnet builders creates a significant 'overbuild' besides the incumbent operators.

However, with the onset of inflation and higher borrowing costs there have been some signs of stress in the industry:

• In Germany: Liberty recently confirmed the exit of its HelloFiber operation in German, filing for insolvency and citing "changed macroeconomic conditions [including inflation, interest rates and access to capital]". Glasfaser Direkt, another FTTH altnet, also filed for insolvency recently having previously committed E1bn in investments.
• In the UK: The biggest altnet, CityFibre, recently announced redundancies of around 1/5 of its workforce. Since then in March press reports indicated UK cable incumbent VMO2 could pursue a combination of its network infrastructure with CityFibre to build fibre.

The latter is potentially a function of increasing financial pressures at CityFibre who targets 8m homes in the UK by 2025. But perhaps more of a driver is the lagging build at VMO2 with only 1m cable homes overbuilt and another 1.5m expected during 2023 versus a target of 15m by 2028. To meet this target would require 12.5m homes passed in 5 years = 2.5m homes per annum. BT is currently running around 3.5m homes as a fully operational and national fibre operator, thus we doubt if VMO2 has the scale of engineering resource to meet its targets. Given the rate of BT build and, more crucially, overbuild we think an inorganic solution comprising CityFibre could be credible.

Nevertheless ad hoc commentary from wider operators also supports a view of an industry in distress, with US infrastructure investor DigitalBridge publicly expressing an interest in consolidating operators at reduced valuation levels as macro conditions become more challenging.

Case study: Generic fibre altnet

UK operator VMO2 recently discussed the potential need for Altent wholesale network load of up to 40% to provide a return on capital. We look closer at the viability of altnet models in a rising-rates environment (and with arguably some cost inflation as well). Hence, we have built a simplified altnet model, looking on IRR and NPV based on a number of assumptions:

• Network roll-out, penetration & revenues: We model a total of 1,000 homes passed at an annual run-rate of 100 (i.e. complete after 10 years). We expect penetration of each passed home at 10% in year 1, 20% in year 2 & steady-state penetration of 30% in year 3 as our base case (that we subsequently flex to prove/disprove the VMO2 statement). Hence, full network penetration is reached in maturity year 10 + 2 years = year 12. We then expect wholesale revenues of E20/month for each home taken up (c2 years payback, if taken up) and apply an annual +1% inflationary adjustment.
• Capex needs: We assume capex of around E500/home for the roll-out. Post roll-out, we model c10% maintenance capex/sales. What's the precedent? Note that capex costs vary significantly by country/roll-out methodology with some in the low hundreds (e.g. Spain/Portugal) vs >E1k for example in Germany, Switzerland & the Scandis (and even much higher in some rural areas). Wholesale rates should be however higher in those high-cost countries, accounting for this.
• Margin profile: We assume a contribution margin of 70% in our base case & assume a tax rate of 25% (post interest & D&A, so limited over the first 25 years). What's the precedent? There are not many pure-play NetCos out there, but a couple that report their margin profile: (1) TDC reports an EBITDA margin of 68% for 2022 & an EBITDAaL margin of 64%. (2) Telenet indicates a 76% adjusted EBITDA margin for its NetCo. (3) Telecom Italia targets to reach a 50% EBITDAaL margin for its NetCo, up from c40% expected for 2025. The examples of TDC/TI in our view show that there could be even some downside risk to our base case on an after-lease basis.
• Financing: We assume that out of a total investment volume of E500k around one quarter is financed through an initial equity infusion (E125k) and the remainder based on debt. The implies a peak of leverage at around 5.5x net debt during the roll-out phase and a return to a lower end of 4.0x in the long-run. In terms of financing costs, we base our assumption on 10-year EURIBOR +150bps, which implies a cost of debt around 4.5%.

We show the main elements of our base case below: The roll-out takes place for the first ten years with full penetration from year 12. From year 11 onwards, the unlevered FCF profile changes drastically due to the lack of roll-out capex and turns positive.

At the equity level, we assume that after the initial injection of E125k, there is no dividend pay-out until the NetCo comes back down to 4.0x leverage from year 11, when returns start to flow back to investors. For simplicity, we thereafter expect re-leverage each year to 4.0x & all excess cash flows to be distributed to equity holders.

On an unlevered basis, cumulative FCF reaches break-even from year 17 (Equity FCF already from year 14 only due to re-leveraging).

However, we believe that the key factor for equity investors is the NPV/IRR of the investment given the cost of equity and the long time horizon of the returns. Assuming a cost of equity hurdle rate of around 10%, we reach a positive NPV of E33k (including E125k as the initial investment). Similarly, our model results in an IRR of 11-12% based on the assumptions above.

Having said that, results are highly sensitive for changes in assumptions - particularly the wholesale penetration: As we show below, NPV turns negative for any double-digit required rate of return when assuming that penetration is only 25% instead of 30%.

We show a similar sensitivity for IRR, adding the cost of debt as an additional variable: We find that at any penetration level, an additional percentage point in interest rates (all else equal) reduces the IRR by around -0.5%. Hence, we also see difficulties to reach double-digit IRRs with a 25% penetration rate with an interest rate of >4%.

And note too that with around 15-20% of homes currently mobile only (and a potential beneficiary of niche FWA products), 30% penetration is arguably 30/80% = 37.5% of marketable homes passed.

Illustrating this:

Altnet outlook - Implications of broadband market structures

We believe there is a clear implications from our finding above: Ideally, the number of fibre broadband networks should not exceed two in any area & the outlook for fibre altnets is difficult in areas that are already penetrated by the (copper/fibre) incumbent and a cable operator as well. Our thought process below:

• Not all households take fixed broadband - margin of error is thin: Not all households take broadband services, e.g. in Germany the broadband internet penetration is still <90%. This can be a result of demographics, but also some households being mobile-only. Implication: A required penetration of 30% (per above) actually implies a market share of at least one third. Hence, with three networks that would imply the need for equally-split market shares - however with no room for deviation.
• Incumbents (and cable) should have a retail advantage: We assume that in areas in which fixed incumbents (and cable operators) already have established a customer base, fibre altnets face an initial disadvantage & need to catch up on customer awareness. If the competition sticks to (old) copper/cable technology, this should over time erode and the altnet should benefit from its technology advantage. If however the existing retail players catch up relatively quickly with fibre roll-out themselves, we believe there is a risk of not reaching retail penetration targets: Competition can build on its existing customer base (to migrate to fibre) & larger players likely have a cost advantage (sourcing/administration), making pricing potentially more competitive.

Implications for altnets? We believe the case for altnets is difficult, when rolling out (1) as a third network with (2) existing retail players having credible plans to upgrade their footprint to fibre.

What areas can altnets target? Having two competing (fibre) infrastructures in each area appears to be a viable option in most areas based on our standardised model. In our view, this can indeed back the case for fibre altnets, but only in specific regions:

• Areas with no cable footprint: Nearly half of the European market is covered by at least DOCSIS 3.0 & around one third by DOCSIS 3.1. Per 'End of the cable honeymoon', 13 Sept 2022, we believe there is a clear need for these cable operators to upgrade their footprint to fibre-to-the-home. Assuming a fibre roll-out by the competition takes place, the altnet would face direct competition from two established players: The incumbent & cable. Markets with relatively low cable penetration include Italy, France, Sweden & Finland vs the Netherlands, Belgium, Switzerland & Germany at the upper end.
• Rural areas with prohibitive cost to build/a clear speed advantage: (Very) rural areas, which often do not have cable penetration anyway, likely remain in focus as well. There, we believe the case can be amplified by two elements: (1) Government subsidies for the first operator to build in very rural areas are in place in many countries. Securing those as an altnet can provide a clear cost advantage & completely different economics vs followers. (2) Underserved copper areas, in which fibre upgrade by the incumbent takes time, can give the altnet enough time to establish its own retail take-up. Securing contracts with customers even before the roll-out can improve visibility (as done e.g. by Deutsche Glasfaser in Germany).

 The consumer is not an obvious casualty

Any potential shift in regulation is unlikely to weigh heavily on the consumer. Previous studies have shown that expenditure on telecoms services comes in around 2% on average.

As a potential measure of how much utility is derived from this spend, statista notes that the average person in the UK spent 5 hours and 47 minutes per day access the internet via any device. Bearing in mind the average person spends around 7 hours per day asleep (YouGov statistics), then for just 2% of spend, the average user is absorbed for around 1/3 of their day.

 Taxing tech - too little / too late

EU Commissioner Thierry Breton recently launched a consultation on a whether regulators should charge tech companies for their use of telecoms networks in Europe, known as the "sender fee", paid direct to the telco operators.

The logic underlying this is that a significant proportion of internet traffic is dominated by a small number of tech companies, notably the digital content platforms Netflix and YouTube. And that they pay nothing towards the investment in the platforms to carry the traffic.

Per commentary in a joint statement from the CEOs of Telefonica, Deutsche Telekom, Vodafone and Orange:

Our view - too little, too late

There is clear momentum at the EU level and as the telco management teams point out, other markets are considering a similar levy, notably in South Korea as one of the world's leading FTTH markets following the surge in traffic from the series 'Squid Game'.

However, this seems a flawed argument to us; content providers are not 'pushing data', consumers (of broadband services) are 'pulling' data and paying broadband services directly. And we have some sympathy for the argument that there is a symbiosis to both parties' involvement; that without content providers there would be no demand for telco network usage growth, but without telco investment there would be no platform to stream traffic … etc.

Ultimately, we take the view that regulatory interference is a bigger driver of telecoms poor return on capital over time, catalysing more competition and thus price deflation that meant that the telcos were not able to monetise capacity growth and thus capex/sales ratios have risen and squeezed cash flow. We take the view that regulation must understand this dynamic before a retaliatory approach through taxation.

And thus, we do not forecast any benefit from this potential levy in our financial modelling.

 Wider national interests could pre-empt Telco breakup

The extrapolation of AI into an increasingly and eventually wholly digital society presents multiple challenges for regulators and governments.

 Cyberattacks: could be used to generate malware

Human-written defensive software may not be enough to combat AI-generated malware which has drastic implications for cybersecurity. Our US Aerospace and Defense Electronics colleagues have a negative outlook for ChatGPT in terms of cybersecurity: The weakest link in the cybersecurity chain, regardless of how much companies spend, are the human users. Many foreign attackers struggle with English as a second language, and their scams can be relatively easy to spot. Forbes reported cybercriminal have already caught on to these use cases and showcased malicious code written by ChatGPT which can be used to exploit victims.

Ethical concerns: accountability, privacy, IP

• ESG: In January 2023, it was reported that, to develop ChatGPT, OpenAI used outsourced workers from Kenya earning less than US$2 per hour to label data that had toxic content. This was to prevent the application from generating violent, sexist, and racist remarks.
• Privacy and misuse of data: The model could accidently reveal sensitive information and the output can be misused e.g., tracking individuals. If data is misused, then it could be the case that the model violates privacy laws e.g., EU's GDPR. Implemented in May 2018, the EU's General Data Protection Regulation seeks to give natural persons 'control of their own data'. This legislation aims to give consumers various rights including protection of their personal data, to be forgotten, and to data portability.
• Accountability and ownership: e.g., students have been using ChatGPT to complete assignments. This also leads on to a question of intellectual property and copyright - should the chatbot be attributed or the developer or the source data? Perhaps even the user since they still exercise discretion.

Whether generative AI is eligible for copyright protection varies country to country. For example, in Korea, the Copyright Act protects works that are 'creative productions expressing human thoughts and emotions', which means that since AI is nonhuman, it cannot exercise legal rights. An amendment to this act has been proposed to recognise the copyright of generative AI but it has been pending for three years.

In the US, copyright laws do not protect work that has been solely created by a machine but if there was substantial human involvement then it could receive copyright protection (source: AI Multiple).

In China, the creation of products by generative AI cannot be copyrighted, yet Tencent is entitled to copyright for an AI-generated article.

 Emissions: significant computing power to train the models

A large language model like GPT-3 requires a significant amount of energy and computing power to train it. This is due to the limited memory capacity of even the largest GPUs, which means that multiple processors must run in parallel.   It would take 288 years for a single V100 Nvidia GPU to train GPT-3. Training an AI model creates more 57x more CO2 emissions than a human generates in a year.

Could Governments seek more influence?

As a potential future development, we question whether political sensitivities to the factors defined above could compel a desire to have more influence over national broadband infrastructure. Today Government influence, whether through a direct holding or state controlled/funded entities remains significant across the sector.

And over time we have witnessed multiple efforts by foreign entities to control telecom fixed line incumbent operators that have been rebuffed or at least investigated by political figures, while other Governments are now working more proactively to gain influence. Considering some examples over time:

• KPN: Back in 2013, in response to an effort by Mexican Telco America Movil to acquire Dutch operator KPN, the Dutch Minister of Economic Affairs wrote that an acquisition by a "foreign company" could have consequences for national security.
• BT Group: More recently in 2022 the UK Government exercised its rights under the 2021 National Security and Investment Bill to investigate the acquisition of 18% of BT Group by French investor Patrick Drahi. The act gives direct power to review transactions that could "reasonably suspect give rise to or may give rise to a risk to national security", and in certain circumstances would empower the Business Secretary to "impose certain conditions, block or unwind it completely". The investigation was found to give the Govt no undue concern, however the Bill reserves the right to review any increase in stake holding by Mr Drahi.
• Telecom Italia: In response to a long running saga involving multiple owners of Telecom Italia over time, the Italian Government is currently involved indirectly through state lender CDP in a process to acquire TI's fixed line infrastructure. Industry Minister Urso was quoted saying "We need the network to be under public control", quoting the Governments ambition to speed up the digitalisation of tis economy.
• Government stakes in telco (incumbents) are still common: Particularly with a view to their past as state-owned (fixed-line) monopolies, most European incumbent operators are still partly owned by their respective governments or their investment vehicles. These ownership stakes add an additional interest/influence on top of local legislation. Telco operators in our coverage with government ownership stakes include: Deutsche Telekom (Germany), Orange (France), Proximus (Belgium), Swisscom (Switzerland), Telecom Italia (Italy through CDP), Telenor (Norway) & Telia (Sweden).

In addition, we have observed more widespread moves to outlaw Chinese Vendor Huawei from mobile infrastructure with multiple Governments forcing the replacement of associated equipment from sensitive core areas of the network and wider 'edge' network over time.

Break up?

Thus, the debate around telecoms conglomerate operations comprising networks through service layers is likely to come under increased scrutiny over time as Governments potentially seek more control and influence over communications infrastructure, notably fixed line.

For equity investors this may not be the worst outcome, consider:

• Operationally: More and more telco operators are considering the sale of passive infrastructure assets, taking the view that these are not the direct driver of differentiation.
• Regulation: There is a view, and one directly expressed by Telecom Italia in its recent restructuring, that incumbent fixed line regulation 'infects' the service layer, preventing price and product agility.
• Valuation: infrastructure assets are proving to trade at a significantly higher valuation level when stood alone than within the telco conglomerate, with notable mobile tower transactions at 26/27x EV/EBITDA versus current telco sector operator valuations closet to 7x.

 Case studies Telecom Italia and BT Group

As potential case studies we consider events at Telecom italia where the board has committed to review offers for its fixed network operations (NetCo) from a consortium comprising state lender CDP, Cassa Depositi e Prestiti, and Macquarie versus another offer from KKR. Recent press articles have suggested a valuation of E20-22bn that could enable the residual 'ServCo' to de-lever and potential accrue value through a re-rating and/or cash distribution.

   Similarly at BT .Group, previous press reports debate by management have suggested a spin-off of BT's Openreach network division could unlock value to shareholders with precedent asset valuation above current BT's own.

 Application

The digitalisation of telecoms as a concept is not new, but progress has been slow. Even today and despite the digitalisation of retail that has seen the growth of Amazon et al, telco remains an industry dominated by high street retail sales with only around 25% of sales through digital channels. Efforts to automate customer care have been slightly more successful with the onset of chatbots. In this section we conclude:

• AI provides a potential catalyst to expedite these opportunities and further embed itself into modern network architectures to support network planning, deployment, and efficient operations.
• We estimate that actions could deliver 2ppts of ROCE as a vital addition to current telco ROCE around 6%, pivoting into +ve territory vs. a c7% cost of capital hurdle.
• Observing operator case studies we think Deutsche Telekom, Elisa and Telefonica are amongst the more advanced players that could benefit sooner from initiatives.

 AI: Digitalising sales & customer care

Digitalising the customer journey in both Consumer & Enterprise confers multiple potential benefits that could drive revenue growth and cost reductions, in our view, and these look set to benefit from the application of AI tools and techniques.

Direct opex savings derive from reduced human interactions and expensive high street retail footprint, while more indirect benefits could accrue with lower churn as customer care is improved. Revenue benefits could derive from better understanding of the customer needs and usage that could apply more suitable products and upsell additional services, although we are more hesitant to attempt to quantify this as there could be an opposite effect too as customers out of contract or on legacy products are migrated to more suitable products that could even derive lower spend.

Current state of play

Current digitalisation of customer interactions varies significantly across the sector and disclosure is limited & not like-for-like. However, as operators realise the operational benefits from increased digitalisation, some have published explicit medium-term targets (e.g. DTE, Proximus, TI & Vodafone). We note a few key observations on the current state of digital sales & customer care and the announced targets:

• Most sales still non-digital: Most telecom operators still conduct the majority of sales in-person (e.g. in a store or over the phone). For the larger telecom operators/incumbents we note that the share of digital sales generally ranges around 20-30% of total volume. At the lower end, we find operations such as Deutsche Telekom's EU business outside Germany in the teens (but already in 2020), whereas Freenet already has around half of gross adds online (the share was even higher during the Covid-19 pandemic).
• Higher digital adoption by challenger brands, possibly due to a different customer profile: Although with limited disclosure, we find a higher share of digital sales for challenger brands compared to premium operators. (1) Swisscom reports that >60% of sales for its second brand were done digitally vs only 11% for its first brand. (2) Freenet shifted the majority of its gross adds online during the Covid-19 pandemic & still has around half of its sales online. Management commentary by Freenet also indicated some relationship between the online share & the share of lower-ARPU SIM-only sales (without devices). However, this does not necessarily imply a lower relative/absolute margin for those customers.
• Domestic operations appear to be first digitalisation targets: Deutsche Telekom & Orange have both reported the share of digital sales for their domestic business compared to the rest of Europe - with both operators at a firmly higher online share in their home countries (24%/23%) vs the remaining operations (13%/18%), there could be a read-across for the wider sector. A potential explanation could be the initial focus on better operations/efficiency in the home & main market. Having said that, regional differences (e.g. non-domestic operations in developing countries) could also play a role.

• Higher digital penetration in customer care: Generally, we see a higher degree of automation in customer care than for direct sales. Only few operators report 'like-for-like', but we see (1) Proximus with 25% automated customer care vs 20% online sales and (2) Swisscom with 65% customer care online for the own brand vs 61%/11% of sales for the first/second brand. Swisscom is a good example of driving customer onto the online experience, now offering monthly discounts for online-only subscriptions. Other datapoints also suggest a somewhat higher online penetration in customer care with DTE reporting 30% of calsl shifted to digital and Elisa with >30% automated consumer customer processes.
• Digitalisation targets highlight room & ambition for growth: Having said that, a number of operators have already announced targets to improve digitalisation in both, sales & customer care. Examples include Deutsche Telekom, Proximus, TI & Vodafone, which all gave medium-term targets to raise the digital exposure. In sales, the targets range mostly around a 30%+ share (compared to 20-30% at the moment), while in customer care the digital share targets are more around 40%+ (e.g. for Proximus or DTE's calls).

Potential benefits

As discussed above, a further move towards digitalisation should improve both, customer services and cost efficiency. A number of case studies by telecom operators can help to understand the potential scope for improvements.

• Elisa: As part of the Elisa capital markets day in 2023, Elisa reported on its current state of digitalisation and the impact on the consumer business. (1) The operator found that for calls, customers actually prefer some voicebot interaction first (which can make a resolution more efficient), scoring higher than human-only interactions. (2) Elisa showed in its presentation that automation of customer processes by +9pp (to 33%) over two years was actually able to reduce the costs for the same by -8% over the same period.
• Deutsche Telekom: Deutsche Telekom has probably been most open about its current state of digitalisation & its targets. For 2024, it announced a number of ambitions during its last DT capital markets day in 2021. Targets include not only a digitalisation of sales/customer care, but also improvements in network/IT performance & broadband experience. Not splitting up the savings between the different segments, the operator however expects an EBITDAaL impact of >E300m over 2021-24.

• BT: In its Consumer business briefing in 2022 showed the benefits from more efficient customer management on the cost structure. Over FY18-22, the operator showed that a reduction of service minutes (per customer) by -31% was accompanied by a reduction of frontline staff by -14% and of retail stores by -20%. With more cost efficiency, the operator was able to achieve a higher number of customers per service agent (+33%) and hence a higher profitability per employee (EBITDA up +18%).
• Telenet: Belgian operator Telenet has also updated on the current status of its automation push at the most recent capital markets day in September 2022. It showed that over 2019-22, it was able to improve its digital sales share by +10pp to 33% and nearly doubled its self-installation share to 67%. During a similar timeframe (2018-21), the operator reported a reduction of human service interactions by -24% - linking that to an opex reduction for IT/residential customers by -15% as well.

• Vodafone: Looking back to September 2019 and Vodafone's 'Digital First' event, the company identified 50% of Group opex at the time that could be affected by digital initiatives and set operational targets for the years ahead:

Progress has been variable:

• Digital sales channels missed the 40% FY 21 target by some degree and remaining some distance away at just 26% at H1 23 (+2ppts YoY).
• Customer care via chatbot looks more encouraging, but cannot be measured clearly against original targets, now quoted as a completion rate which refers to the number of interactions solved without any human contact.
• There has been no formal update on whether the ambition to cut stores has been achieved, but the limited change in digital channel sales may indicate it has not.

This could indicate that a firmer commitment to achieve original digital sales targets (supporting fewer stores and lower commissions) could derive net benefits beyond current expectations. Running the numbers is less obvious, however assuming these costs are proportionally equivalent, then adding 14% to digital channel sales to achieve VOD's original target could reduce costs around E5-600m, c. 3% of current EBITDAaL.

 AI: Optimising energy consumption, network planning & performance

AI's predictive capabilities can also help manage energy consumption across networks, and serve as a pre-emptive factor in network planning, deployment and operations.

AI & Energy consumption

Operators are increasingly utilising AI to optimise network energy costs that could account for up to 95% of total energy consumption (ref Vodafone). Underlying this Nokia estimates that within mobile networks, only 15% of consumption is ultimately used to transfer data, with the remainder absorbed by heating and cooling:

Case study - Equipment manufacturer

Nokia provides a solution to energy management through its Nokia AVA service that claims up to 30% of energy saving and reduced CO2 emissions for telecoms radio networks. Underlying capabilities include:

• Dynamic shutdown of unused active and passive elements in low-traffic situations: Using AI to learn and predict traffic patterns rather than just fixed time slots. Alternative patterns can be evolved across rural and suburban footprints.
• Automated control of the antenna angle: reducing the need for physical intervention.
• AI powered energy management across active radio and passive equipment: Using AI to identify patterns and trends to establish 'benchmarks' for monitoring.
• AI powered cooling: active sensors can predict cooling needs, with a potential 70% reduction in cooling costs.
• AI savings simulation: Using AI to predict the effect of changes, and where these could be more relevant or where modernisation could apply.

As an illustration of energy saving, the following chart shows how the Nokia service can optimise energy usage with more dynamic power up/down time:

Case study - Operators

• Vodafone 2023 is using a big data analytics platform, developed in-house, to drive energy efficiencies across its owned 11,500 radio base stations. Artificial intelligence and machine learning help energy specialists identify 'consumption anomalies' - parts of the network using more energy than expected. Targeted actions can be then delivered to make the sites more energy efficient.

Another initiative - the Vodafone Smart Sites programme - uses Internet of Things (IoT) and AI on Vodafone's highest energy-consuming radio base stations, enabling it to manage the sites remotely and reduce the need for engineer visits. As of March 2023, 1,300 sites have been connected.

• Telefonica 2023: Telefónica Spain has been the first operator worldwide to test Ericsson's Radio Deep Sleep Mode energy saving functionality, tested in Madrid at a site with 5G deployment configuration. Supported by Artificial Intelligence and Machine Learning algorithms, the company achieved savings of up to 8%, considering the site's total 24-hour consumption, and up to 26% in low traffic hours. In addition, the Micro Sleep Tx functionality was tested in 4G and 5G technologies in the city of Talavera de la Reina. With continuous operation throughout the day, an energy reduction of up to 16% was achieved.
• Tele2 2023: The Swedish operator has led the European research project investigating how AI can be used to reduce energy consumption can be reduced in mobile networks. The three-year long study analysed data submitted by telecom operators, resulted in live testing of the energy savings in the 5G network & showed that smarter mobile networks could reduce energy consumption by 30-40% in the long term.
• Elisa has gone one step further & other operators (such as DTE) are pursuing similar initiatives: Transitioning from own energy-saving AI applications towards commercial energy management for third parties. During its capital markets day, the operator disclosed an AI-enabled distributed energy storage start-up, which is generating an additional revenue stream for the group. It generates revenues through stabilisation services & cost saving opportunities from load-shifting, having 14 patents granted/pending and assuming a payback time of 2-3 years. How big is the potential opportunity? Elisa estimates the European telecom energy storage market to reach E1bn by 2030.

AI & Network operations

In a study conducted by Ericsson on the application of AI, capital expenditure reduction was the dominant focus of service providers. Considering the following benefits identified by advanced fibre operator Telefonica in its annual report (we summarise):

• Network planning: based on current usage and traffic or load forecasts implemented automatically with AI, it is possible to size network growth needs more accurately, with greater granularity and faster than before.
• Testing and deployment: There are more and more pieces in the network, with more vendors in play and more software components with new versions available at shorter intervals. Here, AI could help in recommending the set of tests to run and the best time to do it, as well as could help to write the code of these tests automatically.
• Network operations: work is being done on automatic real-time monitoring, where AI is capable of detecting anomalies, extracting patterns and identifying the root cause of a problem or degradation in network performance automatically, both by identifying the point in the network presenting the problem (radio access, transport network, core, etc.) and the specific casuistry (interference problems, congestion, non-optimized configurations, high load, etc.). Then:

1. AI can propose possible solutions, with the goal of self-repair in those cases where the problem can be solved remotely, automating this task.
2. AI-enhanced automation can streamline diagnostic and self-repair processes in the most common and repetitive cases, allowing experts to focus on the analysis of more complex or less common problems that AI is not capable of solving today.
3. AI offers the possibility of predictive maintenance, anticipating the occurrence of events or failures in the network and trying to prevent them or at least having time for planning a faster solution.

• Energy efficiency: new automation capabilities allow the network to be reconfigured to vary consumption according to traffic. AI at this point can predict the expected traffic at each node by time slots and recommend the best configuration parameters to achieve savings in consumption.

 AI: Modelling the benefits

We analyse the potential benefits of AI and their impact on telecoms ROCE, subdividing the following categories:

Sales & customer care

We aim to quantify the direct cost impact from digitalising the customer journey and find some common ground on savings potential based on the case studies cited above.

• Deutsche Telekom - Germany/EU: -0.3% annualised reduction in opex. Guidance of >E300m in EBITDAaL gains over four years (2024 vs 2020) through digitalisation come mainly from Germany/Europe. Different to some other operators, they include savings from better network/IT/provisioning. Based on reported 2020 numbers, the savings equate to 1.4% of segmental opex/2.3% of EBITDAaL. Hence this implies annual opex savings of -0.3% & EBITDA boost of +0.6% with an increase of digital sales by >2pp per year.
• BT - Consumer: -0.3% annualised reduction in opex. BT indicates possible savings of GBP60m over three years from digitising the Consumer segment - equivalent to 0.8% of the segment's opex & 2.7% of EBITDA. Hence, annualised savings are -0.3% with an EBITDA boost of +0.9%. A driver would be to reach digital sales penetration of 50%. Assuming that BT was largely in-line with other incumbents, that would imply an improvement by +8pp each year.
• Elisa - Consumer: -0.3% annualised reduction in opex (estimated). Elisa indicates that it increased automation of consumer customer processes by +9pp over 2020-22 (to 33%) & at the same time reduced costs by -4.1% in both years. Assuming proportionate savings in the customer opex (based on a Vodafone precedent of c7%), this implies a direct annual opex reduction by -0.3% & EBITDA boost of +0.4%.
• Telenet: -0.3% annualised reduction in opex (estimated). Telenet reported a reduction of opex for IT & residential customer operations by -15% over 2018-21. Assuming this cost base accounts for c7% of the group opex, this also translates into a reduction of -0.3% p.a. & +0.3% annual EBITDA boost, while increasing digital sales by >3pp each year.

In our model, we start our base case at 25% of sales being digital - largely in-line with telco incumbents at the moment. We assume that without the boost from AI, digitalisation should naturally go up to 40% over the next five years & stay at that level thereafter. However, we believe that in an AI-boosted scenario, telecom operators can shift sales into a digital environment faster & deeper. In that scenario, we assume an accelerated adoption and reaching 60% digital sales penetration by year 7: In total that drives an EBITDA margin expansion by +2.0%, assuming a similar proportionate impact as in the case studies above.

Improved customer retention

On top of cost saving opportunities from digital sales/service, we see an additional potential benefit from AI on the sales side: Smart pricing. Knowing exactly the right contract to offer to customers that are out of contract (or as an add-on) should allow for a better success rate in customer acquisitions. Freenet offered some insights into a potential upside from this, when it presented its smart pricing model: It expects an overall uplift of >20% for the full subscriber base, which is partly driven by +5% higher gross adds & +25% better retention.

In our model, we assume subscriber acquisition costs at 5% of revenues (e.g. for marketing or commissions). Smart pricing could help to generate more attractive, tailored offers and improve retention across the sector in our view. With a lower overall churn pool, we model that subscriber acquisition costs could fall by -20%, providing a +1% boost to EBITDA margins in an AI-boosted scenario.

Energy efficiency

Per commentary above, Nokia claims up to 30% energy savings through its AVA services and this could be broadly consistent with TEF's peak 24% energy savings in low traffic hours. On more conservative assumptions, TEF expects an average of 16% reduction in energy usage across 4/5G networks using Ericsson AI-powered services. On the other hand, Tele2's latest 5G case study appears more ambitious & points to mobile energy savings as high as 30-40% in the long term.

Whether the same savings can be applied across fixed networks is less obvious however data from BT has suggested that fixed traffic patterns are even 'peakier' than mobile, perhaps conferring similar or even better potential for power-down operations and reduced cooling etc.

Energy opex as a percentage of revenues across Euro telcos ranges from lows around 1.5% at DT (but this also includes lower priced US operations, while DT Germany is hedged at pre crisis levels) to Telenor's Q4 4.4%. We assume an average 3%.

In our model we thus apply a 25% reduction (average between TEF & Tele2 assumptions) x 95% network energy x 3% revenue exposure = 0.86bps EBITDA margin boost that we assume is managed into infrastructure over the long term (8-10 years).

Network deployment / fault management

Quantifying the benefits of network planning and testing is difficult and we exclude this here. To some extent this could be part captured in sales digitalisation opex savings that we defined above & that use AI network planning to close stores.

Quantifying fault management is also difficult, and once again there could be benefits already accrued in customer care savings above as incoming calls are reduced and/or handled by AI-bots. However:

• Data from Ericsson has indicated that its Energy Infrastructure Operations solution can achieve a15 percent reduction in site visits related to passive infrastructure, and an approximate 30 percent reduction in energy-related outages.
• Analysis from the US Fiber Broadband association indicated that 48% of direct copper network opex (under gross margin) is due to engineer truck rolls. We know fiber efficiency is around 50% improved, with 24% a potential proxy to reduce.

In our model we thus apply a 15% site visit reduction x 24% opex expenditure x 40% average telco COGS = 1,4% EBITDA margin benefit.

Capex efficiencies

Artificial intelligence should help on the capex side as well in optimising network efficiency, routing traffic and balancing workloads: Fault detection, prediction & resolution can all help in improving speed of network recovery, customer experience & an operator's cost profile. A precedent is Elisa, which for many years has not exceeded its own 12% capex/sales target - far below the sector average of >18%. A report by STL Partners highlights that Elisa, through rolling out a self-organising RAN, did not only reduce mobile network customer complaints by -20%, but also improved capex efficiency by -2%.

In our model, we assume a long-term benefit in capex intensity of c2% capex/sales, providing a +2% boost to OpFCF & lower capital employed over time.

Aggregated conclusions

Putting our assumptions together, we present a telecom sector model to highlight the potential benefits from AI integration over a ten-year period. We start at a 35% EBITDA margin, 15% capex/sales, 20% EBIT margin & a 6.0% ROCE.

Over a ten-year horizon, we see scope for a +5% expansion in the EBITDA margin from AI integration with digital sales adding +2pp, lower churn +1pp, lower energy usage +0.6pp & lower network opex (e.g. fewer faults) with another +1.4pp. To this we add potential efficiencies in capex deployment, adding another +2% to the OpFCF margin. In such a scenario, sector EBITDA margins could go up to 40% and OpFCF to 27% of sales.

We believe that the overall telco sector continues to earn ROCE below its costs of capital. However, better margins & lower capital intensity (hence, lower D&A and capital employed) could change this picture: On our estimates, we see the sector proxy reaching a ROCE of 8.0%, up by +2pp & more than covering current costs of capital.

We show below the contribution of the main factors, for which we see scope for a boost by AI. Combined, these could drive an improvement by up to +2pp.

 Wider operator summaries

• Deutsche Telekom: DT is one of the more advanced operators, with an all-IP network supporting its 'Telco as a platform' concept with software layers supporting the integration of AI into the digital journey from network planning through implementation, customer installation and management. Examples of efficiencies include the 10x acceleration of network planning from up to 4months to less than two weeks using AI to scan and plan potential rollout possibilities and further manage the engineering process. Also using AI within network maintenance has reduced downtime by 27% with 31% customers less impacted. Numerically Deutsche Telekom has probably been most open about its current state of digitalisation & its targets. For 2024, it announced a number of ambitions during its last DT capital markets day in 2021. Targets include not only a digitalisation of sales/customer care, but also improvements in network/IT performance & broadband experience. Not splitting up the savings between the different segments, the operator however expects an EBITDAaL impact of >E300m over 2021-24
• Swisscom: Swisscom is actively pursuing a higher online-only share of its B2C customer base. Since its tariff revamp in 2Q22, the Swiss operator is offering a CHF10/month discount to subscribers, that opt for online-only services. Actual online sales penetration varies still significantly between Swisscom's first brand (11% in 2022) and second brand (62%). For customer care, Swisscom already reports a majority of cases for its own brand being solved online (65% in 2022, growing steadily from 56% in 2020). A driver behind this is likely an increasing adoption of Swisscom's app, whose installations more than doubled from 0.9m in 2020 to 2.2m in 2022. Chatbots play a limited role within online penetration so far (<1m interactions), but the company indicates that usage is constantly increasing & the success rate is going up as well. In general, Swisscom flags that online & in-person interactions still go hand-in-hand with the former for information purposes and the latter (through 115 Swiss stores) for advice. In the B2B segment, the operator is also flagging some benefits from increased online usage with a focus particularly on mobile/billing/SMEs.
• Telenor: Telenor is already one of the most efficient European telecoms, when considering the labour intensity for the whole business: In 2022, it kept its wage costs at c10% of revenues - far below most other incumbents with ratios ranging from the teens to c30%. However, disclosure is limited on the extent to which digitalisation is helping to achieve that: Telenor does not report the extent of digitalisation in the sales/customer care channels. The main disclosure is on technology operations, for which Telenor has reported 58% automation coverage in 2021.
• Telefonica: TEF has established a global AI lab to support the integration of AI into day-to-day operations and is vocal on the use of technology in network planning, maintenance and deployment (with more disclosure given above). And at the group level TEF has >30% of sales now via digital channels with ranges of 30-50% across lines of business. Regarding the implantation fo various technologies, Telefónica Spain has been the first operator worldwide to test Ericsson's Radio Deep Sleep Mode energy saving functionality, tested in Madrid at a site with 5G deployment configuration. Supported by Artificial Intelligence and Machine Learning algorithms, the company achieved savings of up to 8%, considering the site's total 24-hour consumption, and up to 26% in low traffic hours. In addition, the Micro Sleep Tx functionality was tested in 4G and 5G technologies in the city of Talavera de la Reina. With continuous operation throughout the day, an energy reduction of up to 16% was achieved.
• Telia: Digitalisation & automation have been key pillars of Telia's transformation strategy, announced in early 2021 that targeted to reduce opex by SEK2bn (now reduced to exclude energy costs), including lower sales & marketing expenses. Digitalisation-related transformation targets include intelligent automation, -20%/-50% lower cycle time/error rate and up-to-date analytics tools made available to all staff. Overall, we find that disclosure on the current status of digitalisation of customer journeys (and network management) is limited, but note that Telia provides some data on results so far: (1) The number of automated campaigns nearly tripled by 2022 vs 2020 (+197%), to which management its -21% reduction in customer contacts over the same period. (2) Improvements and automation of processes decreased the number of incidents by -18% over 2020-22 & at the same time increased the hours saved.
• Elisa: Elisa provides more colour on the benefits from its customer digitalisation than its Nordic peers: As discussed above, around one third of consumer customer processes are already automated with a recent increase by +9pp over 2020-22 reducing costs by -8%. Moreover, during its 2023 capital markets day, management put numbers behind that trend: In 2022, Elisa resolved 130k cases with its chatbot & 70k with its voicebot, while the latter is allowing for 10k fewer call transfers per month. Overall, online contacts for Elisa have increased by +12.8% p.a. over 2020-22, while at the same time human interactions fell by -8.3%. In addition to the improvements in B2C/B2B customer journeys, Elisa is somewhat special within European telecoms through its International Digital Services segment: Accounting for <4% of group revenues, but growing by double-digits, the sub-segment provides a mix of software services to the telecom industry (e.g. network automation/monitoring or performance management) & manufacturing (e.g. analytics and data management). Both have the potential to benefit heavily from improvements in AI, but require openness from competitors to outsource such services (we have heard positive feedback) and require heavy investments into research & development: The segment still has not reached EBITDA break-even despite the recent acquisitions of mostly profitable businesses, but management now guides for improvement in profitability as the focus of investments now shifts towards sales for a highly scalable business (see 'CMD feedback - Outlook in-line, focus on digital services', 10 March 2023).
• Tele2: Simplification of IT, particularly with the integration of ComHem & TDC, has been a main element of Tele2's cost saving strategy over the last years. While the operator provides only relatively limited disclosure on KPIs around automation of its customer journey/network management, its latest case study on AI for energy usage highlights Tele2's ambition and could work as a good proxy for the overall sector - pointing to an energy reduction by c30-40% in the long term.
• Freenet: With no own infrastructure (aside from Media Broadcast), Freenet's digitalisation targets mainly focus on a more efficient customer journey. And indeed, the operator shows the extent to which telcos can leverage its online presence: On top of brick-and-mortar through 500 own shops & c400 Media-Saturn shops (plus independent retailers), the operator has shifted a significant portion of sales online. In 2022, around half of gross adds were online - coming down from an even higher share during the Covid-19 pandemic. A similar share of customer requests was also resolved automatically, e.g. with the use of Chatbots. Additionally, Freenet uses WhatsApp as an additional communication channel, in which 30% of interactions have already been automised. Moreover, Freenet uses automation in the management of its customer base through a new focus on customer lifetime value (per its 2021 capital markets day): A data-driven smart pricing model is targeted to optimise compound margins & allow for one-to-one offers. For the full base, management expects the new approach to uplift by >20% through smart pricing (e.g. with higher gross adds/better retention). Overall, we expect Freenet's business model to remain hybrid (with particularly SIM-only online), but aided by big data. Importantly, already achieved levels of digital sales should underscore Freenet's business proposition even in an online environment.
• Orange: Digitalisation and AI continue to be a central part of the incumbent ambitious of network and customer excellence. More specifically at its latest CMD, Orange targeted digital sales in France >40% by 2025 (they were at 23% in 2021) and digital aftersales >65%. For Orange, AI and digital efforts "reinvent phygital customer experience through excellence of digital channels and expertise of customer advisors empowered by AI to (a) ensure 360° customer view (b) augment customer autonomy and (c) transform Shops". The efforts is equally strong for networks, with a new Group industrial model that, thanks to network softwarisation, automation & AI is translating into a 30x to 300x faster "model". Lastly, Orange is also able to leverage the digital age and AI tools to boost ambitious in its Enterprise unit and more specifically, through Cybersecurity.