Introduction: The Day a Mobile Network Began Becoming an Intelligence Platform
On the morning of October 21, 2026, in a conference hall in Berlin, Jensen Huang stood before an audience of telecommunications executives, engineers, and investors and described a call that had taken place thousands of miles away in Bellevue, Washington. The call itself was, by every outward measure, unremarkable. A commercial smartphone connected to a commercial cell site operating in licensed spectrum, the voice was carried over T-Mobile’s existing 5G Standalone core, and the conversation proceeded as billions of conversations proceed every day. What made the call worth announcing at one of the most closely watched technology conferences of the year was not what the user heard but what the infrastructure underneath the call had become. The radio software that processed the signal, Nokia’s CUDA-accelerated anyRAN software, was running not on purpose-built baseband silicon but on NVIDIA’s Aerial RAN Computer, a general-purpose accelerated computing platform designed from the beginning to support artificial intelligence workloads alongside radio functions.[1] NVIDIA, Nokia, and T-Mobile described it as the world’s first AI-RAN call on a commercial network, and Nokia’s own engineers were careful to add that it was also, more precisely, the world’s first software-defined radio call completed using accelerated computing in a live operator environment.[2]
The significance of a single call should not be exaggerated, and Nokia’s own commentary was admirably restrained on this point. The companies did not disclose throughput, latency, or spectral-efficiency improvements attributable to the AI-RAN configuration, and independent observers were quick to note that the demonstration proved coexistence rather than superiority: it showed that a commercial radio stack could operate on GPU-based infrastructure in a subscriber-facing network without dropping the call, not that artificial intelligence had yet improved the quality of that call in any measurable way.[3] Nokia’s blog put it plainly:
“One successful live-network call is not the destination.” — Nokia Corporation, Nokia Blog [2]
Yet the restraint of the engineers should not obscure the magnitude of the shift that the call represents. For more than a century, the telecommunications network has been designed, financed, regulated, and valued as a system that moves information from one place to another. The Bellevue call demonstrated that the same physical infrastructure, the same cell sites and switching facilities and fiber-connected aggregation points, could host the kind of computing platform on which modern artificial intelligence runs. Jensen Huang framed the moment in characteristically expansive terms:
“Now, they will also carry the world’s intelligence.” — Jensen Huang, Founder and CEO, NVIDIA, on telecom networks built to carry conversations [4]
Across the Atlantic and the Pacific, the Bellevue call was surrounded by a constellation of related developments that, taken together, suggest an industry in transition rather than a vendor in search of a headline. Deutsche Telekom reported that laboratory measurements conducted in Bonn and in Nokia’s facility in Oulu had shown typical uplink throughput improvements of 10–15 percent from AI-based channel estimation running in real time alongside the radio workload, with larger gains under more complex radio conditions and field validation planned for 2027.[5] Elisa reported that it had run voice, mobile data, and video analytics simultaneously on the same accelerated platform over a live 5G Advanced network in Finland, which it described as Europe’s first field trial of its kind, and announced that it was deploying GPU-equipped servers at multiple network locations to create distributed AI computing capacity.[6] BT disclosed plans for live trials in Birmingham in 2027. Orange Business and NVIDIA described a pilot of distributed voice-AI inference spanning central cloud and regional points of presence, including an emergency-call scenario in which AI translates between a foreign-language caller and an operator in real time.[7] NVIDIA reported that nearly twenty operators worldwide had publicly announced AI-RAN collaborations using Nokia’s software on its accelerated platform.[1] On the previous day, SoftBank and Nokia had agreed to conduct a commercial AI-RAN field trial in Japan in 2027 under live traffic conditions.[8] And six days later, on October 27, T-Mobile’s chief technology officer John Saw published an essay describing the company’s work with Figure AI, Serve Robotics, the City of San Jose, and Amazon Web Services, in which he disclosed that T-Mobile and Figure were assessing whether portions of Figure’s Helix S1 and S2 robot models could run on compute closer to the network edge while preserving the responsiveness that physical machines require.[9]
The reason these developments matter extends well beyond the telecommunications industry’s perennial search for a new revenue story. Consider the industrial and urban environment that is now taking shape: humanoid robots being trained for warehouse and factory work, sidewalk delivery robots navigating pedestrian traffic, inspection drones patrolling utility corridors, connected vehicles exchanging information with roadside infrastructure, and AI assistants embedded in glasses and earbuds that continuously upload images and audio to obtain contextual help. Each of these machines requires intelligence, and until recently the architecture for supplying that intelligence has been binary: either the model runs on the device, within the strict limits of battery, heat, weight, and cost, or the device transmits its data to a distant hyperscale datacenter, waits for processing, and receives a response through the network. That arrangement remains perfectly adequate for a great many applications, but physical systems increasingly demand predictable responsiveness, continuous coordination with other machines, and dependable access to intelligence while they are moving through environments that change from one moment to the next. A robot navigating a crowded sidewalk cannot treat every unexpected obstacle as an ordinary cloud request. An industrial machine coordinating with a dozen others on a factory floor cannot assume that every decision should make a round trip to a hyperscale facility in another state.
The emerging alternative is a layered, distributed computing architecture in which certain AI workloads execute on the device itself, others execute at telecommunications edge facilities located close to where machines operate, and the most computationally demanding activities, including the training of frontier models, remain in regional and hyperscale datacenters. Stanford’s Fei-Fei Li has argued with growing insistence that the frontier of artificial intelligence is shifting from language toward the physical world, toward what she calls spatial intelligence, and that this shift changes the kind of infrastructure the field will require:
“AGI will not be complete without spatial intelligence.” — Fei-Fei Li, Stanford University and World Labs [10]
If she is right, and if the machines that embody spatial intelligence are mobile, battery-constrained, and numerous, then the question of where their intelligence executes becomes an infrastructure question of the first order. It is that question, and the possibility that telecommunications networks could supply part of the answer, that this paper calls Tower Inference.
Why I Chose the Title “Tower Inference”
I chose the title Tower Inference because the combination of these two words captures an economic and technological transition that the conventional vocabulary of edge computing, 5G Advanced, and AI-native networking does not fully express. Each of those existing terms describes a piece of the picture, but none of them conveys the structural idea that the physical apparatus of telecommunications could itself become part of the computational system that produces intelligence rather than merely the conduit through which intelligence is delivered.
The word Tower stands for the geographically distributed telecommunications system in its physical entirety: the radio sites with their antennas and baseband cabinets, the central offices and switching facilities that have anchored the network for a century, the fiber-connected aggregation points that gather traffic from dozens of cells, the regional computing locations that operators have built for their own network functions, and the operational apparatus of power, cooling, security, and maintenance that keeps all of it running around the clock. The word is intentionally concrete. The telecommunications network is not an abstraction; it is a vast inventory of real estate, electrical service, and connectivity that exists today in every metropolitan area, suburb, industrial corridor, transportation route, and rural community that a national operator serves.
The word Inference stands for the practical execution of trained artificial intelligence models: the moment at which a model interprets a camera frame, transcribes a voice, predicts a trajectory, recommends an action, or generates a response. Inference is distinct from training, which is the computationally gargantuan process of producing a model in the first place and which will remain concentrated in hyperscale facilities for the foreseeable future. Inference, by contrast, is the part of artificial intelligence that must happen wherever and whenever a user or a machine needs an answer, and it is therefore the part of artificial intelligence whose geography is genuinely open to redesign. Stanford’s AI Index has documented the steep and continuing decline in the cost of inference, which is precisely what makes its distribution across many smaller locations economically conceivable.[11]
Bringing the two words together identifies a developing market in which telecommunications infrastructure could become a geographically distributed resource for computing intelligence. The title does not imply, and this paper will repeatedly caution against assuming, that every radio tower will contain powerful GPUs. Much of the computing described here will reside in nearby central offices, metropolitan aggregation facilities, and regional edge locations that serve many radio sites at once, because those facilities already possess the electrical capacity, cooling, physical security, and fiber connectivity that a rooftop installation lacks. Tower is a synecdoche for the network as a whole, not a literal prescription for where the hardware will go.
The term is particularly appropriate within the Five-Layer AI Economy framework that organizes this research series, because it describes a new interaction among energy, semiconductors, datacenters, models, and applications. In the conventional hyperscale model, electricity and advanced chips are concentrated in enormous computing campuses where frontier models are trained and deployed, and the network serves those campuses as a logistics system serves a factory. Tower Inference introduces a complementary possibility: that existing telecommunications networks could distribute selected computing workloads across locations that already possess connectivity, spectrum access, facilities, and operational support, thereby turning the third layer of the AI economy, the datacenter layer, from a small number of very large nodes into a wide and heterogeneous mesh. NVIDIA, Nokia, T-Mobile, Ericsson, Cisco, AWS, Google Cloud, and a growing list of other participants would therefore find themselves competing and collaborating over an increasingly valuable set of questions: where should intelligence execute, who should operate the infrastructure on which it executes, and who should capture the economic value generated by intelligent machines? The title is not merely a description of faster wireless networks. It defines a research framework for examining the possible transformation of telecommunications into a commercially significant component of artificial intelligence infrastructure through 2030 and beyond.
Section 1: The Transformation of Telecommunications Networks — From Carrying Information to Computing Intelligence
The first task of any serious examination of Tower Inference is to understand what, precisely, is being transformed. Telecommunications networks are among the oldest and most capital-intensive infrastructures of the modern economy, and their economic identity has been remarkably stable across more than a century of technological change. This section begins with that historical identity, because the magnitude of the proposed shift can only be appreciated against the persistence of what came before. It then introduces the three distinct forms in which artificial intelligence and the radio access network are converging, examines the geography of distributed intelligence and why that geography matters, analyzes the way artificial intelligence is changing the physical character of mobile traffic, and finally situates the telecommunications network within the Five-Layer AI Economy as a potential operator of distributed computing assets rather than merely a transport provider.
1.1. The Historical Function of Telecommunications Infrastructure
For more than a century, telecommunications networks have been evaluated almost entirely according to their ability to connect people, businesses, and devices to one another. The industry’s technological evolution, from analog telephony and electromechanical switching to digital switching, from circuit-switched voice to packet-switched data, from dial-up to broadband, and from the first generation of cellular networks through 4G and 5G, has repeatedly expanded the capacity of the network and the range of services that connectivity could support. Each generation brought dramatic improvements in throughput, latency, reliability, and coverage, and each generation required enormous new investment in spectrum, equipment, and construction.
Nevertheless, beneath all of this technological change, the underlying commercial relationship remained remarkably consistent. Telecommunications companies operated networks that transported information, and other industries developed most of the applications that generated economic value from that information. The operator sold minutes, then messages, then megabytes; the application developer, the content producer, and the software company sold what people actually wanted to do with those minutes, messages, and megabytes. The network was indispensable, but it was indispensable in the way that a road is indispensable to a trucking company: essential to the business, but not the business itself.
The internet economy intensified this distinction to a degree that has shaped the strategic anxieties of the telecommunications industry for two decades. Technology companies including Google, Amazon, Microsoft, Meta, and Apple built some of the most profitable enterprises in history on top of telecommunications infrastructure that they did not own and, for the most part, did not pay for in proportion to the value they extracted. Mobile network operators supplied connectivity, but application providers, cloud platforms, digital advertisers, and software developers controlled the customer experiences that commanded higher margins and compounding growth. The operator’s investment produced the substrate; the platform’s investment produced the profit. This asymmetry became so familiar that it acquired a name in industry discourse, the “dumb pipe” problem, and successive generations of operator strategy have been organized, often unsuccessfully, around escaping it.
The arrival of artificial intelligence as the dominant workload of the next decade could alter this relationship, and it is important to be precise about why. As applications become increasingly dependent on model execution rather than on the retrieval and display of stored content, the network itself may acquire a second economically meaningful role: hosting the computational functions required by the services it connects. A telecommunications operator could eventually supply connectivity, inference capacity, network coordination, location-sensitive services, security controls, and operational intelligence as elements of an integrated platform rather than as a single commoditized input to someone else’s platform. Cisco’s Gurudatt Shenoy framed the strategic question facing service providers in exactly these terms at GTC Berlin, asking whether the network would carry someone else’s value or become the platform on which that value is created, and arguing that the answer depends on whether operators can
“solve utilization through placement and routing rather than more infrastructure.” — Gurudatt Shenoy, Cisco [12]
This would represent an important change in the industry’s economic identity. Instead of participating principally as a transportation provider within the digital economy, the telecommunications industry could become a direct infrastructure participant in the production and delivery of artificial intelligence. Michael Dell, whose company supplies the servers for Nokia’s AI-RAN systems, made the same point from the vendor’s side of the table:
“The telecommunications industry owns the most valuable real estate for AI — the edge.” — Michael Dell, Chairman and CEO, Dell Technologies [13]
The emerging opportunity is not guaranteed, and the history of the industry counsels humility. Telecommunications operators have previously struggled to capture the value of digital services despite investing hundreds of billions of dollars in mobile broadband infrastructure that made those services possible. Operator-branded app stores, operator-branded cloud services, operator-branded content platforms, and operator-branded payment systems have, with a handful of exceptions, failed to achieve the scale and margins of their platform competitors. Tower Inference therefore raises a consequential question that this paper will keep returning to: will artificial intelligence create a genuinely new revenue model for the telecommunications industry, or will telecommunications companies once again provide essential infrastructure while other companies capture most of the profits? The answer will depend less on the technology than on the structure of the commercial relationships, the control of orchestration software, the ownership of the enterprise customer, and the discipline with which operators evaluate where distributed computing creates value and where it does not.
1.2. Understanding AI-RAN: Three Different Forms of Telecommunications Intelligence
The AI-RAN Alliance, established in February 2024 at Mobile World Congress with founding members including NVIDIA, Nokia, Ericsson, Samsung, SoftBank, T-Mobile, Arm, Microsoft, DeepSig, and Northeastern University, organizes the convergence of artificial intelligence and radio networks into three related but distinct technical concepts, and the distinction among them is the single most important conceptual tool for evaluating any AI-RAN claim.[14] The three concepts were formalized in the academic literature by Kundu, Lin, Gadiyar, Lacasse, and Chowdhury of NVIDIA in a January 2025 paper published in IEEE Network, which described the radio access network as undergoing a shift away from
“communication-centric infrastructures towards compute-communication platforms.” — Lopamudra Kundu et al., NVIDIA Research, IEEE Network [15]
The first form is AI-for-RAN: using artificial intelligence to improve the network itself. In this form, AI algorithms assist with radio resource management, interference mitigation, traffic prediction, beamforming, channel estimation, scheduling, energy savings, and operational automation. The value created is internal to the operator: better network performance, higher spectral efficiency, and lower operating costs. Deutsche Telekom’s 10–15 percent uplink gain from AI-based channel estimation is a pure AI-for-RAN result, as is Ericsson’s AI-native scheduler, which the company reports delivered roughly 10 percent spectral-efficiency gains and up to 20 percent downlink throughput gains in more than fifteen deployments and trials.[5][16] AI-for-RAN is the least controversial and most commercially mature of the three forms, and it has been pursued vigorously by vendors using both general-purpose accelerators and purpose-built silicon.
The second form is AI-and-RAN: sharing computing infrastructure between network functions and external AI applications. In this form, an accelerated computing platform installed at a network location supports both the telecommunications processing that the operator requires and additional AI workloads that have nothing to do with the radio network. The economic logic is utilization: radio workloads vary with time of day, location, and events, and a platform purchased to handle peak radio demand will have idle capacity during off-peak periods that could, in principle, be allocated to other work. SoftBank’s AITRAS orchestrator, which in February 2026 was extended in collaboration with Nokia Bell Labs to allocate computing resources to external customers’ AI workloads alongside internal RAN control, is the most explicit commercial expression of AI-and-RAN to date.[17]
The third form is AI-on-RAN: running AI applications using computing resources associated with the radio network, in order to deliver services to end users and enterprises. In this form, the distributed infrastructure supports video analytics, robotic coordination, industrial monitoring, conversational voice assistants, and other applications that benefit from executing close to the user. This is the form that most directly expresses the commercial opportunity represented by Tower Inference, because it is the form in which the operator sells intelligence rather than merely consuming it or renting out spare cycles. Elisa’s demonstration of video analytics running alongside voice and data on a shared platform is an AI-on-RAN result; so is Orange’s distributed voice-AI pilot; so is T-Mobile’s exploration of running portions of Figure’s Helix models at the network edge.[6][7][9]
Northeastern University’s Tommaso Melodia, whose Institute for the Wireless Internet of Things is a founding academic member of the AI-RAN Alliance and whose group built the Colosseum emulator and the X5G testbed on NVIDIA ARC hardware, has emphasized that these three forms should not be seen as a hierarchy of ambition but as a set of architectural problems that must be solved together. In a 2025 paper proposing a converged O-RAN and AI-RAN architecture, Polese, Mohamadi, D’Oro, Bonati, and Melodia argued that data-intensive AI applications at the network edge demand a shift in RAN design
“from merely consuming AI for network optimization, to actively enabling distributed AI workloads.” — Michele Polese, Tommaso Melodia et al., Northeastern University [18]
Melodia has elsewhere described artificial intelligence as more than a performance-enhancement tool for the radio, characterizing it as “a driving force for edge computing” in its own right.[19]
The distinctions matter because all three approaches can exist independently, and because conflating them is the most common error in both promotional and critical commentary. A telecommunications company might successfully improve radio performance with AI-for-RAN without generating any revenue from external inference, and Ericsson’s decision to deliver its AI in RAN capabilities as a software subscription running on its own purpose-built silicon, explicitly without GPUs, is a strategy built on exactly that premise.[16] Another operator might deploy shared accelerated computing in the AI-and-RAN model and discover that commercial demand for its spare capacity is thin or intermittent. A third could profitably offer AI-on-RAN edge inference services from regional facilities without placing any computing hardware at individual cell sites at all. Consequently, rigorous AI-RAN research must distinguish among technological feasibility, network efficiency, commercial deployment, and actual profitability. These are separate milestones, achieved by different evidence, and the Bellevue call was a milestone of the first kind.
1.3. The Emerging Geography of Distributed Intelligence
The geographical distribution of computing infrastructure is fundamental to understanding Tower Inference, and it is here that the loose language of “putting AI at the tower” does the most damage to clear thinking. Traditional hyperscale AI infrastructure concentrates enormous computing resources in facilities designed around high-capacity electrical interconnection, specialized liquid cooling, sophisticated internal networking, and large accelerator clusters that can be scheduled as a single machine. These facilities remain indispensable for large-scale model training and for many forms of inference, and nothing in the Tower Inference thesis suggests otherwise. NVIDIA’s own financial results make the point with unusual force: in the quarter ended July 26, 2026, the company reported Data Center revenue of $89.0 billion, up 117 percent from a year earlier, against Edge Computing revenue of $7.2 billion, a category that grew a healthy 27 percent but remains an order of magnitude smaller.[20] The center of gravity of AI computing is the hyperscale campus, and it will remain so.
Telecommunications networks possess a different physical advantage, which is geographic reach. A national mobile operator already has facilities distributed across metropolitan areas, suburbs, industrial corridors, transportation routes, and rural communities, and it already has the fiber, the power, the physical security, and the operational staff to run those facilities. However, the practical suitability of these locations for AI workloads varies enormously. A rooftop radio installation may have a few square feet of cabinet space, a modest electrical feed, passive cooling, and no on-site staff; it is a poor candidate for a multi-kilowatt GPU server. A metropolitan switching center, by contrast, may have thousands of square feet of conditioned floor space, redundant power, and existing fiber interconnection to dozens of cell sites; it is a far better candidate. A regional telecommunications datacenter could potentially host substantial inference capacity while maintaining direct network interconnection to an entire market.
This variation suggests a differentiated computing hierarchy rather than a uniform deployment, and it is this hierarchy that the term Tower Inference is meant to encompass. Cisco’s 2026 inference-architecture white paper formalizes a four-tier placement model of far edge, metro edge, central edge, and external cloud or peering, and treats every inference location, from a single edge cabinet to a regional facility, as a datacenter in its own right with its own routing, security, and assurance requirements.[21] The framework proposed in this paper extends that model to include the device itself and the hyperscale campus at the two extremes, because the economics of Tower Inference depend on understanding the whole continuum rather than only the middle of it.
| Location | Most suitable computing role | Typical constraints |
| On-device | Immediate sensing, safety-critical control, lightweight models, offline operation | Battery, heat, weight, cost, model size |
| Radio site or nearby far edge | Selected low-latency inference, radio-signal intelligence (channel estimation, beamforming), event detection, local coordination | Space, power, cooling, no on-site staff, uplink congestion |
| Metro aggregation facility or central office | Shared AI services, industrial applications, video analytics, voice AI, regional inference capacity serving many sites | Electrical upgrades, thermal retrofits, orchestration complexity |
| Regional telecommunications datacenter | Larger models, fleet coordination, model serving, elastic computing, sovereign or jurisdiction-bound workloads | Competition with other datacenter demand for power and labor |
| Hyperscale AI campus | Frontier-model training, large-scale inference, global applications, extensive model development | Distance from users, transport latency, concentration risk |
Table 1. The Tower Inference computing hierarchy. Analytical framework proposed for this paper, extending the four-tier placement model in Cisco’s AI Inference Architecture for Wide Area Networks. Placement depends on workload, power, connectivity, resilience, data sovereignty, and economics.
This hierarchy challenges the assumption that all meaningful artificial intelligence must originate from a small number of exceptionally large computing campuses, but it does not imply their replacement. Instead, the AI economy may evolve toward a coordinated system in which large facilities produce and serve substantial quantities of intelligence while smaller distributed locations execute selected applications with demanding geographic or operational requirements. NVIDIA’s October 21 announcement explicitly identifies centralized datacenters, metropolitan sites, and cell sites as potential locations within this distributed architecture, and the new Aerial RAN Computer configurations it introduced at GTC Berlin, the MGX 2U and MGX 4U powered by RTX PRO 6000 Blackwell Server Edition GPUs with x86 or Vera CPU options and server offerings from Dell, Quanta Cloud Technology, and Supermicro, are designed for high-capacity sites rather than for individual rooftops.[1] Elisa’s decision to deploy such servers “at multiple network locations” rather than at every cell is consistent with the same logic.[6]
The academic literature anticipated this geography well before the vendors commercialized it. Letaief, Shi, Lu, and Lu, writing in the IEEE Journal on Selected Areas in Communications in 2022, described the evolution of wireless networks “from connected things to connected intelligence” and argued that embedding training and inference capabilities into the network edge would allow 6G to integrate sensing, communication, computation, and intelligence into a single system.[22] Xiao, Shi, Li, Saad, and Princeton’s H. Vincent Poor, in a 2020 paper on self-learning edge intelligence, identified the absence of edge-native AI as
“one of the key missing components in 5G.” — Yong Xiao, Walid Saad, H. Vincent Poor et al., Princeton University and Virginia Tech [23]
And Wen Tong and Imperial College’s Geoffrey Ye Li, in their widely cited 2021 paper on the nine challenges of AI and wireless for 6G, drew a distinction between AI-for-Network, which they associated with 5G, and Network-for-AI, which they argued would be the mainstay of 6G, going so far as to predict that
“AI will revolutionize 6G and will also be the fundamental business for 6G.” — Wen Tong and Geoffrey Ye Li, Imperial College London [24]
Tower Inference is, in a sense, the commercial name for the Network-for-AI proposition that these researchers advanced half a decade ago.
1.4. Why Artificial Intelligence Changes Mobile Network Traffic
The traditional mobile internet was heavily oriented toward downloading information. Users streamed videos, browsed websites, downloaded applications, and consumed content that had been generated elsewhere and stored in distant servers, and the networks were engineered accordingly, with downlink capacity dominant and uplink capacity treated as an afterthought. Ericsson’s measurements across a sample of operator networks found that uplink traffic averaged only about 8 percent of total traffic volume, and in video-dominated traffic the uplink share fell to roughly 3 percent.[25][26] A network built on these proportions is a network built to deliver, not to listen.
Physical artificial intelligence changes this relationship, and the change is structural rather than incremental. An autonomous robot may continuously upload visual information, sensor readings, location changes, and operational context in order to obtain guidance or to contribute to a shared model of its environment. An industrial monitoring system may transmit inspection images and equipment telemetry at high resolution and high frequency. An AI-enabled wearable may periodically upload images or audio to obtain contextual assistance from a model that is too large to run in a pair of glasses. In every one of these cases, the device is a source of data rather than a sink for content, and the network’s uplink becomes the binding constraint. Ericsson’s September 2026 analysis opened with a sentence that could serve as the epigraph for this entire section:
“Most mobile networks were built for a downlink-heavy world.” — Ericsson, Ericsson Blog [27]
The same analysis argued that AI-enabled applications could create substantially greater uplink demand, particularly when devices depend on remote inference because of battery, thermal, or processing limitations, and Ericsson’s June 2026 Mobility Report projected, on the basis of scenario modeling, that additional AI traffic alone could make uplink volumes three times higher in 2031 than in 2025.[28] More than three quarters of operators already experienced faster uplink growth than downlink growth in 2025.[25] T-Mobile’s John Saw, speaking on the company’s third-quarter 2026 earnings call, offered a characteristically concrete description of what he expects the traffic mix to look like by the end of the decade, driven by AI wearables, AI agents, and what he called the explosive growth of physical-AI devices such as humanoids, robots, delivery drones, vehicles, and factory machines:
“still downlink heavy but increasingly uplink rich” — John Saw, Chief Technology Officer, T-Mobile US [29]
It should be said that not every observer accepts the premise that an uplink surge is imminent or that it would require major new investment. William Webb, the independent consultant and former IEEE president, has argued that a doubling or even tripling of uplink traffic could likely be accommodated within existing networks because
“it would just be filling spare capacity.” — William Webb, independent consultant and former IEEE President [26]
The disagreement is not academic, because Ericsson has explicitly tied its 6G investment thesis to the uplink story, and Light Reading has observed with some skepticism that there has famously been no correlation between downlink traffic growth and the connectivity revenues earned by operators, which raises the uncomfortable possibility that uplink growth will follow the same pattern.[30] The resolution of this debate will depend on data that does not yet exist at scale, and this paper treats the uplink question as an open empirical matter rather than a settled fact.
What is not open to debate is that the location of computing resources influences the communications path required to serve an intelligent device, and that this relationship cuts in both directions. Moving an inference workload closer to the device may reduce transport distance and avoid certain network segments, but it does not automatically eliminate the need for substantial radio capacity. A system transmitting continuous high-resolution video may still place significant demands on wireless uplink resources even when its inference model runs in a facility a few kilometers away. Likewise, placing computing hardware at a cell site does not guarantee acceptable application latency if the wireless connection is congested or if the accelerator is overloaded by competing requests. The relevant engineering objective therefore involves the entire processing and communications path, which for a remotely executed inference task can be expressed as:
Ltotal = Luplink + Ltransport + Lqueue + Linference + Ldownlink
In this expression, Luplink represents transmission from the device into the radio network, including any delay from contention for uplink resources in a busy sector; Ltransport represents movement to the processing location and the associated traversal of fronthaul, midhaul, backhaul, and transport segments; Lqueue represents waiting for available computing capacity at the chosen inference location; Linference represents the actual execution of the model; and Ldownlink represents the return path to the device. An economically and technically successful Tower Inference system must optimize these combined elements rather than any one of them in isolation, and the uncomfortable truth embedded in the equation is that the term most under the operator’s control, Ltransport, is often not the term that dominates. Saw himself identified the uplink as the component most in need of attention, telling analysts that if there is one thing to keep an eye on it is the radio uplink in busy sectors, particularly with physical AI and AI agents, while noting that carrier aggregation, transmit switching, higher-power transmission, and new spectrum are being brought to bear on the problem.[31]
Equally important, operators must measure consistency and tail latency, not simply average latency. A system that usually responds in twenty milliseconds but occasionally takes two seconds is, for a robot deciding whether to stop, a system that responds in two seconds. The distribution of Ltotal across percentiles, under congestion, during handovers, and in the presence of failures, is the figure of merit that will determine which physical-AI applications can safely depend on network inference and which cannot. Jeffrey Andrews of the University of Texas at Austin, director of 6G@UT and one of the most cited researchers in wireless systems, made a closely related observation in his keynote on deep learning in the 6G air interface, noting that autonomous vehicles and robots need situational awareness that extends beyond what they can sense alone, because otherwise
“they’ll have to drive too slowly, too conservatively.” — Jeffrey G. Andrews, Truchard Family Endowed Chair in Engineering, The University of Texas at Austin [32]
The implication for Tower Inference is that network-supplied intelligence is valuable precisely when it is reliable enough to be trusted, and that reliability is a property of the whole path, not of the accelerator alone.
1.5. Telecommunications as a New Participant in the Five-Layer AI Economy
The Five-Layer AI Economy provides a useful structure for understanding why distributed telecommunications computing could become economically important, and it is worth walking through each layer deliberately, because the significance of Tower Inference lies in how it changes the relationships among the layers rather than in any single layer taken alone.
Layer 1 is Energy. Distributed AI facilities require dependable electricity, electrical switchgear, cooling, backup power, and increasingly sophisticated energy management, and they require these things at many locations rather than at a few. The International Energy Agency has estimated that global data-center electricity consumption grew 17 percent in 2025, more than five times the global average rate of electricity demand growth, and projects that consumption will roughly double by 2030 with AI-focused facilities tripling their draw.[33] Tower Inference adds to this picture not a new quantity of demand so much as a new geography of demand, scattered across central offices and aggregation sites whose electrical service was designed for telecommunications equipment rather than for GPUs.
Layer 2 is Chips. NVIDIA accelerators, Ericsson’s purpose-built RAN silicon with neural-network accelerators, Marvell and other merchant silicon, Arm-based CPUs, and embedded processors on devices together determine which models and which network functions can operate efficiently at each location in the hierarchy. The competition between general-purpose GPUs and purpose-built silicon, which this paper examines in Section 4, is at bottom a Layer 2 competition over which architecture will win the next RAN cycle.
Layer 3 is Datacenters. Telecommunications central offices, metropolitan computing facilities, and AI-ready radio sites could become extensions of the broader datacenter layer, and this is the layer that Tower Inference most directly transforms. The structural significance is that Layer 3 would no longer be identified primarily with enormous hyperscale campuses. Instead, it could include a wider network of computing facilities with different capacities, ownership arrangements, energy requirements, and geographic advantages, operated by companies that have never previously thought of themselves as datacenter businesses.
Layer 4 is Models. Open-weight models such as NVIDIA’s Nemotron family, proprietary AI services such as Google’s Gemini, specialized speech and vision systems such as NVIDIA Riva, vision-language models, robot-control models such as Figure’s Helix, and compact domain-specific models provide the computational intelligence executed at appropriate locations. The trend toward smaller, more efficient models that can run on a single server rather than a cluster is what makes distributed execution feasible; the Stanford AI Index 2026 reports that model quality at fixed compute has improved roughly fivefold since the release window of GPT-4, which means that a given edge server can deliver far more useful intelligence today than it could two years ago.[11]
Layer 5 is Applications and Agents. Autonomous robots, industrial inspection systems, emergency communications, city-monitoring applications, voice assistants, and enterprise agents generate the demand for distributed processing, and it is the characteristics of these applications, their latency tolerance, their data volume, their safety requirements, their geographic footprint, that determine whether Tower Inference makes economic sense for any given use.
The structural significance of Tower Inference, viewed through this framework, is that telecommunications companies would become potential operators of distributed Layer 3 assets while participating more directly in the commercial activities of Layers 4 and 5 than they ever have before. Nokia’s chief executive Justin Hotard captured the ambition in a single phrase when announcing NVIDIA’s $1 billion investment in his company in October 2025:
“put an AI data center into everyone’s pocket.” — Justin Hotard, President and CEO, Nokia [34]
This produces the paper’s first major proposition: the future AI infrastructure economy may be organized not only by the quantity of computing capacity available, but also by the geographical location, accessibility, reliability, and operational suitability of that capacity. Quantity has dominated the discourse of the past three years, measured in gigawatts and in billions of dollars of accelerator purchases, and quantity will continue to matter. But as the AI economy shifts from training toward inference, and as inference shifts from chatbots toward machines that move through the physical world, the question of where capacity sits, and how reliably it can be reached, becomes a dimension of competition in its own right. The telecommunications network is the only infrastructure in the economy whose entire design philosophy has been organized, for a century, around that very question.
Section 2: AI-RAN Economics — Shared Computing, Network Productivity, Operator Balance Sheets, and the Search for New Telecommunications Revenue
If Section 1 established what Tower Inference is, this section asks the harder question of whether it can pay. The telecommunications industry has never lacked for technological visions; what it has lacked, with painful regularity, is a mechanism for converting those visions into returns that justify the capital they consume. The economic analysis of AI-RAN must therefore begin with the structural investment problem that operators face, proceed through the specific mechanisms by which shared computing, spectral efficiency, and new service models might generate value, and then test those mechanisms against the hardest evidence available: the earnings reports that the largest operators and vendors filed in October 2026. This section also proposes an explicit investment framework for evaluating a distributed inference deployment, and it closes with the risk that the telecommunications industry’s long asset lives are fundamentally mismatched to the short economic lives of AI accelerators.
2.1. The Economic Problem Facing Telecommunications Operators
Telecommunications operators face a longstanding investment challenge that has become, if anything, more acute in the 5G era. Mobile networks require continuous spending on spectrum, radio equipment, fiber, network software, maintenance, energy, and technological upgrades, and the scale of that spending is enormous. T-Mobile maintained a capital-expenditure plan of approximately $10 billion for 2026; Verizon held its guidance at $16.0 to $16.5 billion; AT&T guided to $23–24 billion of capital investment in 2026 and $22–23 billion in each of 2027 and 2028, on top of $22.8 billion of spectrum acquisition expenditure recorded in the third quarter alone, principally reflecting its EchoStar transaction.[35][36][37] Yet improvements in connectivity do not necessarily produce proportional increases in revenue. Consumers have become accustomed to effectively unlimited mobile data, competition constrains pricing, and the industry’s experience with 5G, which delivered dramatic capacity improvements without a correspondingly dramatic new consumer revenue stream, has made investors wary of any thesis that assumes traffic growth will translate into monetization.
Enterprise services offer opportunities for differentiation, but developing them requires commercial relationships, sales capabilities, and technical competencies beyond ordinary consumer connectivity, and operators have historically been better at building networks than at building enterprise software businesses. The strongest recent operator results have come not from new services but from pricing discipline and operational efficiency: Verizon’s record third-quarter 2026 adjusted EBITDA of $13.8 billion, up 8 percent, was attributed by its chief executive Dan Schulman to a disciplined approach to device subsidies and a structural evolution of the business model, and T-Mobile’s 11 percent Core Adjusted EBITDA growth to $9.6 billion was driven substantially by a 4 percent increase in postpaid average revenue per account to $155.73.[36][35] These are impressive results, but they are results of running the existing business better, not of inventing a new one.
Artificial intelligence introduces a possible alternative. Rather than using computing resources exclusively to operate mobile networks, telecommunications providers could deploy infrastructure capable of supporting both network functions and commercially valuable AI services, which would create the possibility of multiple revenue or cost-saving mechanisms arising from a single shared capital investment. An operator pursuing this path might benefit from improved radio-network capacity and efficiency, from lower operating costs through network automation, from revenue for hosting external inference workloads, from enterprise contracts for managed edge applications, and from premium service agreements that bundle assured connectivity with assured computing performance. However, these benefits must be evaluated individually and with skepticism. The existence of five possible revenue streams does not establish that any of the five will materialize at scale, and the history of operator diversification is a history of revenue streams that looked plausible on a slide and proved elusive in a market.
2.2. The Shared-Computing Model
One of the most interesting economic questions raised by AI-RAN is whether telecommunications infrastructure can use expensive computing equipment efficiently, and the question deserves careful treatment because the answer is less obvious than vendors sometimes imply. Network traffic varies according to location, time, customer behavior, and specific events: a cell serving a business district is busy during working hours and nearly idle at three in the morning, while a cell serving a stadium is idle most of the week and saturated for three hours on a Sunday afternoon. External AI workloads also have changing computational requirements, and some of them, such as overnight batch processing of video archives or model fine-tuning, are flexible about when they run. If radio workloads and external applications can share accelerated computing capacity without compromising network performance, the operator may improve utilization substantially, and improved utilization is the difference between a GPU that earns its keep and one that depreciates in a cabinet.
SoftBank’s February 2026 collaboration with Nokia directly addresses this issue, and it is especially valuable as evidence because it focuses on the economic mechanism behind AI-RAN rather than merely demonstrating that accelerated computing can run telecommunications software. The companies described an extension of SoftBank’s AITRAS Orchestrator that integrates Nokia Bell Labs’ AI-RAN External Compute Engine, a platform that brokers and manages computing resources, so that the orchestrator can optimize the allocation of resources across both internal RAN control and external customers’ AI workloads. External customers can access computing on the AI-RAN infrastructure on demand without owning their own platforms, and SoftBank stated plainly that the enhancement was intended to let it treat those resources “as a new revenue source.”[17] ABI Research characterized the announcement as the first commercial application of the AI-on-RAN concept and as a shift in AI-RAN’s economic purpose from a cost-efficiency measure to a platform-economics model, while cautioning that specific applications had not yet been formally defined and recommending that initial deployments focus on urban or industrial locations where edge AI workloads would be generated at a constant rate.[38] The October 20, 2026 agreement to extend this work into a commercial-network field trial in Japan in 2027, with the AITRAS Orchestrator integrated with Nokia’s MantaRay management platform to coordinate AI and RAN resources, moves the shared-computing model from laboratory to live traffic.[8] SoftBank’s chief technology officer Ryuji Wakikawa described the planned trial as
“a significant milestone in AI-RAN’s transition from concept to execution.” — Ryuji Wakikawa, Senior Vice President and CTO, SoftBank Corp. [8]
The underlying logic is straightforward. A computing platform purchased for network operations will have periods of available capacity, and instead of leaving that capacity unused, a properly designed system could allocate selected resources to commercial AI applications. There is, however, an important limitation that the elegance of the logic tends to obscure: telecommunications processing has demanding service obligations. Radio functions operate on timescales of microseconds and milliseconds, they are subject to regulatory and contractual reliability requirements, and they cannot be preempted or delayed to accommodate a third party’s video-analytics job. Computing resources therefore cannot be reassigned to external applications if doing so compromises critical radio-network functions, which means that the shared-computing model requires priority rules, workload isolation, performance guarantees, and mechanisms for rapidly reclaiming capacity when network demand increases. The Elisa trial, which ran voice over New Radio, 5G Advanced data, and video analytics simultaneously, and the Nokia–SoftBank trial, which will explicitly validate the preservation of carrier-grade RAN operational requirements while AI workloads share the platform, are the first serious tests of whether these isolation mechanisms work under real conditions.[6][39] Until they report results, the shared-computing model remains a well-reasoned hypothesis rather than a demonstrated business.
2.3. The Economics of Spectrum Efficiency
The financial opportunity extends beyond selling computing services, and in the near term the more certain value may come from a less glamorous source: artificial intelligence helping operators extract more useful capacity from spectrum they already own. Radio spectrum is scarce, valuable, and expensive. Operators spend tens of billions of dollars acquiring access rights, as AT&T’s $22.8 billion EchoStar transaction in the third quarter of 2026 illustrates, and they spend billions more on equipment capable of using assigned frequencies efficiently.[37] Improvements in spectral efficiency therefore possess considerable economic value even when they are expressed in single-digit percentages, because a 10 percent improvement in the efficiency of a spectrum portfolio worth tens of billions of dollars is, in effect, billions of dollars of spectrum created out of software.
Nokia and Deutsche Telekom’s laboratory work on AI-enhanced uplink channel estimation for the Physical Uplink Shared Channel, which carries user data from devices to the network, showed typical throughput improvements of approximately 10–15 percent in measurements conducted in Bonn and Oulu, with gains increasing as radio channel complexity increased, and the companies emphasized that the technique introduces AI directly into the real-time signal-processing chain of the receiver rather than into monitoring or management layers.[5] Deutsche Telekom’s framing was explicitly economic: the gains demonstrate how AI can help operators
“get more performance from the spectrum and networks they already own.” — Deutsche Telekom AG, Media Information [5]
These results should not be interpreted as demonstrated, network-wide commercial improvements. They are preliminary measurements from specified laboratory conditions, and the companies themselves describe field validation as the next step rather than as an accomplished fact. But the potential financial mechanism is important and, unlike external inference revenue, it does not depend on finding a customer. If an operator can serve more traffic using existing spectrum and infrastructure, it may defer certain network-expansion investments or improve customer experience without proportionate increases in expenditure, and that benefit accrues whether or not the operator ever sells a single inference cycle to a third party.
Ericsson has pursued the same prize by a different route, embedding AI models directly into basebands and radios running on its own silicon and delivering the capability as a software subscription called AI in RAN, which the company says has been proven in more than fifteen deployments and trials and delivers up to 20 percent higher downlink throughput and up to 10 percent better spectral efficiency, with T-Mobile targeting full commercial deployment of the AI-native scheduler in the third quarter of 2026.[16] Nokia, for its part, has set a public target of more than 100 percent spectral-efficiency gains by 2028 through staged software releases, of which roughly 20 percent has been demonstrated, 50 percent is targeted for 2027, and the remainder for 2028, and Futurum Group’s Nick Patience has observed that this is a roadmap target rather than a measured result and that Nokia’s claimed gains are five to ten times the scale of what Ericsson has already put into commercial service.[40] The gap between the two vendors’ numbers is itself informative: it reflects not only different measurement conditions but a genuine architectural wager about whether the more compute-intensive algorithms that GPUs make possible will deliver proportionally larger gains than the leaner algorithms that purpose-built silicon can run.
The economics of AI-RAN therefore contain two separate possibilities that should be kept distinct in every investment analysis. Internal value arises from better network performance and operating efficiency, and it is available today in commercially shipping software. External value arises from commercial AI services delivered using distributed computing infrastructure, and it remains, as of October 2026, in trial. A durable investment thesis would ideally demonstrate both, but a prudent one would be financed on the first and treat the second as an option.
2.4. From Connectivity Revenue to Intelligence-Service Revenue
A mobile operator traditionally charges customers for connectivity through subscriptions, data plans, or enterprise network agreements, and the entire apparatus of operator billing, pricing, and customer management is organized around those units. Tower Inference could create additional commercial arrangements that do not fit neatly into that apparatus, and understanding what those arrangements might look like is essential to understanding whether operators are organizationally prepared to sell them.
Consider a business operating hundreds of autonomous inspection machines across a metropolitan area. The business might require continuous wireless access across the entire area, occasional bursts of high-compute inference when a machine encounters something it cannot classify on its own, device authentication and identity management, location-sensitive workload routing so that inference happens at the nearest available facility, and emergency escalation to a human operator when confidence falls below a threshold. A telecommunications operator could potentially offer all of these capabilities through an integrated service agreement, and instead of selling a fixed amount of mobile data it might charge according to reserved inference capacity, completed AI processing tasks, the number of managed intelligent devices, availability and service-level commitments, enterprise-specific model hosting, location-sensitive network services, or end-to-end managed solutions that bundle all of the above.
The commercial opportunity is already influencing operator discussions at the highest level. During T-Mobile’s October 28, 2026 earnings call, analysts from Evercore ISI, BNP Paribas, and Goldman Sachs pressed management on how physical AI could be made to pay, and management described early monetization as involving integrated solutions rather than separate charges for connectivity and compute, with chief executive Srini Gopalan describing a world of millions of robots and drones as essentially a given and the company’s chief technology officer arguing that the infrastructure build-out to support physical AI needs at least three things: low-latency connectivity, inference at the edge, and, for many use cases, voice.[31] Saw’s October 27 essay added a particularly revealing dimension to the value proposition, noting that shifting some processing off the robot could reduce battery use, computing demands, heat, weight, and cost, and could let machines run more capable AI without needing more hardware or power.[41]
This is a particularly important development for the economics of Tower Inference, because it identifies a source of value that does not depend on the operator charging more for connectivity. If network-based intelligence allows a customer to purchase less expensive devices, the network operator may be able to participate in the resulting economic savings. A fleet operator could compare two alternatives. Under the first, every robot carries a relatively expensive onboard processing system sized for the most demanding task it might ever encounter. Under the second, each robot retains the essential local intelligence required for safety and basic autonomy but accesses additional computational capacity from a shared edge platform when it needs more. The second arrangement may reduce device cost, battery consumption, weight, and hardware-upgrade requirements, and because the onboard hardware is sized for the typical case rather than the worst case, the savings compound across a fleet. However, the second arrangement introduces recurring network and computing charges, a dependence on connectivity, and a new category of failure mode. The correct economic comparison must include the full lifecycle cost of both approaches, including the cost of the fallback behavior when the network is unavailable, and the advantage should not be assumed.
2.5. What the Third-Quarter 2026 Earnings Reports Say About the Economics of Distributed Intelligence
The most useful corrective to both the enthusiasm and the skepticism surrounding Tower Inference is to read the financial statements that the principal participants filed in the autumn of 2026, because those statements reveal both how strong the operators’ core businesses are and how little of their current economics depends on anything resembling AI-RAN revenue. The picture that emerges is of an industry with the balance-sheet capacity to experiment and the earnings momentum to be patient, but with no evidence yet that distributed inference has moved the needle on any income statement.
| Company (period) | Revenue | Profit measure | Capital investment | Notable AI-related disclosure |
| T-Mobile US (Q3 2026) | $23.58B, +7.4% YoY; service revenue $19.4B, +6% | Net income $3.46B, +27%; Core Adjusted EBITDA $9.6B, +11%; diluted EPS $3.22 | ~$10B 2026 capex plan maintained | Kinetic Edge with AWS; Figure AI, Serve Robotics, City of San Jose pilots; first AI-RAN call; raised FY guidance |
| Verizon (Q3 2026) | $34.3B, +1.5% YoY; mobility & broadband service revenue $23.5B, +2.9% | Net income $5.3B, +5.2%; adjusted EBITDA $13.8B, +8% (record) | $16.0–16.5B 2026 capex guidance maintained | $2B+ AWS agreement to scale distributed compute for agentic and generative AI; ~$4B of AI Connect contracts in six months |
| AT&T (Q3 2026) | $32.3B, +5.3% YoY; service revenue $26.3B, +3.8% | Adjusted EBITDA $12.5B, +5.4%; adjusted operating income $7.6B | $5.1B Q3 capex; $23–24B 2026 capital investment; $22.8B Q3 spectrum acquisition | Fiber to 60M+ locations by 2030; converged fiber-and-5G strategy positioned “for the AI era” |
| Nokia (Q3 2026) | €5.154B, +10% reported, +9% constant currency | Comparable operating profit €605M, +35%; comparable operating margin 11.7%, +220 bps; free cash flow −€99M | n/a | AI & Cloud customer net sales +123%, order intake +118% to €1.5B; Mobile Infrastructure +5%; AI-RAN pilots end-2026, commercial 2027 |
| Ericsson (Q3 2026) | SEK 59.5B, +6% reported, +7% organic | EBIT SEK 8.9B, 14.9% margin (vs. 26.9% prior year, which included one-time items); net income SEK 5.8B | n/a | AI in RAN software subscription live in 15+ deployments; purpose-built silicon strategy |
| NVIDIA (Q2 FY2027, ended July 26, 2026) | $96.2B, +106% YoY; Data Center $89.0B, +117% | GAAP net income $59.7B; gross margin 75.0% | n/a | Edge Computing revenue $7.2B, +27% YoY; Q3 guidance $108B; ACIE customers $40.3B, +138% |
Table 2. Third-quarter 2026 financial results of principal Tower Inference participants, compiled from company earnings releases and SEC filings. NVIDIA’s fiscal second quarter 2027 is its most recent reported period as of October 31, 2026. Sources: T-Mobile US Form 8-K; Verizon 3Q 2026 earnings release; AT&T Form 8-K; Nokia Interim Report Q3 2026; Ericsson Third Quarter Report 2026; NVIDIA Q2 FY2027 release.[35][36][42][43][44][20]
Several observations follow from the table. First, the three largest American operators are all growing service revenue, expanding margins, and raising guidance, which means that none of them is pursuing AI-RAN out of financial desperation; they are pursuing it from a position of strength, which historically has produced more patient and more disciplined experimentation than crisis-driven diversification. T-Mobile’s chief executive described the company’s position in terms that place AI squarely in the growth narrative without depending on it:
“We have a tremendous growth runway ahead across core wireless, broadband, and smart adjacencies.” — Srini Gopalan, CEO, T-Mobile US [35]
Second, the most concrete AI-related revenue that any operator disclosed in the quarter had nothing to do with the radio network at all. Verizon’s agreement with AWS, valued at more than $2 billion and described as a commercial framework to scale distributed compute and meet the real-time demands of agentic and generative AI workloads, is substantially a fiber-connectivity agreement, following a July 2026 dark-fiber agreement with Google worth more than $1 billion; Schulman told analysts that Verizon had signed roughly $4 billion of such AI Connect contracts in six months.[45] He described the pattern in which the largest builders of AI infrastructure are
“choosing our fiber network to power the next generation of AI infrastructure.” — Dan Schulman, CEO, Verizon [45]
This is an important and somewhat sobering data point for Tower Inference, because it suggests that the first wave of operator AI revenue is accruing to the layer of the network that connects hyperscale datacenters to one another, not to the layer that connects devices to distributed inference. The road to AI revenue, for Verizon at least, currently runs through long-haul fiber rather than through GPU-equipped cell sites.
Third, the vendors’ results reveal the asymmetry of the AI opportunity within the telecommunications supply chain. Nokia’s net sales to AI and cloud customers grew 123 percent, but that growth came overwhelmingly from optical and IP networking equipment sold into datacenter build-outs, while its Mobile Infrastructure segment grew a modest 5 percent and the company’s free cash flow turned negative as working capital and restructuring charges absorbed the operating improvement.[43][46] Ericsson grew organically by 7 percent on the strength of network modernization and mission-critical networks, but its reported margins compressed sharply against a prior-year quarter that had included large one-time gains.[44] For both vendors, AI-RAN is a strategic bet on the next cycle, financed by a current business that is healthy but not booming. NVIDIA’s results, by contrast, show a company for which telecommunications is a rounding error on a $96 billion quarter, and its chief executive described the current moment in terms that explain why the company is nonetheless investing in the sector:
“Now, compute is revenue.” — Jensen Huang, Founder and CEO, NVIDIA [20]
If compute is revenue, then every location that can host compute is a potential revenue location, and the telecommunications network has more such locations than any other infrastructure in the economy. NVIDIA’s $1 billion equity investment in Nokia, announced October 28, 2025, at $6.01 per share for a 2.9 percent stake, should be read in this light: it is a relatively small cheque for NVIDIA, which one analyst called a rounding error next to its OpenAI and Intel commitments, but it bought credibility for an AI-RAN effort that had seen limited uptake since the Aerial platform was first announced in 2019.[47] As Fierce Network’s analyst commentary put it:
“Before this deal, Nvidia Aerial was largely a SoftBank showcase.” — Fierce Network (analyst commentary) [47]
The earnings evidence, taken together, supports a measured conclusion. The operators have the financial capacity to pursue Tower Inference, the vendors have the strategic motivation, and NVIDIA has the platform and the patience. But the revenue that would validate the thesis has not yet appeared, and the AI revenue that has appeared is flowing to fiber rather than to the radio edge.
2.6. A Proposed Tower Inference Investment Framework
This paper proposes evaluating a distributed inference deployment through the following economic relationship, which is intended as a discipline for operator investment committees rather than as a predictive model:
VTower = RAI + Vnetwork + Soperations − Cincremental − Crisk
Here VTower is the net value of the deployment over its economic life. RAI is revenue attributable to external AI and enterprise services, which must be contracted or credibly forecast rather than assumed from vendor market-size estimates. Vnetwork is the financial value of improved radio performance or deferred network investments, which should be computed from measured spectral-efficiency gains applied to the operator’s actual capacity-expansion plan. Soperations represents verifiable savings from automation and efficiency, net of the cost of the automation itself. Cincremental is the additional cost of computing hardware, deployment, power, cooling, networking, maintenance, orchestration software, and the reserve capacity that must be held back to protect radio functions. Crisk is the expected cost associated with outages, security incidents, contractual penalties, underutilization, and technological obsolescence.
The framework deliberately separates revenue, operational improvements, and financial risks, and it deliberately refuses to assume that all available computing capacity can be sold externally. A credible investment case should be evaluated against the operator’s existing infrastructure and operating obligations, site by site and workload by workload, because the terms of the equation vary enormously across the computing hierarchy described in Table 1. An AI-RAN project might reduce certain network costs while failing to generate meaningful external revenue, in which case Vnetwork and Soperations must carry the investment alone. Another deployment might attract substantial enterprise demand but require such expensive computing equipment and such extensive facility upgrades that Cincremental overwhelms RAI. A third might work economically at a regional aggregation facility, where power and cooling are available and many cells can share the capacity, but fail at individual cell sites because of electrical, thermal, and maintenance constraints. These differences imply that Tower Inference will develop through selective deployment rather than uniform construction across every radio location, and any vendor roadmap that assumes otherwise should be discounted accordingly.
The framework also invites a broader question about whether artificial intelligence will generate the aggregate economic value on which all of these revenue hopes ultimately rest, and here the economics profession is sharply divided. MIT’s Daron Acemoglu, the 2024 Nobel laureate, has estimated that AI will deliver roughly 0.55 percent of total factor productivity gains over the coming decade, a figure far below the expectations embedded in current infrastructure investment, and he has repeatedly warned that the gap between laboratory capability and deployed productivity is wider than the industry presumes:
“it’s not as easy as sometimes it’s presumed.” — Daron Acemoglu, Institute Professor, MIT; 2024 Nobel Laureate in Economic Sciences [48]
Stanford’s Erik Brynjolfsson, by contrast, has argued that the United States is now emerging from the trough of the productivity J-curve, in which heavy initial investment in a general-purpose technology obscures early gains before they appear in aggregate statistics, and he pointed to 2025 productivity growth of roughly 2.7 percent, nearly double the ten-year average, as evidence that the
“fog may finally be lifting.” — Erik Brynjolfsson, Director, Stanford Digital Economy Lab [49]
The relevance of this debate to Tower Inference is direct. If Acemoglu is right, the enterprise demand for distributed inference will grow slowly and selectively, and operators should size their investments to the internal value of AI-for-RAN with external revenue treated as upside. If Brynjolfsson is right, the demand will arrive faster than infrastructure can be built, and operators who have positioned distributed capacity in the right metropolitan markets will find themselves with a scarce and valuable asset. The investment framework above is designed to produce sensible decisions under either scenario, by insisting that the certain terms be measured and the uncertain terms be bounded.
2.7. The Competitive Risk of Technological Obsolescence
Distributed AI computing also exposes telecommunications operators to a financial risk that is familiar in the broader AI infrastructure industry but novel in the telecommunications context: rapid technological change in the computing substrate. Radio systems often remain operational for many years, and the depreciation schedules, financing structures, and procurement cycles of the telecommunications industry are built around asset lives measured in a decade or more. Accelerated computing platforms, by contrast, may experience much shorter periods of economic competitiveness, as successive chip architectures, model optimizations, and software systems improve performance per watt and per dollar at a pace that has repeatedly surprised even the companies producing them. NVIDIA has moved from Hopper to Blackwell to Vera Rubin in roughly three years, and each generation has rendered its predecessor economically inferior for new deployments even while the older hardware remains perfectly functional.
This creates a possible mismatch between telecommunications investment cycles and AI equipment economics. A provider installing expensive GPUs at distributed sites in 2027 must consider whether those systems will remain cost-effective in 2029 or 2030, when the applications they were purchased to serve may be running on models three times as efficient and competitors may be deploying accelerators with several times the throughput. The financial analysis should therefore distinguish physical equipment life from economic usefulness, because an accelerator may remain functional while becoming substantially less competitive against newer equipment, and the Crisk term in the investment framework must include the probability-weighted cost of early replacement.
Operators should evaluate upgrade costs, software portability, hardware replacement logistics across hundreds of unstaffed sites, utilization trends, and potential residual values before committing to large distributed deployments, and they should weigh the counterargument that vendors advance in favor of the GPU model: that a software-defined radio running on general-purpose accelerators can be upgraded from 5G Advanced to 6G through software rather than through a forklift replacement of baseband hardware, which Nokia and NVIDIA have explicitly positioned as a central benefit of the architecture.[50] If that promise holds, the shorter economic life of the accelerator may be offset by the longer economic life of the software investment that runs on it. The most important economic advantage may ultimately come from flexible infrastructure that allows computing systems to evolve without rebuilding the underlying telecommunications network, and the vendor that can credibly demonstrate that flexibility, rather than the vendor with the fastest chip, may win the operator’s confidence.
The second major proposition of this paper is therefore that Tower Inference becomes commercially attractive only when the combined value of network improvement, operational savings, and externally purchased AI services exceeds the incremental cost and risk of distributed computing, evaluated honestly and site by site, with the obsolescence risk of the computing substrate priced in rather than wished away.
Section 3: Physical Intelligence Applications — Robotics, Autonomous Machines, Emergency Communications, and Real-Time Coordination
The economic analysis of Section 2 concluded that external AI revenue is the least certain term in the Tower Inference equation, and that it will materialize, if it materializes at all, from applications whose characteristics genuinely reward proximity. This section examines those applications in detail. It begins with the conceptual shift from digital to physical artificial intelligence and the architectural question that shift raises about where intelligence should reside, then proceeds through humanoid robotics, autonomous delivery fleets, industrial automation, autonomous vehicles, emergency communications, and network sensing, drawing on the specific pilots that T-Mobile, Deutsche Telekom, Elisa, and Orange disclosed in October 2026. It concludes with the principle that must govern every one of these applications: safety must survive the failure of the network.
3.1. Why Physical Artificial Intelligence Changes the Infrastructure Question
The next stage of artificial intelligence will increasingly involve systems that interact with physical environments, and this represents an important departure from applications primarily designed to generate text, images, software code, and digital analysis. A physical AI system must interpret changing surroundings and produce outputs that influence real-world operations in real time, and the consequences of its mistakes may therefore extend beyond incorrect information to physical harm. An industrial robot can damage equipment or injure a worker. An autonomous vehicle can encounter road conditions that no training set anticipated. A delivery robot can obstruct a wheelchair ramp. A drone can lose connectivity over a crowd. These applications introduce stronger requirements for predictable performance, resilience, system integrity, and operational safety than any chatbot ever faced, and they create a fundamental architectural question: how much intelligence should reside inside an autonomous device, and how much should be supplied by external computing infrastructure?
Fei-Fei Li has described the current generation of language models, for all their fluency, as fundamentally ungrounded in the three-dimensional world in which bodies move, and she has argued that physical data is scarce, unstandardized, and hard to observe in ways that text is not, which means that spatial AI will not scale at the pace that language models did and will require different infrastructure to collect its data and to deploy its models.[51] Qualcomm’s executives, approaching the same problem from the device side, have described a transition from centralized AI models in datacenters toward distributed, autonomous agents in devices, cars, wearables, and humanoid robots, and have been blunt about the implications for the network:
“It’s not enough just to add more compute. We need more advanced networks.” — Qualcomm executive, Snapdragon Summit [52]
An entirely onboard architecture offers independence from network connectivity but is limited by available power, processing capacity, heat dissipation, and physical size, and those limits bind hardest precisely on the machines, humanoids, drones, and small delivery robots, whose form factors leave the least room for computing hardware. An entirely remote architecture permits access to arbitrarily powerful models but introduces dependence on connectivity, network reliability, and external processing availability, and that dependence is intolerable for any function whose failure could cause physical harm. The likely solution for many applications is a hybrid system in which safety-critical perception and control functions remain onboard, while larger reasoning models, fleet coordination, environmental analysis, and selected advanced services operate at nearby or regional computing facilities. The precise division will depend on the application, on the latency distribution described in Section 1.4, and on the cost comparison described in Section 2.4. Tower Inference is the potential infrastructure supporting that division, and the applications examined below are the first experiments in drawing the line.
3.2. Humanoid Robots and the Economics of Shared Intelligence
Humanoid robotics provides the most vivid example of the hybrid architecture, because the physical constraints of a humanoid robot differ so sharply from those of a datacenter. A robot must carry its own power supply, and every watt consumed by onboard computing is a watt not available for locomotion and manipulation. It must manage heat within a compact body that is often sealed against dust and moisture. Additional computing equipment increases weight, which increases the energy required to move, which increases the battery required, which increases the weight further. And every component adds cost to a machine whose economic case depends on being cheaper than the human labor it supplements. These constraints create a powerful interest in distributing selected computing tasks beyond the robot itself.
On October 27, 2026, T-Mobile’s John Saw described the company’s work with Figure AI in terms that make the architectural question concrete. The relationship began with connectivity, in the form of eSIM service for Figure’s humanoid robots, progressed to supporting the robots’ training loop, and has now reached the point at which T-Mobile and Figure are assessing whether portions of the Helix S1 and S2 models, the systems that control the robots’ movement and interpret visual information and language instructions, can run on compute closer to the edge while preserving the responsiveness that machines in the physical world require.[9] Saw described the broader strategic logic as a recognition that physical AI needs intelligence in more than one place and that the network is what connects those places into a working system, and he was explicit that intelligence will live on the device, at the edge, and in the cloud simultaneously.[9]
The commercial significance lies in the possibility of separating basic robotic autonomy from supplemental intelligence. A robot could retain the capabilities required for immediate perception, balance, collision avoidance, and essential control, all of which must function at millisecond timescales regardless of connectivity, while more computationally intensive functions, such as complex scene interpretation, open-ended language interaction, coordination among multiple robots, or computationally demanding task planning, are supported by shared infrastructure. A distributed architecture of this kind could potentially improve the amount of useful intelligence available to a robot without requiring equivalent increases in onboard computing hardware, and because the shared infrastructure serves many robots, the capital cost of that intelligence is amortized across a fleet rather than borne by each machine.
Nevertheless, the engineering limitations are substantial and should not be minimized. A robot cannot depend on remote inference for a safety-critical action that must occur faster than the network can reliably respond, and the tail of the latency distribution, not its mean, defines what “reliably” means. Moreover, many advanced robotic models, including the vision-language-action architectures that Figure’s Helix family represents, maintain continuous internal state and require closely coordinated processing across perception and control, and partitioning such a model between a robot and a network facility may introduce communication overhead, synchronization complexity, and failure modes that offset the benefits of external computing. The research question that T-Mobile and Figure are actually investigating is therefore not whether robotic intelligence can migrate to the network but which specific robotic functions are suitable for remote execution, at what latency, with what fallback behavior, and at what cost. The answer to that question, when it is published, will be one of the most important pieces of evidence in the entire Tower Inference debate.
3.3. Autonomous Delivery Fleets: From Individual Intelligence to Collective Coordination
Autonomous delivery robots introduce another commercially relevant use case, and one whose economics are already being tested at scale on the sidewalks of American cities. The economics of a delivery fleet depend on the number of completed deliveries per robot per day, on robot utilization, on maintenance costs, on the frequency and cost of human intervention, and on the ability to operate reliably across diverse and changing environments. A small fleet may be manageable with significant human supervision, with remote operators watching feeds and intervening when a robot encounters something it cannot handle. A much larger fleet requires more efficient methods of identifying unusual circumstances and coordinating responses, because the ratio of human supervisors to robots cannot scale linearly without destroying the economic case for automation in the first place.
T-Mobile’s announced collaboration with Serve Robotics addresses exactly this scaling problem. The company is evaluating the feasibility of conducting monitoring at the network edge in order to assess which problems can be solved by humans and which need more AI input, and it is exploring an edge-AI layer that could handle the kinds of unexpected obstructions, a construction barrier that appeared overnight or a wandering pet, that would otherwise require human attention, along with more natural conversational interaction between robots and the people they encounter.[9][41] Saw noted directly that employing people to assist robots becomes challenging when thousands of machines are involved.[41]
The potential role of Tower Inference is especially interesting when considered at fleet scale, because fleet scale is where the economics of shared infrastructure are most favorable. Rather than treating every robot as an isolated intelligent device, an operator could use distributed infrastructure to support a collective system. A regional inference platform might process selected events from many robots, identify recurring obstacles and share them across the fleet, coordinate route changes in response to a street closure, or decide when human intervention is necessary and route the request to an available supervisor. This produces a possible economic advantage that individual onboard computing cannot match: computational resources shared across devices whose needs vary over time, so that the capacity required to handle the fleet’s peak aggregate demand is far smaller than the sum of the capacity each robot would need to handle its own worst case.
However, sharing is beneficial only when the demand profiles, network conditions, and model-serving architecture actually permit it, and fleet-level deployments introduce correlated risks that individual deployments do not. A fleet of robots encountering difficult situations simultaneously, as they might when a parade closes a dozen streets at once, might create computing peaks that exhaust the shared capacity. A weather event could produce correlated demand across an entire metropolitan area. A communications disruption could affect many devices in the same geographic area at the same moment, which is exactly the moment when they most need guidance. Fleet-level intelligence therefore requires careful capacity planning, resilience engineering, and graceful degradation, and a particularly important research question concerns the ratio between the number of robots operating in a region and the amount of shared inference capacity needed to support them under realistic correlation assumptions. That ratio could become a new planning variable for telecommunications companies entering the physical-AI market, as fundamental to their business as the ratio of subscribers to cell sites has been for forty years.
3.4. Industrial Automation: When the Factory Becomes a Distributed Intelligent System
Manufacturing represents another promising environment for edge intelligence, and in some respects the most natural one, because modern industrial facilities already contain machinery, sensors, inspection systems, robotic equipment, and computerized control platforms, and many already rely on sophisticated local networks and industrial computing systems. The question in manufacturing is not whether intelligence should be distributed but whether the telecommunications operator’s distributed infrastructure adds anything to the distributed infrastructure that the factory already has.
Tower Inference could complement existing industrial arrangements by providing additional processing resources and coordinated services, particularly for applications that span multiple facilities or that extend beyond the factory fence. Possible applications include equipment inspection, predictive maintenance, machine-vision analysis, inventory monitoring, and coordination among mobile industrial robots that move between buildings or across a campus. Deutsche Telekom’s expanding Industrial AI Cloud, which hosts open models including NVIDIA Nemotron and serves robotics companies such as Agile Robots, and its collaboration with NVIDIA, Nokia, and Supermicro on distributed infrastructure for physical-AI applications, are the clearest European expression of this direction.[1][53] In the United States, T-Mobile and NVIDIA described pilot activity involving physical-AI developers and industrial applications, including automated utility inspections and computer-vision systems operating on distributed edge infrastructure, earlier in 2026.[9]
The important distinction in industrial settings is between operational inference and industrial safety control. Machine-learning models may identify maintenance problems, flag anomalies in inspection images, or recommend corrective actions, and these functions tolerate latency measured in hundreds of milliseconds or even seconds. But many industrial control systems require deterministic performance and validated safety protections, often governed by functional-safety standards, that should not depend on an ordinary wireless connection or on a shared computing platform whose capacity may be claimed by another tenant. A sophisticated deployment would therefore separate immediate machine protection and emergency controls, which remain in hard-wired or dedicated local systems; local industrial computing and automation, which remain in the facility; telecommunications-based inference and coordination, which the operator’s distributed infrastructure might supply; regional analytics and enterprise applications, which might run in a regional datacenter; and hyperscale training and model development, which remain in the cloud.
This architecture is particularly relevant to the Five-Layer AI Economy because it connects Layer 5 industrial applications with geographically distributed Layer 3 infrastructure through a chain of explicit design decisions rather than through a single undifferentiated “edge.” The economic benefits may arise from better maintenance decisions, improved worker productivity, fewer inspection delays, and more efficient management of complex multi-site operations, and these benefits should be measured through actual industrial outcomes, such as reduced unplanned downtime or improved first-pass yield, not simply through the volume of AI processing performed. A fixed manufacturing facility with an established private 5G network and on-premises computing may need very little from a public operator’s distributed inference platform; the operator’s value is likely to be greatest for industrial customers whose assets are mobile, whose operations are geographically dispersed, or whose scale does not justify building their own edge.
3.5. Autonomous Vehicles and the Limits of Network-Based Intelligence
Autonomous vehicles are frequently mentioned as potential beneficiaries of edge computing, and the superficial logic is appealing: vehicles require sophisticated perception, localization, prediction, planning, and control, and all of these are computationally intensive. However, a fundamental safety principle must govern this discussion and should be stated without qualification. An autonomous vehicle should not depend on a remote mobile-network response to perform immediate safety-critical driving actions. Braking, steering, obstacle avoidance, and collision mitigation must remain available, at full capability, when network service is interrupted, degraded, or delayed, because network service will be interrupted, degraded, and delayed, and a vehicle that requires the network to avoid a pedestrian is a vehicle that will eventually fail to avoid one.
Tower Inference therefore should not be presented as a replacement for onboard vehicle intelligence, and the credible vendors do not present it that way. Its more realistic role involves supplemental functions whose failure degrades convenience rather than safety: sharing selected infrastructure observations, such as a roadside camera’s view of an occluded intersection; providing updated information about road conditions, construction, and weather; supporting fleet-management and dispatch decisions; processing less time-critical environmental data for mapping and model improvement; coordinating logistics, charging, and depot operations; and delivering additional contextual information to vehicle systems that can use it when available and ignore it when not. Andrews’s observation that vehicles need to know about things outside their own field of vision points exactly to this supplemental role: the network can extend what the vehicle sees, but it cannot be allowed to become what the vehicle depends on.[32]
Some future applications could benefit from cooperative sensing, in which vehicles and roadside infrastructure share information to build a common picture of a complex environment, and the integrated sensing and communications capabilities being developed for 6G could eventually allow the radio network itself to contribute to that picture. But these systems require robust security against spoofed observations, data-quality assurance, communication standards that span manufacturers and jurisdictions, and clear rules concerning responsibility when shared information is wrong. The distinction between advisory information and operational control is especially important here, and it should be reflected in liability frameworks as well as in system architecture. The future of autonomous mobility may involve substantial cooperation among onboard processors, mobile networks, roadside computing, and hyperscale infrastructure, but that cooperation does not imply that every component should exercise authority over the vehicle, and the component that exercises authority must be the one that cannot be disconnected.
3.6. Emergency Communications and Real-Time Voice Intelligence
A particularly practical application of distributed intelligence, and one that may arrive commercially well before autonomous machines do, concerns emergency communications and real-time voice. Public-safety operations often involve rapidly changing circumstances, language barriers, incomplete information, and an absolute requirement for reliable communications, and AI-supported speech recognition, translation, summarization, and situational assistance could improve selected workflows in ways that save lives.
On October 21, 2026, Orange Business and NVIDIA described a pilot of distributed voice-AI inference across multiple sites in Orange’s network, with NVIDIA Blackwell systems in the central cloud and Aerial RAN Computer platforms at distributed locations, in which an AI Distributed Inference platform manages workloads across central-cloud and regional points of presence and models are accessed through Orange’s Live Intelligence multi-LLM platform.[7] The demonstration used a wildfire emergency scenario in which AI translates in real time between a foreign-language caller and an emergency operator while also transcribing the call and preparing summaries that capture key information for dispatch. NVIDIA’s Ronnie Vasishta described the strategic logic from the vendor’s perspective:
“Telcos are uniquely positioned to become distributed infrastructure service providers.” — Ronnie Vasishta, Senior Vice President of Telecom, NVIDIA [7]
A companion proof of concept developed by Gcore, Orange Business, and NVIDIA for fire and rescue call handling makes the design philosophy explicit, and its design philosophy is the correct one. The emergency call never depends on the AI assistant; if the AI is unavailable, the operator carries on exactly as they would without it. The system is human-supervised, the configuration is approved and version-controlled across every point of presence, and the motivating principle is stated in the title of the work itself:
“In an emergency call, a late answer is a missing answer.” — Gcore, Orange Business and NVIDIA, Fire and Rescue Call-Handling Proof of Concept [54]
This suggests that some commercially useful Tower Inference services may emerge from voice and communications applications before more ambitious autonomous-machine systems become widespread, and voice applications offer several structural advantages for an operator. They integrate naturally with existing communications services that operators already sell. They can use smaller, specialized speech and translation models rather than frontier reasoning systems, which means they can run on a single edge server. They have clear enterprise use cases in contact centers, field service, and public safety. And they generate measurable operational benefits, such as reduced translation delays or improved information accessibility, against which the value of the service can be priced. Elisa’s distributed inference ambitions in Finland similarly list real-time voice AI alongside video surveillance and industrial automation as candidate services, and its chief operating officer framed the broader thesis:
“AI is becoming a critical layer of digital infrastructure.” — Sami Komulainen, COO and EVP, Technology and Operations, Elisa [6]
Nevertheless, emergency applications carry substantial risks that the enthusiasm around voice AI should not obscure. Translation errors in an emergency may have grave consequences. Speech recognition performs differently across accents, languages, noise conditions, and technical vocabulary, and the conditions of an emergency call, a panicked caller, background noise, a poor connection, are precisely the conditions under which it performs worst. Unreliable AI summaries may omit the single critical detail that a human dispatcher would have caught. Deployment should therefore incorporate rigorous quality testing across demographic and acoustic conditions, human verification where appropriate, complete recordkeeping, and explicit limitations on automated decision-making. A communications network that becomes an inference platform must preserve its traditional reliability obligations, which in the emergency context are among the most stringent in any industry, while taking on new responsibilities for the accuracy of AI outputs that it has never previously borne.
3.7. Telecommunications Networks as Environmental Sensors
One of the less-explored possibilities of Tower Inference extends beyond conventional inference services into a capability that only the radio network can supply: integrated sensing and communications. In such architectures, the radio signals used for communications also contribute information about aspects of the physical environment, so that the network becomes a distributed sensor as well as a distributed computer. Nokia’s October 21 announcement identified integrated sensing and communications as part of its broader AI-RAN development and demonstrated at GTC Berlin how the RAN can detect and understand activity in the physical environment, and it described an open ecosystem of applications being built on the platform, including DeepSig’s spectrum-sensing and radio-frequency threat-detection applications and ZaiNar’s positioning software, which enables precise network-based location using existing wireless infrastructure without requiring changes to connected devices or radio hardware.[55] 3GPP’s Release 20 study program includes integrated sensing and communications among the usage scenarios being examined for 6G, which means that what is today a vendor demonstration could become a standardized network capability by the end of the decade.[56]
These activities suggest a potentially important development. Telecommunications networks may become platforms not only for transmitting data and processing intelligence but also for generating selected categories of environmental information, which could create applications in logistics, industrial operations, infrastructure monitoring, drone detection, and public safety. Qualcomm has described a sensing-enabled “aerial economy” in which networks interpret radio signals to detect objects and environmental conditions for city-scale mapping and drone detection, and T-Mobile’s work with the City of San Jose, which uses edge AI to combine the views of separate 5G-connected cameras into an integrated picture of traffic, is a step toward the same kind of network-derived situational awareness built from conventional sensors.[57][9]
However, sensing functionality raises additional concerns that are distinct from, and in some ways more serious than, those raised by inference. Network-derived information about device locations, movements, or surrounding activity can become extraordinarily sensitive, and even when an application does not involve conventional cameras, it may create privacy implications that users neither expect nor consent to. A network that can locate devices precisely without their cooperation, or that can detect the presence and movement of people who are not carrying devices at all, is a network with surveillance capabilities that no previous generation of telecommunications infrastructure possessed. The commercial use of such information requires appropriate controls, lawful processing, transparency, and safeguards against unauthorized surveillance, and the expansion of network sensing should therefore be treated as a distinct governance challenge within Tower Inference rather than folded into the general discussion of edge computing.
3.8. The Fundamental Requirement: Safety Must Survive Network Failure
The greatest engineering mistake in the entire Tower Inference enterprise would be to assume that lower average latency automatically permits unrestricted remote control of physical machines. Average latency is a statistic about the typical case; safety is a property of the worst case. A safe distributed architecture must account for equipment failures, service interruptions, congestion, software errors, cyberattacks, model drift, and changes in model behavior after updates, and it must do so not by hoping these events are rare but by designing every physical-AI application so that the machine remains safe when they occur.
Applications should therefore identify, explicitly and in advance, which functions can safely depend on external infrastructure and which cannot, and the dividing line should be drawn by the consequence of failure rather than by the convenience of the architecture. For a physical-AI system, the relevant performance measures include far more than the time required to generate a model output. They include the end-to-end response-time distribution across percentiles; network availability and geographic coverage along the machine’s actual routes; performance during congestion and mass events; model reliability under distributional shift; recovery time after component failure; security and identity verification for every instruction the machine receives; graceful degradation when external computing becomes unavailable; and the ability to maintain essential local safety functions indefinitely without any network support at all.
These requirements may limit the commercial feasibility of certain applications, and that limitation should be welcomed rather than resisted, because it channels Tower Inference toward the applications where it belongs. Yet the same requirements also create opportunities for telecommunications operators capable of delivering differentiated reliability and service assurance, because an operator that can credibly guarantee the tail of the latency distribution, document its coverage along a fleet’s routes, and demonstrate its failure-recovery behavior is selling something that a hyperscale cloud, reached over the public internet, cannot sell. The third major proposition of this paper follows directly: physical AI will create demand for geographically distributed intelligence, but the most valuable deployments will preserve local autonomy while using network inference to extend capability, coordination, and efficiency, and the operators who understand that distinction will build businesses while those who ignore it will build liabilities.
Section 4: Infrastructure Competition — NVIDIA, Nokia, Ericsson, Cisco, the Hyperscalers, and the Contest for Distributed AI
The applications examined in Section 3 will be delivered, if they are delivered, through a supply chain that is being contested as vigorously as any in the technology industry. Telecommunications operators own the network, but they do not own the accelerators, the radio software, the orchestration layer, the cloud platforms, or the models, and the allocation of economic value among the companies that do own those things will determine whether Tower Inference enriches the operators or merely passes through them. This section examines the competitive position of each category of participant, beginning with the operators’ distinctive assets, proceeding through NVIDIA’s platform strategy, the divergent paths of Nokia and Ericsson, Cisco’s bid to control inference routing, and the ambiguous role of the hyperscalers, and concluding with the strategic importance of software portability and the international dimension of the contest.
4.1. Telecommunications Operators Possess a Different Infrastructure Advantage
Hyperscale technology companies possess formidable advantages in computing infrastructure. They control enormous datacenter campuses, extensive private fiber networks, mature cloud platforms, specialized software, model-serving systems, and sophisticated developer ecosystems, and their economies of scale in purchasing accelerators and electricity are unmatched. NVIDIA’s own disclosure that its AI Clouds, industrial, and enterprise customers accounted for $40.3 billion of quarterly revenue, up 138 percent, alongside its hyperscale customers, illustrates how concentrated the demand for accelerated computing remains among a small number of very large buyers.[58]
Telecommunications operators possess a different collection of assets. They have established billing relationships with hundreds of millions of mobile customers, distributed physical facilities in every market they serve, licensed spectrum that no cloud provider can replicate, deep radio-network expertise, and networks that reach geographically dispersed locations that no datacenter campus will ever be built near. The opportunity for Tower Inference emerges from combining these assets with accelerated computing, and the operators’ strongest position is in applications where connectivity and computing are complements rather than substitutes.
Importantly, operators should not assume that owning a radio network automatically provides a competitive advantage in AI services. They must establish that the network contributes something commercially valuable that customers cannot obtain as effectively through existing cloud services, private industrial networks, or local computing systems. Possible differentiators include geographic proximity and the latency it affords, network-controlled quality of service that can prioritize an inference request over a video stream, mobility management that keeps a session alive as a robot moves between cells, local data processing that satisfies sovereignty and privacy requirements, and integrated operational support that a cloud provider cannot offer for a physical device in the field. The strength of these advantages will vary by application. A fixed manufacturing facility with an established private network may not need a mobile operator’s distributed inference platform at all, while a delivery robot moving across an urban area may benefit directly from services integrated with a wide-area mobile network. This distinction suggests that telecommunications operators should initially prioritize applications where connectivity and distributed computing produce complementary value, and T-Mobile’s chosen pilots, humanoid robots, sidewalk delivery robots, and city-wide camera networks, are all applications of exactly that kind.
4.2. NVIDIA’s Strategy: Extending Accelerated Computing Beyond Hyperscale Datacenters
NVIDIA’s telecommunications strategy reflects its broader interest in expanding the addressable market for accelerated computing beyond the hyperscale customers who currently dominate its revenue. Historically, much of the company’s AI-related commercial prominence has been associated with GPUs deployed in large datacenters, and the telecommunications network, with its hundreds of thousands of sites and its tens of thousands of aggregation facilities, represents a potential source of demand that is both large and structurally distinct from the hyperscale market. Jensen Huang has framed the opportunity in national-strategic terms, describing telecommunications as critical national infrastructure and AI-RAN as
“a generational platform shift” — Jensen Huang, Founder and CEO, NVIDIA [50]
that could allow the United States to regain leadership in an infrastructure technology where its vendors have lost ground over two decades.
At GTC Berlin, NVIDIA introduced the MGX 2U and MGX 4U configurations of its Aerial RAN Computer for high-capacity sites, powered by RTX PRO 6000 Blackwell Server Edition GPUs, with hardware partners including Dell Technologies, Quanta Cloud Technology, and Supermicro and with Red Hat bringing the stack onto OpenShift.[1] The architecture combines accelerated computing with software-defined radio capabilities and the capacity to run external AI applications, and NVIDIA has cited Omdia’s forecast that the RAN market will exceed a cumulative $200 billion by 2030 with AI-RAN as its fastest-growing segment.[13] The economic implication for NVIDIA is significant: if operators adopt accelerated computing across even a fraction of their networks, infrastructure spending on GPUs extends beyond conventional hyperscale procurement into a new and recurring customer base, and NVIDIA’s software, including Nemotron open models and the Agent Toolkit now integrated into Nokia’s and Cisco’s network-management products, extends with it.[1]
However, this also raises a question concerning platform dependence that operators should weigh carefully. NVIDIA supplies the hardware, the CUDA software environment on which Nokia’s radio software now depends, model technologies, and other infrastructure components. A telecommunications operator that builds extensively around one accelerated computing ecosystem may gain substantial integration benefits and a single accountable partner, but it may also face significant costs when attempting to introduce alternative hardware, migrate workloads, or change suppliers, and the history of the telecommunications industry’s relationship with its equipment vendors is a long record of operators discovering too late how expensive lock-in can be. The long-term question is whether a common accelerated platform will improve operator bargaining power by creating new service opportunities, or concentrate more technological influence in the hands of a single infrastructure supplier whose interests are not identical to the operator’s.
4.3. Nokia, Ericsson, and the Two Paths Toward AI-Native Networks
The two largest Western radio-equipment vendors have made opposite architectural bets, and the contest between them is the most consequential industrial rivalry within Tower Inference. Nokia’s strategy emphasizes software-defined radio networks running on NVIDIA’s merchant silicon, shared accelerated computing between RAN and AI workloads, and the integration of AI applications with communications infrastructure. In July 2026 it announced what it called the industry’s first commercial AI-RAN platform, built on its AI-native anyRAN software and NVIDIA’s Aerial platform, with three hardware deployment paths, a value-based software subscription model, and a roadmap targeting more than 100 percent spectral-efficiency gains by 2028; pilot deployments are scheduled for the end of 2026 and commercial availability for 2027, with T-Mobile US named at every milestone.[40] Nokia’s chief technology and AI officer, Pallavi Mahajan, has framed the strategy as a bet on convergence itself:
“As the boundaries between network and compute continue to blur” — Pallavi Mahajan, Chief Technology and AI Officer, Nokia [59]
distributed infrastructure, she argued, can become a powerful platform for both connectivity and AI.
Ericsson has taken the opposite path. Its AI in RAN software, launched June 11, 2026, embeds telco-grade AI models directly into basebands and radios running on Ericsson Silicon, including neural-network accelerators that the company says increased AI inference capability in its Massive MIMO radios tenfold, and it is delivered as a software subscription that activates on existing AI-ready hardware without a capital-intensive hardware replacement and explicitly without GPUs.[16] Ericsson’s models are designed for microsecond-scale inference at the radio, and the company has emphasized energy efficiency as a central advantage of purpose-built silicon over general-purpose accelerators. Ericsson has nonetheless also demonstrated its Cloud RAN software running on NVIDIA infrastructure, and T-Mobile reported that tests in New York of Ericsson’s silicon-based AI beamforming improved downlink user throughput by 16 percent and spectrum efficiency by 12 percent during peak congestion, announced the day before the Bellevue call.[3] Ericsson’s own executives have described a future in which AI models run across devices, cell sites, the edge, and centralized systems, with model size and placement determined by latency, power, and performance requirements rather than by a single architectural doctrine.[60]
The distinction matters because it is a genuine disagreement about engineering economics rather than a marketing difference. Some network improvements may be achieved most efficiently with specialized processors inside radio equipment, where power budgets are tight and latency requirements are measured in microseconds. Others may benefit from more general-purpose accelerated computing that can run larger and more complex algorithms and that can be repurposed for external workloads when radio demand is low. Still others may require regional computing systems. There is unlikely to be one universally optimal arrangement, and Futurum’s analysis frames the question precisely as whether custom silicon or merchant silicon wins the next RAN architecture cycle.[40] For the Tower Inference thesis specifically, the Ericsson path delivers AI-for-RAN value with minimal new capital but offers no shared platform for AI-on-RAN services, while the Nokia path requires new capital and bets that the shared platform will be worth it. Competition between the two could therefore extend beyond traditional radio-network performance into the design of AI-native computing architectures, and operators who hedge, as T-Mobile has done by testing both, may be making the wisest choice available.
4.4. Cisco and the Emerging Market for Intelligent Inference Routing
Cisco introduces another dimension that is easy to overlook but may prove decisive: the question of deciding where an inference request should execute. In October 2026, Cisco published an AI Inference Routing blueprint developed with NVIDIA, built on mechanisms that service providers already operate, including Segment Routing over IPv6, Flexible Algorithm, and BGP, and designed to incorporate workload requirements and policies into network-routing decisions so that providers can steer each inference request toward an eligible location along a path that reflects its latency, throughput, cost, and jurisdictional requirements.[12] The accompanying white paper treats every inference location as a datacenter, defines four placement tiers, and addresses slicing, assurance, security, and automation for inference connectivity across access, metro, core, and edge domains.[21]
This addresses a fundamental challenge in distributed computing that the hardware announcements tend to skip past. Simply installing AI accelerators at multiple network locations does not guarantee that applications will use those resources effectively, because an inference request may have suitable capacity nearby and still travel farther than necessary if the network lacks context about the workload. A system must determine which computing site is suitable for a particular task, and for any given request the choices may include a nearby facility with limited available capacity, a larger regional center with more processing resources, a remote hyperscale datacenter offering lower computing costs, or an enterprise-owned system operating within a regulated environment. The optimal destination changes according to network conditions, inference requirements, power availability, model placement, and commercial agreements, and it changes from moment to moment.
This creates a possible new infrastructure capability, intelligent workload placement across geographically distributed computing resources, and it also introduces a form of economic competition that the operators have not yet fully reckoned with. A company controlling inference orchestration could influence which computing providers receive workloads, which means that commercial value may not belong exclusively to the owner of the GPUs or the operator of the radio network. It may also accrue to the software platform that coordinates the execution of AI services, in the same way that value in the application economy accrued to the platforms that coordinated distribution rather than to the networks that carried the bits. The SoftBank AITRAS Orchestrator, Nokia’s MantaRay, Cisco’s Crosswork AI, and the hyperscalers’ own orchestration layers are all contenders for this role, and the operator that cedes orchestration to a third party may find that it has ceded pricing power along with it.
4.5. The Role of Hyperscalers: Collaboration and Competition
The expansion of Tower Inference does not necessarily threaten the hyperscale model, and the two infrastructures may become increasingly interdependent rather than adversarial. AWS, Microsoft, and Google already possess extensive experience managing cloud applications, distributed software, identity systems, model services, and enterprise workloads, and telecommunications operators can contribute local facilities, mobile-network integration, and geographic reach that the hyperscalers lack. AWS Wavelength, which embeds AWS compute and storage within telecommunications providers’ facilities at the edge of 5G networks for latency-sensitive applications, demonstrates an established model for this kind of cooperation.[61] T-Mobile’s Kinetic Edge initiative adds a new chapter: the company is partnering with AWS to build the first instances of its distributed intelligence infrastructure for physical AI, combining T-Mobile’s nationwide 5G Advanced network and distributed edge with AWS’s cloud orchestration and AI services to bring what Saw calls deterministic connectivity, compute, and intelligence closer to where machines operate.[9] Google Cloud is pursuing related opportunities through Google Distributed Cloud, which in April 2026 added Gemini models, a new A4 machine family with 2.25 times the peak compute of its predecessor, and a sovereign agentic AI architecture for on-premises, edge, and air-gapped environments, in partnership with NVIDIA and Dell.[62]
These developments suggest three possible commercial relationships, and the operator’s returns differ dramatically among them. In a partnership model, telecommunications operators provide facilities and connectivity while hyperscalers supply computing platforms and software services, and the operator earns a hosting and connectivity margin. In a competition model, operators develop independent inference businesses that compete with certain cloud offerings, and the operator earns platform margins but bears platform risks. In a hybrid integration model, operators and cloud providers jointly deliver solutions combining regional computing, mobile-network functions, and enterprise applications, and the margin is split according to each party’s bargaining power. The third model may become particularly important because each participant contributes capabilities that would be expensive to reproduce independently, and T-Mobile’s Kinetic Edge is explicitly structured this way.
Nevertheless, the allocation of economic returns in the hybrid model remains an open question, and the history of operator–hyperscaler partnerships is not encouraging for the operators. If the hyperscaler controls application development, model distribution, customer billing, and workload orchestration, the telecommunications operator may receive only a limited portion of the total service value, essentially a real-estate and connectivity fee. If the operator owns an indispensable network capability, such as deterministic quality of service or mobility management, and establishes direct enterprise relationships, its bargaining position may improve substantially. Tower Inference therefore requires careful examination of commercial contracts, customer ownership, software control, and pricing authority, and the operator that enters these partnerships without clarity on all four will discover, as operators discovered in the app-store era, that the partner’s business model has quietly become its own.
4.6. The Strategic Importance of Software Portability
A distributed inference market will become more economically sustainable if applications and radio software can operate across different computing architectures, because portability reduces dependence on a single vendor and permits operators to upgrade computing equipment without completely rebuilding their software environments. However, portability is complicated in practice. Different AI accelerators have distinct memory architectures, software frameworks, development tools, and performance characteristics, and applications optimized for one platform may not perform equally well on another. Radio-network software introduces additional constraints, because a shared platform must protect network operations while supporting external workloads, and the isolation mechanisms that provide that protection are themselves platform-specific.
These requirements make interoperability a commercial issue rather than merely a technical preference, and the SoftBank–Nokia field trial is relevant precisely because it tests portability under live conditions. The trial will validate Nokia’s anyRAN software running as the RAN component of SoftBank’s AITRAS platform, which was designed to host SoftBank’s own RAN software, third-party RAN software, and edge AI applications on common GPU infrastructure, with the explicit goal of assessing AITRAS as a common platform for hosting software from multiple vendors without the limitations imposed by any specific RAN implementation.[8][63] The Polese–Melodia architecture pursues the same goal from the standards side, extending O-RAN’s modular, disaggregated, cloud-native principles to AI workloads so that AI-RAN sites can be managed through open interfaces with multi-vendor interoperability.[18] The long-term economic benefit would be greater operator flexibility in choosing technology suppliers and a more competitive market for the accelerators themselves. Yet the success of such arrangements will depend on actual commercial performance under commercial traffic, not on architectural promises, and eWeek’s assessment of the SoftBank trial is the right one: it is a milestone to watch rather than a confirmed production rollout, and evidence of sustained reliability and economic value will determine how compelling AI-RAN becomes.[64]
4.7. The International Competition for Distributed Intelligence Infrastructure
Tower Inference is not exclusively an American development, and the international dimension of the contest is in some respects more consequential than the vendor rivalry. Europe, Japan, and other regions are actively exploring AI-native network architectures, and they are doing so with different priorities. Deutsche Telekom is advancing AI-RAN field trials for 2027 alongside its Industrial AI Cloud; Elisa has conducted shared-workload field trials and plans commercial validation in 2027; BT is preparing live trials in Birmingham; Orange is developing distributed voice-AI services and testing Nokia software on NVIDIA platforms for spectral efficiency and edge services in live networks; and SoftBank is pursuing its AITRAS architecture in Japan with a 2027 commercial-network trial.[1][8] NVIDIA counted nearly twenty operators globally with public AI-RAN collaborations, including ten in Europe, the Middle East, and Africa, and nine EMEA operators building or expanding NVIDIA-powered infrastructure across centralized and distributed locations.[1]
These projects represent different potential pathways toward commercial deployment. European operators place considerable emphasis on data localization, sovereignty, enterprise applications, and regulated industries, and Nokia’s launch of a sovereign agentic AI platform for telco autonomy, which allows operators to automate network operations while retaining control over operational data and AI decisions, speaks directly to that priority.[55] Japanese operators are exploring computing architectures relevant to robotics, industrial systems, and advanced mobile networks. American operators combine nationwide mobile infrastructure with major hyperscale partnerships and the most aggressive physical-AI pilots. China’s telecommunications industry warrants separate examination because its advanced mobile networks, industrial AI initiatives, domestic equipment suppliers, and domestic accelerator ecosystem could support alternative distributed-computing architectures that are not dependent on NVIDIA at all, and the Stanford AI Index 2026 reports that the performance gap between American and Chinese models has nearly vanished and that China leads in industrial robotics deployment, which is precisely the application domain in which Tower Inference is expected to matter most.[11]
The Center for Strategic and International Studies has argued that the AI technological revolution will be inseparable from the evolution of wireless technology to 6G, and that because wireless networks will increasingly run AI workloads on the same infrastructure they use to carry data, control over communications will affect which nations and companies control computing and ultimately the entire AI stack.[65] The White House has adopted a similar view, describing 6G in its December 2025 memorandum as
“foundational to the national security, foreign policy, and economic prosperity of the United States.” — The White House, National Security Presidential Memorandum NSPM-8, “Winning the 6G Race” [66]
International comparisons must nonetheless distinguish confirmed deployments from proposed services and announced trials, and a reported demonstration does not indicate a commercially scaled platform. The geopolitical significance lies partly in the possibility that different countries will adopt different combinations of semiconductors, network equipment, software, models, and governance, and that such divergence could complicate interoperability and create strategic dependencies throughout the Five-Layer AI Economy. The question extends beyond who manufactures the most advanced radio equipment. It concerns who controls the infrastructure through which intelligent machines communicate, execute, and coordinate. The fourth major proposition of this paper is that Tower Inference could produce a new arena of infrastructure competition in which network ownership, semiconductor performance, cloud orchestration, software portability, and enterprise relationships jointly determine economic advantage, and in which no single participant, not the operator, not NVIDIA, not the hyperscaler, holds all of the cards.
Section 5: The 2027–2030 Outlook — Commercial Adoption, Investment Geography, Energy, and Government Policy
The preceding four sections have examined Tower Inference as a technology, an economic proposition, a set of applications, and a competitive arena. This section turns to the future, and specifically to the period between 2027 and 2030 during which the thesis will either be validated by commercial evidence or relegated to the long list of telecommunications visions that outran their business cases. It proposes a four-year development scenario, examines where in the United States distributed inference investment is most likely to concentrate, analyzes the energy requirements that distributed computing imposes, situates the thesis within federal and state policy, addresses cybersecurity and operational governance, and closes with the most important unanswered question of all: where the commercial boundary of distributed inference actually lies.
5.1. The Transition From Experiments to Commercial Deployments
The period between 2027 and 2030 may determine whether Tower Inference becomes a significant infrastructure category or remains a specialized collection of network innovations. As of October 31, 2026, the industry has produced important technical demonstrations: the Bellevue AI-RAN call demonstrated software-defined radio processing on accelerated computing in a live network; Elisa demonstrated three concurrent workload types on shared hardware over the air; Deutsche Telekom measured double-digit uplink gains in the laboratory; SoftBank demonstrated external-workload orchestration and is preparing a commercial-traffic trial; and physical-AI developers are exploring edge-based applications with the largest American operator. These developments create an expanding body of technical evidence. They do not yet establish that distributed inference will become a major contributor to telecommunications revenue, and the earnings analysis in Section 2.5 found no such contribution anywhere in the third-quarter 2026 results.
The next phase must produce commercially meaningful evidence of four kinds. Operators will need to demonstrate that shared infrastructure can support demanding network operations and external AI workloads simultaneously under commercial traffic, which is what the 2027 trials at SoftBank, Deutsche Telekom, BT, and Elisa are designed to test. They must show that the system performs reliably during heavy traffic, equipment failures, and changing demand, which requires months of operation rather than a single demonstration. They must establish that customers will actually pay for the service, which requires signed enterprise contracts rather than letters of intent. And they must demonstrate that the incremental investment produces acceptable financial returns under the framework proposed in Section 2.6. The transition from proof of concept to commercial deployment is therefore likely to be selective and uneven, with certain metropolitan areas and industrial applications becoming commercially attractive earlier than others, and Nokia’s own stated timeline, pilots at the end of 2026, first commercial release in 2027, and mass deployment in 2028, should be read as the optimistic case rather than the base case.[67][40]
5.2. A Proposed Four-Year Development Roadmap
The following is an analytical scenario, not a prediction that the listed milestones will necessarily occur, and its most important feature is that commercialization should precede extensive infrastructure expansion at every stage.
| Year | Phase | What must be demonstrated | Evidence that would validate the phase |
| 2027 | Commercial validation | Workload isolation, technical reliability, radio performance, and initial customer applications in live networks with commercial traffic | Published results from SoftBank (Japan), Deutsche Telekom, BT (Birmingham), Elisa, and T-Mobile field trials; first commercial AI-RAN software releases from Nokia; Ericsson AI in RAN at scale |
| 2028 | Selective enterprise expansion | Regional inference services for targeted industrial, robotics, voice, and infrastructure applications, priced and contracted | Recurring enterprise revenue disclosed in operator earnings; measured spectral-efficiency gains in commercial networks; Qualcomm-led 6G pre-commercial demonstrations |
| 2029 | Multi-site orchestration | Workload placement across far-edge, metro, regional, and hyperscale resources with policy-aware routing | Production deployment of inference-routing systems; multi-vendor portability demonstrated; first 6G commercial infrastructure and devices |
| 2030 | Infrastructure specialization | Differentiation between high-capacity regional edge facilities, specialized low-latency sites, and conventional connectivity sites, guided by demand | Capital allocation visibly tied to measured utilization; 3GPP Release 21 6G specifications; initial commercial 6G systems |
Table 3. Tower Inference development scenario, 2027–2030. Analytical framework proposed for this paper; the timing of 6G milestones follows 3GPP’s stated plan to enable initial commercial 6G systems by around 2030 and Qualcomm’s stated 2029 commercialization target.[68][57]
An operator should not assume that a successful demonstration justifies deploying accelerated computing across thousands of sites. Instead, each phase should establish evidence sufficient to support the next, and the operator’s investment committee should treat the roadmap as a series of options to be exercised only when the preceding option has paid off.
5.3. The Geography of Investment: Why Some Regions May Move First
Distributed inference infrastructure will not develop uniformly across the United States. Its location will depend on the interaction of enterprise demand, network coverage, computing costs, and available electrical infrastructure, and several regional opportunities deserve attention as hypotheses for further investigation rather than as claims that investment commitments have been made.
California, and specifically the San Francisco Bay Area and Southern California, contains substantial technology, robotics, logistics, and AI-development activity, and T-Mobile’s pilots with Figure AI, Serve Robotics, and the City of San Jose are all Californian. The state also presents important questions concerning energy prices, permitting, environmental requirements, and the regulatory treatment of autonomous systems that could either accelerate or impede deployment. Texas combines major metropolitan markets, industrial facilities, logistics operations, semiconductor development, and extensive energy infrastructure that could support targeted distributed-computing applications, although the state’s electricity system is already absorbing extraordinary data-center demand: ERCOT was tracking roughly 475 gigawatts of large-load interconnection requests as of June 2026, about 90 percent of them associated with data centers.[69] Arizona’s semiconductor manufacturing investments and advanced-manufacturing activity may create specialized demand for private-network and edge-AI services. Virginia’s existing datacenter and telecommunications infrastructure could support regional computing opportunities, although power availability, community opposition, and competing infrastructure demand are severe constraints. Ohio and Indiana’s manufacturing, logistics, and distribution facilities could produce demand for enterprise applications that combine reliable communications with localized processing.
These regional observations suggest that Tower Inference should be examined alongside existing datacenter development rather than treated as a completely separate infrastructure economy. A metropolitan inference facility may compete for electricity, equipment, and technical labor with other computing projects in the same market, and in the most constrained markets the telecommunications operator’s facilities may be valuable precisely because they already have interconnection agreements that a new campus would wait years to obtain. Conversely, a distributed architecture could make certain applications practical without requiring all processing to occur in a distant hyperscale campus. The net effect on total energy consumption remains uncertain, because distributing computing capacity can reduce some data-transfer requirements while increasing the overhead associated with operating many smaller facilities at lower utilization.
5.4. Energy Infrastructure: The Hidden Requirement of Distributed Computing
The Five-Layer AI Economy places energy at the foundation of the entire system, and Tower Inference reinforces rather than relaxes that relationship. Advanced computing equipment requires electricity, cooling, networking, and operational support, and these requirements become more complicated, not less, when infrastructure is distributed across many locations. A hyperscale datacenter is designed from the ground up to concentrate substantial computing resources within an integrated electrical and cooling environment. A telecommunications facility was designed to support communications equipment with very different power densities and thermal profiles, and introducing accelerated computing can require electrical upgrades, additional cooling, equipment replacement, and revised maintenance arrangements at each site.
Operators must therefore assess whether an existing site can economically accommodate new computational workloads, and the answer will vary across the hierarchy in Table 1. For a regional computing facility, the costs may be manageable; at a constrained rooftop installation, the economics may be prohibitive. The Department of Energy’s draft 2026 National Transmission Needs Study, released July 9, 2026 for public comment, found what it described as a
“pressing need for additional electric transmission infrastructure” — U.S. Department of Energy, Office of Electricity, 2026 Draft National Transmission Needs Study [70]
driven by load growth from data centers, expanding domestic manufacturing, large industrial loads, and electrification, and it identified AI data centers as the dominant driver of electricity demand growth, with data centers potentially representing up to 12 percent of national electricity consumption by 2028.[70][69] Grid Strategies estimates that data centers account for roughly 55 percent of forecast peak load growth between 2025 and 2030, about 90 gigawatts of a 166-gigawatt total.[71]
Tower Inference introduces a related but geographically different planning problem. Instead of examining only the electricity requirements of enormous AI campuses, utilities and operators may eventually need to evaluate incremental computing demand at selected distributed network locations, and the cumulative demand might be commercially significant even when individual facilities are relatively small. However, it would be premature to assume that this market will create electricity requirements comparable to hyperscale AI training infrastructure, and the scale of future adoption remains unproven. The IEA has noted that power consumption per AI task is declining at an unprecedented rate even as total consumption rises, which cuts in favor of distributed inference becoming more affordable over time.[33] A useful policy objective would be to encourage operators to disclose the location, energy requirements, and intended utilization of substantial new computing deployments when those projects affect local infrastructure planning, so that utilities can plan for a distributed load rather than discovering it after the fact.
5.5. Federal Policy: 6G, Spectrum, and National Technological Leadership
The relationship between artificial intelligence and mobile-network infrastructure has become a matter of federal policy. On December 19, 2025, the White House issued National Security Presidential Memorandum 8, “Winning the 6G Race,” which declared it the policy of the United States to lead the world in 6G development, identified advanced mobile networks as pivotal to artificial intelligence, robotics, and other emerging technologies, directed immediate planning to relocate federal systems from the 7.125–7.4 GHz band so that it can be cleared for full-power commercial 6G use, gave federal incumbents twelve months to submit relocation plans, ordered studies of additional bands, and directed diplomatic engagement to advance American leadership in 6G standards and deployment.[66][72] In July 2026, the National Telecommunications and Information Administration followed with a “Call to Action for 6G Leadership and Security” that enlisted more than twenty allied countries.[73]
The policy significance extends beyond faster wireless connectivity. If future mobile networks support distributed computing and intelligent applications, communications infrastructure becomes more closely connected to AI capabilities across the economy, and the policy domains that govern it begin to interact. Spectrum policy influences wireless capacity and reliability, and therefore the uplink headroom on which physical AI depends. Semiconductor policy affects the availability of advanced computing equipment at the edge. Cybersecurity policy affects the trustworthiness of distributed infrastructure. International standards influence interoperability. Energy policy affects the feasibility of deploying and operating computing facilities. The White House’s July 2025 AI Action Plan emphasizes investment in American AI infrastructure, energy systems, semiconductors, and the workforce required to build and operate advanced facilities, and Tower Inference belongs squarely within that broader policy environment.[74]
Not every analyst agrees with the memorandum’s emphasis. Dean Bubley of Disruptive Analysis has argued that winning in 6G will not require more spectrum so much as more intelligent use of existing spectrum through AI efficiency and spectrum sharing, a critique that, if correct, makes the AI-for-RAN results discussed in Section 2.3 more strategically important than the spectrum clearances the memorandum prioritizes.[75] Either way, the existence of federal AI and 6G initiatives should not be interpreted as a dedicated federal program to finance GPU-equipped radio towers. That would require separate policy authorization and appropriation, and nothing in the current policy framework provides it.
5.6. State Governments and Local Infrastructure Governance
Governors and state governments could become important participants in the evolution of distributed AI infrastructure, and their role would differ somewhat from their involvement in large hyperscale datacenter projects. Many telecommunications facilities are already established, but significant modifications may still raise questions involving land use, electrical capacity, building regulations, equipment noise, and emergency power, and as advanced computing becomes more widely distributed, local authorities may encounter categories of infrastructure proposals they have not previously evaluated. An existing telecommunications building might require additional cooling equipment that changes its noise profile. A metropolitan aggregation facility might need upgraded electrical service that affects a neighborhood’s distribution network. An industrial customer might seek to combine private wireless infrastructure with dedicated inference computing on a site zoned for neither.
These situations involve overlapping authority among telecommunications regulators, utilities, local governments, and sometimes federal agencies, and the appropriate policy response should depend on the actual nature of the installation. A computing upgrade within an existing building has different consequences from the construction of a major new facility and should not be subjected to the same process. State policymakers should also avoid assuming that distributed edge computing necessarily reduces total infrastructure costs; it may simply relocate portions of computing expenditure from one jurisdiction to another. The economic objective should be to encourage productive investment while maintaining transparency concerning electricity demand, reliability, community effects, and public infrastructure obligations.
5.7. Cybersecurity, Operational Responsibility, and the Governance of Network Intelligence
A telecommunications network that hosts AI services creates new security considerations that compound the traditional ones. Conventional network-security concerns include unauthorized access, service disruption, equipment compromise, and threats to communications integrity. Distributed inference introduces additional risks involving model execution, customer data, software supply chains, workload isolation, and automated decisions, and it introduces them at hundreds or thousands of locations that have historically been physically secured but not operated as multi-tenant computing environments.
If AI applications share infrastructure with critical network functions, security failures in the AI tenant must not be allowed to compromise essential communications services, which means that operators will need strong separation between external workloads and protected network operations, clear permissions, incident-response procedures, continuous monitoring, and accountability mechanisms. The emergence of agentic network management raises the stakes further. Nokia’s sovereign agentic AI platform and Cisco’s Crosswork AI both integrate NVIDIA’s Nemotron models to allow AI agents to perform tasks such as deep network troubleshooting across a distributed footprint, and as these systems become more capable, the boundary between recommending an action and executing it becomes increasingly consequential for the reliability of national infrastructure.[55][12]
NIST’s AI Risk Management Framework offers a useful foundation for examining trustworthy AI deployments, and on April 7, 2026, NIST released a concept note for an AI RMF Profile on Trustworthy AI in Critical Infrastructure, intended to
“guide critical infrastructure operators towards specific risk management practices” — National Institute of Standards and Technology, AI RMF Profile on Trustworthy AI in Critical Infrastructure [76]
when engaging AI-enabled capabilities across information technology, operational technology, and industrial control systems.[76][77] The concept note is developmental rather than a finalized regulatory regime, but communications networks are among the critical infrastructure sectors it addresses, and operators deploying Tower Inference would be well advised to align with it before it becomes auditable. For Tower Inference, a central governance principle should be that network operational safety and essential communications reliability take precedence over opportunistic commercial inference workloads, always and without exception, and that this precedence be enforced in the orchestration layer rather than merely promised in the contract.
5.8. The Most Important Unanswered Question: Where Is the Commercial Boundary?
The long-term success of Tower Inference depends on identifying the point at which distributed computing provides enough additional value to justify its costs, and this boundary will differ by workload in ways that resist generalization. A conventional enterprise chatbot may operate economically from a regional or hyperscale datacenter, because its users tolerate hundreds of milliseconds of latency without noticing. A vision system processing sensitive industrial data might benefit from local computing for reasons of sovereignty and bandwidth rather than latency. A mobile robot may require both onboard intelligence and geographically distributed assistance. An autonomous vehicle may use the network principally for supplemental information and fleet services. An emergency-call translation service may require edge execution for latency but hyperscale execution for the long tail of rare languages.
These differences imply that edge computing will not replace centralized cloud computing, and the Stanford AI Index’s documentation of the continuing collapse in inference costs means that the economic threshold for distribution keeps moving.[11] Instead, the market may evolve toward increasingly sophisticated methods for allocating workloads among devices, local facilities, telecommunications infrastructure, and hyperscale datacenters, which is exactly the capability that Cisco’s routing blueprint and SoftBank’s orchestrator are attempting to build. For each application, the appropriate computing location will depend on required response time and reliability, model size and computational intensity, data volume and transmission cost, energy efficiency, security and geographic restrictions, availability of local infrastructure, total lifecycle cost, and commercial responsibility for service performance. These criteria provide a basis for evaluating the 2027–2030 market without relying on speculative claims about the number of future GPU-equipped towers, and they lead to the fifth major proposition of this paper: Tower Inference will succeed through workload-specific investment and measurable commercial performance, not through the widespread installation of AI accelerators across telecommunications networks. The fundamental economic principle is that computing should be distributed only when its location creates additional value, and the discipline to honor that principle will separate the operators who profit from this transition from those who merely finance it.
Section 6: What Have We Learned? Eight Pillars of Tower Inference
The developments announced during October 2026 suggest that artificial intelligence may fundamentally change the economic role of telecommunications networks, yet the transition remains in its early stages, and the evidence supports a serious investigation of distributed intelligence infrastructure without justifying the assumption that every radio site will become a miniature AI datacenter or that telecommunications operators will automatically capture a large share of AI industry revenue. Eight lessons emerge from the analysis, and together they constitute the framework through which this research series will track Tower Inference through 2030.
Pillar 1. Telecommunications Infrastructure Is Becoming a Potential Computing Asset
The first lesson is that telecommunications networks may become economically valuable for more than transporting information. Existing network facilities possess geographic reach, established connectivity, operational infrastructure, and relationships with enterprises and consumers, and adding suitable accelerated computing systems could allow selected locations to support AI model execution alongside traditional network operations. The value of this transformation depends on whether distributed processing provides measurable advantages, and the most promising applications will be those requiring localized data processing, geographically sensitive coordination, or carefully managed network performance. Infrastructure location thereby becomes an additional dimension of competition within the Five-Layer AI Economy: the relevant question is no longer exclusively how many GPUs an organization owns but whether those GPUs operate in locations appropriate for the applications they serve. For policymakers and investors, this introduces a distinction between computing capacity and economically useful computing capacity that the gigawatt-centric discourse of the past three years has largely ignored.
Pillar 2. Shared Computing Must Produce Verifiable Financial Benefits
The second lesson concerns AI-RAN economics. Combining radio-network workloads with external inference services creates the possibility of higher equipment utilization and additional revenue, but technical feasibility does not guarantee financial success. Operators must evaluate the cost of computing hardware, power, cooling, orchestration, upgrades, maintenance, security, and the reserve capacity needed to protect radio functions, and they must do so under the investment framework proposed in Section 2.6 rather than under vendor market-size forecasts. The third-quarter 2026 earnings of T-Mobile, Verizon, AT&T, Nokia, and Ericsson show strong core businesses and no AI-RAN revenue, and the AI revenue that has appeared, Verizon’s $2 billion AWS agreement, is flowing to long-haul fiber rather than to the radio edge. The objective is not maximum GPU deployment; it is maximum economically productive intelligence delivery, and infrastructure should expand according to verified demand and sustainable unit economics.
Pillar 3. Physical AI Will Require Intelligence Distributed Across Devices, Networks, and Datacenters
The third lesson is that physical artificial intelligence may create genuinely new demand for network-integrated computing, because robots, industrial machines, autonomous systems, and intelligent infrastructure require different combinations of local and remote processing than chatbots ever did. Some functions must remain onboard; others may operate effectively at nearby facilities; larger workloads may continue to benefit from centralized infrastructure. A useful model is not a binary choice between device and cloud but a computing continuum in which workloads are assigned according to performance, cost, reliability, safety, and operational requirements, and the T-Mobile–Figure investigation of which portions of the Helix models can run at the edge is the most important live experiment in drawing that continuum. For enterprises, the practical lesson is to design intelligent systems around application requirements rather than around the availability of a particular computing platform; for operators, it is to develop infrastructure that can accommodate changing workloads without compromising essential network services.
Pillar 4. Control of Distributed Intelligence Will Be Shared, and Contested
The fourth lesson concerns competition. Operators own network assets; NVIDIA and other semiconductor companies provide accelerated computing; Nokia and Ericsson develop radio platforms on opposite silicon strategies; Cisco and others supply routing and orchestration; AWS, Google Cloud, and Microsoft contribute cloud platforms and enterprise ecosystems; and AI developers supply models and applications. No single participant controls every part of the emerging system, which creates possibilities for both cooperation and competition, and the availability of interoperable hardware, software, and network interfaces will influence how much bargaining power each participant retains. The central strategic question is therefore not who owns the radio tower but who controls the orchestration layer, the customer relationship, and the pricing authority that determine how intelligence is delivered and how its value is divided.
Pillar 5. The 2027–2030 Market Will Be Determined by Deployment Economics, Reliability, and Governance
The fifth lesson concerns the transition from experimentation to commercial infrastructure. Early AI-RAN demonstrations provide encouraging technical evidence, but the next phase requires proof that these systems can operate reliably, securely, and profitably under commercial traffic, and the most important evidence will come from the 2027 field trials, from recurring enterprise contracts, from actual equipment utilization, and from transparent measures of technical performance. Government policy will influence the market through spectrum management, telecommunications security, energy infrastructure, and industrial investment; 3GPP’s Release 20 studies on 6G architecture, AI-native radio, and AI/ML for NG-RAN are steps in standards development rather than proof of commercial 6G; and NIST’s critical-infrastructure profile will shape how AI in networks is governed.[78][76] The strongest investment opportunities may emerge from selective facilities serving identifiable commercial demand, and the ultimate opportunity is not to eliminate centralization but to create a more capable and economically efficient system combining centralized and distributed intelligence.
Pillar 6. The Uplink, Not the Accelerator, May Be the Binding Constraint
The sixth lesson, which the October 2026 evidence brought into unexpectedly sharp focus, is that the scarce resource in Tower Inference may turn out to be radio uplink capacity rather than edge computing. Every physical-AI application examined in this paper is a source of data rather than a sink for content, and Ericsson’s projection of a threefold increase in uplink volumes from AI traffic alone by 2031, T-Mobile’s identification of the busy-sector uplink as the single thing to watch, and Deutsche Telekom’s decision to aim its first AI-for-RAN effort specifically at uplink channel estimation all point in the same direction.[28][31][5] The latency equation in Section 1.4 shows why: an edge server that eliminates transport delay accomplishes nothing if the device cannot get its sensor data into the network in the first place. The operators best positioned for Tower Inference may therefore be those with the deepest mid-band spectrum portfolios and the most aggressive uplink-enhancement programs, not merely those with the most GPUs, and the skeptical counterargument that uplink growth will fill spare capacity without new investment deserves to be tested empirically rather than dismissed.
Pillar 7. Measurement Discipline and Disclosure Will Separate Evidence From Promotion
The seventh lesson is methodological. The Bellevue call disclosed no throughput, latency, or efficiency data; Deutsche Telekom’s gains are laboratory measurements awaiting field validation; Nokia’s 100 percent spectral-efficiency target is a 2028 roadmap figure of which roughly 20 percent has been demonstrated; Ericsson’s gains come from a different architecture under different conditions; and the three forms of AI-RAN are routinely conflated in both promotional and critical commentary.[3][5][40] A research program on Tower Inference must therefore insist on distinguishing technological feasibility from network efficiency, network efficiency from commercial deployment, and commercial deployment from profitability, and it must demand that claims be accompanied by the conditions under which they were measured. The companies that publish tail-latency distributions, utilization figures, and contracted revenue will earn the credibility that the sector needs; those that publish only firsts will not.
Pillar 8. Distributed Intelligence Raises Global Questions of Access and Inclusion
The eighth lesson extends beyond the operators and vendors to the global economy in which they operate. The IMF’s Kristalina Georgieva has described artificial intelligence as arriving
“like a tsunami hitting the labor market” — Kristalina Georgieva, Managing Director, International Monetary Fund [79]
with roughly 40 percent of jobs globally and 60 percent in advanced economies exposed to its effects, and the physical-AI applications that Tower Inference enables, in logistics, manufacturing, and delivery, are among the most labor-relevant of all. UNCTAD’s Technology and Innovation Report 2025 warned that access to AI compute and infrastructure remains concentrated in a few economies and that just 100 companies, mostly in the United States and China, account for 40 percent of global private AI research and development investment, with its Secretary-General Rebeca Grynspan observing that 118 countries,
“primarily in the Global South, are absent from key debates” — Rebeca Grynspan, Secretary-General, UNCTAD [80]
on AI governance.[81] The World Bank’s World Development Report 2026 argues that developing economies have more to gain than to fear from AI if they invest in reliable electricity, connectivity, and skills, and its president Ajay Banga, launching the Bank’s first digital strategy in October 2026, described the required foundations as running
“from reliable connectivity and affordable devices to digital skills and useful services.” — Ajay Banga, President, World Bank Group [82]
The World Bank’s report director framed the ambivalence of the moment with precision:
“AI generates unprecedented excitement. AI also generates unprecedented concern.” — Gaurav Nayyar, Director, World Development Report 2026, World Bank Group [83]
Tower Inference is relevant to this global picture in two opposite ways. On one hand, mobile networks are the most widely distributed digital infrastructure in the developing world, and if they can host inference, they could bring AI capability to places that will never host a hyperscale campus, which is precisely the kind of shared compute facility that UNCTAD has called for.[81][84] On the other hand, Tower Inference as currently constituted depends on accelerators, software, and orchestration controlled by a handful of companies in two countries, and its diffusion could deepen rather than narrow the concentration that UNCTAD warns against. Which of these outcomes prevails is a policy choice, not a technological inevitability, and it belongs on the research agenda alongside the questions of latency and unit economics.
The overarching lesson across all eight pillars is that Tower Inference represents the possibility that telecommunications networks will evolve from supporting the movement of information to participating directly in the production, delivery, and coordination of artificial intelligence, and that its commercial importance will depend on how successfully the industry converts that possibility into reliable services with measurable economic value, honestly reported and equitably distributed.
Conclusion: Tower Inference and the Next Geography of Artificial Intelligence
The October 21, 2026 announcements involving NVIDIA, Nokia, T-Mobile, and the European operators marked an important moment in the evolution of mobile communications. The successful AI-RAN call in Bellevue, Washington, demonstrated that accelerated computing and software-defined radio technology could operate together in a commercial network environment without degrading the service that subscribers depend on. Related developments in Germany, Finland, the United Kingdom, France, and Japan, together with T-Mobile’s disclosures about humanoid robots, delivery fleets, and city infrastructure, suggested that the industry was beginning to examine telecommunications networks as more than systems for transporting information. Yet the importance of these developments should not be measured solely by the technological achievement of completing a call, and the companies involved were themselves careful not to measure it that way. Their deeper significance lies in the possibility of restructuring where artificial intelligence is executed and how its economic value is distributed.
For much of the modern AI industry, the central infrastructure narrative has focused on enormous datacenters, advanced GPU clusters, expanding electricity requirements, and the extraordinary capital expenditure of hyperscale technology companies, and NVIDIA’s $89 billion of quarterly datacenter revenue confirms that this narrative remains foundational. Large-scale model training and many forms of inference will continue to benefit from highly concentrated computing resources, and nothing in this paper suggests otherwise. However, a future economy involving millions of intelligent machines may require a broader infrastructure model than the hyperscale campus alone can supply. Autonomous robots, industrial systems, connected vehicles, enterprise applications, emergency services, and intelligent public infrastructure will generate computational requirements that differ in kind from those of a chatbot: some tasks demand immediate local processing, others benefit from nearby shared intelligence, and still others require the considerable capacity of regional or hyperscale systems. The next infrastructure challenge may therefore involve coordinating intelligence across locations rather than simply concentrating additional computing power within ever-larger facilities, and NYU’s Ted Rappaport, who has spent four decades anticipating what each wireless generation would make possible, has described the capabilities that this coordination could unlock as
“things that right now seem like science fiction.” — Theodore S. Rappaport, Founding Director, NYU WIRELESS [85]
Telecommunications networks could become essential participants in that coordination. Their geographic reach provides an existing foundation for connecting devices with distributed computing resources. Their radio systems offer operational capabilities, mobility management, quality of service, and now sensing, that no other infrastructure possesses. Their facilities may accommodate selected inference workloads at a fraction of the cost of new construction. Their enterprise relationships and billing systems could support new commercial services. The academic community anticipated this role years before the vendors commercialized it: the Nokia Bell Labs vision of an AI-native air interface, Letaief’s edge AI for 6G, Poor and Saad’s edge-native intelligence, Tong and Li’s Network-for-AI, Andrews’s deep learning in the 6G air interface, and Melodia’s converged O-RAN and AI-RAN architecture together form an intellectual lineage that gives Tower Inference a firmer foundation than most infrastructure visions enjoy.[86][22][23][24][87][18]
Nevertheless, these opportunities are accompanied by substantial uncertainty, and this paper has tried to give that uncertainty its full weight. The economics of shared computing must be demonstrated under commercial traffic. Hardware and software systems must remain adaptable across accelerator generations whose economic lives are far shorter than the networks they serve. Applications must perform reliably at the tail of the latency distribution, not merely at its mean. The uplink must be able to carry what physical AI will ask it to carry. Cybersecurity and safety responsibilities must be clearly defined and enforced in the orchestration layer. Policymakers must understand the relationship between distributed computing, telecommunications infrastructure, electricity requirements, and local communities. And the resulting investment decisions should be driven by commercial evidence rather than by the assumption that every network location requires powerful AI accelerators, a discipline that will be tested through 2030 as telecommunications operators, cloud providers, semiconductor manufacturers, and AI developers compete for capital that has many other claims on it.
Why the Title “Tower Inference” Fits the Paper
The title brings together two concepts that have historically belonged to different technological and economic domains. The tower represents telecommunications infrastructure, the physical network through which people, businesses, and machines communicate, with all of its real estate, power, spectrum, and operational discipline. Inference represents the execution of artificial intelligence models, the computational process through which trained intelligence becomes useful in applications and operational decisions. Their combination identifies the possible emergence of a new infrastructure function in which a telecommunications network no longer serves only as the connection between a machine and a remote AI system but becomes part of the computing environment that delivers the intelligence the machine uses.
This distinction makes Tower Inference a useful research concept within the Five-Layer AI Economy. Energy infrastructure supports distributed computing facilities whose geography is new. Semiconductors provide the processing capability, and the contest between general-purpose and purpose-built silicon will shape the architecture. Datacenters expand beyond traditional hyperscale campuses to include suitable telecommunications locations, operated by companies new to the role. Models execute across a coordinated computing hierarchy, and their shrinking size and falling cost make distribution feasible. Applications and autonomous agents generate demand for intelligence at different geographic and operational scales, and their safety requirements determine where that intelligence may live. The framework demonstrates how a change in one part of the infrastructure system can influence the economics of the entire AI economy, and its central insight is not that computing will become completely decentralized. Rather, artificial intelligence is likely to require a more sophisticated relationship between centralization and distribution, in which large computing campuses provide scale, distributed network facilities provide selected advantages in location, connectivity, and operational integration, and intelligent applications determine the value of combining them.
By 2030, the telecommunications industry may be judged not only by the speed, coverage, and reliability of its networks but also by its ability to support valuable intelligence services. For investors, this represents a potential expansion of telecommunications infrastructure economics, to be approached with the discipline of the investment framework proposed here. For corporations and startups, it creates opportunities to develop applications that combine physical devices with geographically distributed computing. For policymakers, it introduces new relationships among telecommunications, artificial intelligence, semiconductor infrastructure, energy systems, national security, and global inclusion. For the Five-Layer AI Economy, it reinforces the central importance of understanding artificial intelligence as an interconnected industrial system rather than as a collection of models.
The October 2026 announcements may eventually be remembered as early milestones in a much larger transformation. Whether that transformation produces substantial commercial returns remains to be established, and the evidence that will establish it, or fail to, will arrive in the field-trial results of 2027 and the enterprise contracts of 2028. Nevertheless, the emerging direction is clear enough to deserve sustained examination. Tower Inference describes a future in which mobile networks do more than connect intelligent machines: they become part of the infrastructure through which those machines acquire, share, and apply intelligence. That possibility, not simply faster wireless communication, is what makes Tower Inference an important subject for research extending from 2027 through 2030 and beyond.
Endnotes and References:
[1] NVIDIA Newsroom. October 21, 2026. NVIDIA, Nokia and Leading Telecom Operators Turn Networks Into Distributed Intelligence Platforms (GTC Berlin). https://nvidianews.nvidia.com/news/nvidia-nokia-and-leading-telecom-operators-turn-networks-into-distributed-intelligence-platforms
[2] Nokia Corporation (Nokia Blog). October 21, 2026. From the Lab to the Live Network: Nokia, T-Mobile and NVIDIA Complete World’s First AI-RAN Call. https://www.nokia.com/blog/from-the-lab-to-the-live-network-nokia-t-mobile-and-nvidia-complete-worlds-first-ai-ran-call/
[3] Digital Today (Korea). October 28, 2026. T-Mobile Completes World’s First AI-RAN Call on Commercial 5G Network. https://www.digitaltoday.co.kr/en/view/120334/t-mobile-completes-worlds-first-ai-ran-call-on-commercial-5g-network
[4] Unite.AI (reporting NVIDIA GTC Berlin remarks by Jensen Huang). October 21, 2026. NVIDIA, Nokia and Telecom Operators Turn Networks Into Distributed Intelligence Platforms. https://www.unite.ai/?p=491156
[5] Deutsche Telekom AG (Media Information). October 2026. Deutsche Telekom and Nokia Advance Commercial AI-RAN With Measurable Radio Performance Gains on NVIDIA AI Aerial. https://www.telekom.com/en/newsroom/latest-updates/media-information/2026/10/ai-ran-lab-tests
[6] Elisa Oyj (Press Release, via Inderes). October 21, 2026. Elisa Taps NVIDIA and Nokia for Its Journey Towards Commercial AI-RAN in Finland. https://www.inderes.fi/releases/elisa-oyj-elisa-taps-nvidia-and-nokia-for-its-journey-towards-commercial-ai-ran-in-finland
[7] The Fast Mode (Orange Business and NVIDIA announcement, GTC Berlin). October 21, 2026. Orange Business, NVIDIA Advance Distributed Voice AI Inference for Enterprises. https://www.thefastmode.com/technology-solutions/51180-orange-business-nvidia-advance-distributed-voice-ai-inference-for-enterprises
[8] SoftBank Corp. (Press Release). October 20, 2026. SoftBank Corp. and Nokia Agree to Conduct AI-RAN Field Trial on SoftBank’s Commercial Network. https://www.softbank.jp/en/corp/news/press/sbkk/2026/20261020_01/
[9] John Saw, Chief Technology Officer, T-Mobile US (T-Mobile Newsroom). October 27, 2026. Physical AI Is Leaving the Lab. https://www.t-mobile.com/news/network/physical-ai-is-leaving-the-lab
[10] Fei-Fei Li, Sequoia Professor of Computer Science, Stanford University; Co-Director, Stanford Institute for Human-Centered AI; Co-Founder, World Labs (remarks at AI Startup School, San Francisco). June 16, 2025. AGI Will Not Be Complete Without Spatial Intelligence. https://www.startuphub.ai/news/agi-will-not-be-complete-without-spatial-intelligence-says-fei-fei-li
[11] Stanford Institute for Human-Centered AI (AI Index Steering Committee). April 13, 2026. The AI Index 2026 Annual Report. https://hai.stanford.edu/ai-index/2026-ai-index-report
[12] Gurudatt Shenoy, Cisco (Cisco Blogs). October 21, 2026. Turn Distributed Capacity Into AI Infrastructure (AI Inference Routing Blueprint, NVIDIA GTC Berlin). https://blogs.cisco.com/news/turn-distributed-capacity-into-ai-infrastructure
[13] Advanced Television. October 29, 2025. Nvidia to Invest $1bn in Nokia. https://www.advanced-television.com/?p=273143
[14] AI-RAN Alliance. Founded February 2024. AI-RAN Alliance: AI-for-RAN, AI-and-RAN, AI-on-RAN Working Groups. https://ai-ran.org/
[15] Lopamudra Kundu, Xingqin Lin, Rajesh Gadiyar, Jean-Francois Lacasse and Shuvo Chowdhury (NVIDIA Research; IEEE Network). January 15, 2025. AI-RAN: Transforming RAN with AI-Driven Computing Infrastructure, arXiv:2501.09007. https://arxiv.org/abs/2501.09007
[16] Ericsson (Press Release). June 11, 2026. The RAN Gets Smarter: Ericsson Puts AI Where It Matters — AI in RAN Software Subscription. https://www.ericsson.com/en/news/2026/6/the-ran-gets-smarter-ericsson-puts-ai-where-it-matters
[17] SoftBank Corp. (Press Release). February 25, 2026. SoftBank Corp. and Nokia to Enable Execution of External AI Workloads on AI-RAN. https://www.softbank.jp/en/corp/news/press/sbkk/2026/20260225_01/
[18] Michele Polese, Niloofar Mohamadi, Salvatore D’Oro, Leonardo Bonati and Tommaso Melodia (Institute for the Wireless Internet of Things, Northeastern University). July 2025 (revised December 2025). Beyond Connectivity: An Open Architecture for AI-RAN Convergence in 6G, arXiv:2507.06911. https://arxiv.org/abs/2507.06911
[19] Tommaso Melodia, William Lincoln Smith Professor, Northeastern University (lecture summary, National Yang Ming Chiao Tung University). February 17, 2025. Open 6G: Toward Open, Programmable, and AI-Powered NextG Wireless Systems. https://www.cs.nycu.edu.tw/announcements/detail/12857?locale=en
[20] NVIDIA Corporation (Newsroom). August 26, 2026. NVIDIA Announces Financial Results for Second Quarter Fiscal 2027. https://nvidianews.nvidia.com/news/nvidia-announces-financial-results-for-second-quarter-fiscal-2027
[21] Cisco Systems (White Paper, Version 1.0). 2026. Cisco AI Inference Architecture for Wide Area Networks: Design/High-Level Blueprint. https://www.cisco.com/c/en/us/solutions/collateral/design-zone/ai-inference-wan-wp.html
[22] Khaled B. Letaief, Yuanming Shi, Jianmin Lu and Jianhua Lu (IEEE Journal on Selected Areas in Communications, Vol. 40, No. 1). 2022. Edge Artificial Intelligence for 6G: Vision, Enabling Technologies, and Applications, arXiv:2111.12444. https://arxiv.org/abs/2111.12444
[23] Yong Xiao, Guangming Shi, Yingyu Li, Walid Saad (Virginia Tech) and H. Vincent Poor (Princeton University). October 2020. Towards Self-Learning Edge Intelligence in 6G, arXiv:2010.00176. https://arxiv.org/abs/2010.00176
[24] Wen Tong (Huawei) and Geoffrey Ye Li (Imperial College London). September 2021. Nine Challenges in Artificial Intelligence and Wireless Communications for 6G, arXiv:2109.11320. https://arxiv.org/abs/2109.11320
[25] Ericsson (Ericsson Blog). October 2026. 6G Uplink Performance: Why the Uplink Matters in the AI Era. https://www.ericsson.com/en/blog/2026/10/6g-uplink-performance
[26] Iain Morris, Light Reading (quoting William Webb, independent consultant and former IEEE President). 2026. Ericsson Eyes AI Glasses ‘Uplink’ Boom as Critics See Gloom. https://www.lightreading.com/5g/ericsson-eyes-ai-glasses-uplink-boom-as-critics-see-gloom
[27] Ericsson (Ericsson Blog; Senior Expert, Internet Architecture, and Ericsson ConsumerLab). September 2026. AI Is Flipping the Network Traffic Model. https://www.ericsson.com/en/blog/2026/9/ai-driven-devices-and-uplink
[28] IEEE Communications Society Technology Blog (summarizing the Ericsson Mobility Report, June 2026). June 16, 2026. Ericsson’s June 2026 Mobility Report Highlights + AI Impact on Network Traffic. https://techblog.comsoc.org/2026/06/16/ericssons-june-2026-mobility-report-agentic-ai-impact-on-network-traffic/
[29] Light Reading. October 29, 2026. T-Mobile: We’ll Be First With 6G, and No to SpaceX MVNO. https://www.lightreading.com/6g/t-mobile-we-ll-be-first-with-6g-and-no-to-spacex-mvno
[30] Light Reading. 2026. Ericsson Has Pinned Its 6G Hopes on AI’s Rising ‘Uplink’ Demands. https://www.lightreading.com/6g/ericsson-has-pinned-its-6g-hopes-on-ais-rising-uplink-demands
[31] RCR Wireless News. October 29, 2026. T-Mobile Beats Forecasts, Sets Out Mid-Term AI Build-Out and Pay-Out (Q3 2026 Earnings Call Coverage). https://rcrwireless.com/20261029/ai-infrastructure/t-mobile-beats-forecasts-sets-out-mid-term-ai-build-out-and-pay-out
[32] Jeffrey G. Andrews, Truchard Family Endowed Chair in Engineering, The University of Texas at Austin (keynote at Samsung 6G Forum, reported by AITopics). July 23, 2022. Deep Learning in the 6G Air Interface. https://aitopics.org/doc/news:CB55F210
[33] Capacity Media (reporting the International Energy Agency’s ‘Key Questions on Energy and AI’). April 17, 2026. AI Data Centres Could Triple Electricity Consumption by 2030, IEA Warns. https://capacityglobal.com/news/iea-ai-data-centres-energy-grid-concerns/
[34] Nokia Corporation (Newsroom). October 28, 2025. Nokia Partners With NVIDIA: NVIDIA to Invest $1 Billion in Nokia to Accelerate AI-RAN Innovation. https://www.nokia.com/newsroom/nokia-partners-with-nvidia/
[35] T-Mobile US, Inc. (Form 8-K, Exhibit 99.1, U.S. Securities and Exchange Commission). October 28, 2026. T-Mobile US Reports Third Quarter 2026 Results. https://www.sec.gov/Archives/edgar/data/0001283699/000128369926000128/tmus09302026ex991.htm
[36] Verizon Communications Inc. (Third-Quarter 2026 Earnings Release). October 26, 2026. Verizon Delivers Another Record-Breaking Quarter With 3Q 2026 Results Driven by Strong Postpaid Phone and Account Additions. https://www.verizon.com/about/news/verizon-delivers-another-record-breaking-quarter
[37] TelecomLead. October 2026. AT&T Q3 2026 Revenue Rises 5.3% to $32.3 Billion as Fiber, 5G and Broadband Growth Drives Earnings. https://telecomlead.com/5g/att-q3-2026-revenue-rises-5-3-to-32-3-billion-as-fiber-5g-and-broadband-growth-drives-earnings-128270
[38] ABI Research (Insight). 2026. What the SoftBank/Nokia Announcement Means: From Network Infrastructure to AI Platform Infrastructure. https://www.abiresearch.com/market-research/insight/7787520-what-the-softbanknokia-announcement-means-
[39] Nokia Corporation (Newsroom). October 20, 2026. Nokia and SoftBank Corp. Agree to Conduct Commercial AI-RAN Field Trial in Japan. https://www.nokia.com/newsroom/nokia-and-softbank-corp-agree-to-conduct-commercial-ai-ran-field-trial-in-japan/
[40] Nick Patience, The Futurum Group. July 15, 2026. Nokia’s AI-RAN Platform Puts a Number on the Software-Defined RAN Bet. https://futurumgroup.com/insights/nokias-ai-ran-platform-puts-a-number-on-the-software-defined-ran-bet/
[41] Mobile World Live. October 2026. T-Mobile US Pushes Physical AI to the Edge. https://www.mobileworldlive.com/t-mobile-us/t-mobile-us-pushes-physical-ai-to-the-edge
[42] AT&T Inc. (Form 8-K, Exhibit 99.1, U.S. Securities and Exchange Commission). October 21, 2026. AT&T Delivers Strong Third-Quarter 2026 Results With Its Best Consolidated Service Revenue Growth in More Than Four Years. https://www.sec.gov/Archives/edgar/data/0000732717/000073271726000370/t-3q2026exhibit991.htm
[43] Nokia Corporation (Interim Report). October 22, 2026. Nokia Corporation Interim Report for Q3 2026. https://www.nokia.com/newsroom/nokia-corporation-interim-report-for-q3-2026/
[44] Telefonaktiebolaget LM Ericsson (Third Quarter Report 2026). October 15, 2026. Ericsson Reports Third Quarter Results 2026. https://mb.cision.com/Main/15448/4409101/4319119.pdf
[45] Data Center Dynamics. October 26, 2026. Verizon Signs $2bn AWS AI Infrastructure Deal. https://www.datacenterdynamics.com/en/news/verizon-signs-2bn-aws-ai-infrastructure-deal
[46] Investing.com. October 22, 2026. Nokia Q3 2026 Slides: AI Growth Shines Amid Cash Flow Concerns. https://www.investing.com/news/company-news/nokia-q3-2026-slides-ai-growth-shines-amid-cash-flow-concerns-93CH-4966385
[47] Fierce Network (quoting Stefan Pongratz-era analysis and Fierce commentary by Chua). October 2026. Nokia and Nvidia Keep AI-RAN Momentum Going as They Edge Closer to Commercial Reality. https://www.fierce-network.com/wireless/nokia-nvidia-keep-ai-ran-train-moving-toward-commercial-deployment
[48] Daron Acemoglu, Institute Professor, Massachusetts Institute of Technology; 2024 Nobel Laureate in Economic Sciences (interview, Fortune). June 21, 2026. Nobel Laureate Daron Acemoglu on AI, Productivity, Capitalism and Democracy. https://fortune.com/2026/06/21/nobel-laureate-daron-acemoglu-ai-productivity-capitalism-democracy/
[49] Erik Brynjolfsson, Director, Stanford Digital Economy Lab, Stanford Institute for Human-Centered AI (Financial Times op-ed, reported by Fortune). February 15, 2026. The AI Productivity Take-Off Is Finally Visible / Productivity Liftoff Along the J-Curve. https://fortune.com/2026/02/15/ai-productivity-liftoff-doubling-2025-jobs-report-transition-harvest-phase-j-curve/
[50] Converge! Network Digest. October 2025. NVIDIA to Invest $1B in Nokia: Partnership for AI-RAN and Data Center Networking. https://convergedigest.com/nvidia-to-invest-1b-in-nokia-partnership-for-ai-ran-and-data-center-networking/
[51] PYMNTS (reporting Fei-Fei Li in conversation with Jeetu Patel, Cisco). February 3, 2026. Fei-Fei Li Says AI Progress Now Depends on Physical Context. https://www.pymnts.com/?p=3451498
[52] Computer Weekly (quoting Qualcomm executives). 2026. Qualcomm Plots Out 6G, Wi-Fi 8 Future With AI as the New User Interface. https://www.computerweekly.com/news/366639874/Qualcomm-plots-out-6G-Wi-Fi-8-future-with-AI-as-the-new-user-interface
[53] TelecomLead. October 2026. Elisa and Deutsche Telekom Partner With Nokia and NVIDIA on AI-RAN: 15% Faster 5G Uplink and 2027 Commercial Plans. https://telecomlead.com/telecom-equipment/elisa-and-deutsche-telekom-partner-with-nokia-and-nvidia-on-ai-ran-15-faster-5g-uplink-and-2027-commercial-plans-128272
[54] Gcore (Gcore Blog, with Orange Business and NVIDIA). October 22, 2026. How Gcore Everywhere AI Is Enabling Distributed AI to Fire and Rescue Call Handling: In an Emergency Call, a Late Answer Is a Missing Answer. https://gcore.com/blog/everywhere-ai-emergency
[55] Nokia Corporation (Newsroom). October 21, 2026. Nokia Advances AI-RAN Toward Commercial Deployment and Physical AI, Launches Sovereign Agentic AI Platform for Telco Autonomy With NVIDIA. https://www.nokia.com/newsroom/nokia-advances-ai-ran-toward-commercial-deployment-and-physical-ai-launches-sovereign-agentic-ai-platform-for-telco-autonomy-with-nvidia
[56] 3rd Generation Partnership Project (3GPP). 2025–2026. Release 20: 5G-Advanced Evolution and 6G Studies. https://www.3gpp.org/specifications-technologies/releases/release-20
[57] Broadband Breakfast. March 3, 2026. Qualcomm Sets 2029 6G Timeline as Huawei Calls for Expanding and Advancing 5G. https://broadbandbreakfast.com/qualcomm-sets-2029-6g-timeline-as-huawei-calls-for-expanding-and-advancing-5g
[58] CNBC. August 26, 2026. Nvidia Earnings Takeaways: Huang Forecasts 70% Fiscal 2028 Revenue Growth, Far Above Estimates. https://www.cnbc.com/2026/08/26/nvidia-nvda-earnings-report-q2-2027-live-updates.html
[59] TelecomTV (quoting Pallavi Mahajan, Chief Technology and AI Officer, Nokia). October 20, 2026. Nokia and SoftBank Corp. Agree to Conduct Commercial AI-RAN Field Trial in Japan. https://www.telecomtv.com/content/nokia/nokia-and-softbank-corp-agree-to-conduct-commercial-ai-ran-field-trial-in-japan-56490/
[60] RCR Wireless News. October 5, 2026. Ericsson Says AI RAN Will Require Smarter Networks — and Stronger Uplinks. https://rcrwireless.com/20261005/networks/ericsson-ai-ran-networks
[61] Amazon Web Services. Product documentation (accessed October 2026). AWS Wavelength: 5G Edge Computing Infrastructure. https://aws.amazon.com/wavelength/
[62] Muninder Sambi, Google Cloud (Google Cloud Blog, Google Cloud Next ’26). April 22, 2026. New Innovations in Google Distributed Cloud. https://cloud.google.com/blog/topics/hybrid-cloud/google-distributed-cloud-at-next26
[63] Gigi Onag, Light Reading. October 22, 2026. SoftBank, Nokia Target 2027 for Commercial AI-RAN Trial in Japan. https://lightreading.com/ai-machine-learning/softbank-nokia-target-2027-for-commercial-ai-ran-trial-in-japan
[64] Eric Mboizi, eWeek. October 22, 2026. SoftBank and Nokia’s AI-RAN Trial: What It Needs to Prove. https://www.eweek.com/artificial-intelligence/news-softbank-nokia-ai-ran-trial-apac-japan/
[65] Center for Strategic and International Studies (CSIS). 2025. 6G and AI: How Wireless Technology and Edge Devices Are Shaping the Global AI Race. https://www.csis.org/analysis/6g-and-ai-how-wireless-technology-and-edge-devices-are-shaping-global-ai-race
[66] The White House (National Security Presidential Memorandum NSPM-8). December 19, 2025. Winning the 6G Race. https://www.whitehouse.gov/presidential-actions/2025/12/national-security-presidential-memorandum-nspm-8-0bda
[67] Data Center Dynamics. March 2026. Nokia Maps Out AI-RAN Vision at MWC Ahead of 6G Era. https://www.datacenterdynamics.com/en/news/nokia-maps-out-ai-ran-vision-at-mwc-ahead-of-6g-era/
[68] Xingqin Lin et al. (arXiv). December 2024. 3GPP Evolution From 5G to 6G: A 10-Year Retrospective, arXiv:2412.21077. https://arxiv.org/abs/2412.21077
[69] Data Center Knowledge. 2026. DOE: AI Data Centers Are Transforming America’s Transmission Map. https://www.datacenterknowledge.com/energy-power-supply/doe-says-ai-data-centers-are-rewriting-america-s-transmission-map
[70] U.S. Department of Energy, Office of Electricity. July 9, 2026. DOE’s Office of Electricity Publishes 2026 Draft National Transmission Needs Study to Strengthen America’s Grid. https://www.energy.gov/oe/articles/does-office-electricity-publishes-2026-draft-national-transmission-needs-study
[71] Grid Strategies LLC. 2025. National Load Growth Report 2025: Data Centers Drive About 55% of Forecast Growth. https://gridstrategiesllc.com/wp-content/uploads/Grid-Strategies-National-Load-Growth-Report-2025.pdf
[72] The White House (Fact Sheet). December 19, 2025. President Donald J. Trump Takes Action to Win the 6G Race. https://www.whitehouse.gov/fact-sheets/2025/12/fact-sheet-president-donald-j-trump-takes-action-to-win-the-6g-race
[73] ExecutiveGov (reporting on the National Telecommunications and Information Administration). July 28, 2026. NTIA Rallies 20+ Allied Nations Around New 6G Security Initiative: Call to Action for 6G Leadership and Security. https://www.executivegov.com/articles/ntia-6g-leadership-security-initiative
[74] The White House (Office of Science and Technology Policy). July 2025. Winning the Race: America’s AI Action Plan. https://www.whitehouse.gov/wp-content/uploads/2025/07/Americas-AI-Action-Plan.pdf
[75] Dean Bubley, Disruptive Analysis (Broadband Breakfast expert opinion). January 17, 2026. Winning in 6G Will Not Require More Spectrum. https://broadbandbreakfast.com/dean-bubley-winning-in-6g-will-not-require-more-spectrum/
[76] National Institute of Standards and Technology (Raymond Sheh and Martin Stanley). April 7, 2026. Concept Note: AI RMF Profile on Trustworthy AI in Critical Infrastructure. https://www.nist.gov/programs-projects/concept-note-ai-rmf-profile-trustworthy-ai-critical-infrastructure
[77] National Institute of Standards and Technology. January 2023 (updated 2026). Artificial Intelligence Risk Management Framework (AI RMF 1.0) and Profiles. https://www.nist.gov/itl/ai-risk-management-framework
[78] 3GPP TSG RAN WG3 (Technical Report 38.745, Release 20, V20.0.0). March 2026. Study on Artificial Intelligence (AI)/Machine Learning (ML) for NG-RAN Phase 3. https://whatthespec.net/3gpp/spec/38.745
[79] Kristalina Georgieva, Managing Director, International Monetary Fund (World Economic Forum, Davos; reported by Fortune). January 23, 2026. IMF Chief Warns of an AI ‘Tsunami’ Coming for Young People and Entry-Level Jobs. https://fortune.com/2026/01/23/imf-chief-warns-ai-tsunami-entry-level-jobs-gen-z-middle-class/
[80] Rebeca Grynspan, Secretary-General, UNCTAD (reported by Aduana News). April 2025. UNCTAD Calls for Equitable Sharing of the Benefits of Artificial Intelligence. https://aduananews.com/en/la-unctad-pide-una-distribucion-equitativa-de-los-beneficios-de-la-inteligencia-artificial/
[81] United Nations Conference on Trade and Development (UNCTAD). April 2025. Technology and Innovation Report 2025: Inclusive Artificial Intelligence for Development. https://unctad.org/system/files/official-document/tir2025_en.pdf
[82] Ajay Banga, President, World Bank Group (Global AI and Digital Summit, Seoul; reported by Trend News Agency). October 20, 2026. World Bank Group Unveils First Digital Strategy for Developing Nations. https://www.trend.az/world/4237036.html
[83] Gaurav Nayyar, Director, World Development Report 2026, World Bank Group (UN Audiovisual Library / UNifeed). August 2026. World Bank / World Development Report (soundbites). https://media.un.org/unifeed/en/asset/d361/d3615192
[84] World Bank Group. August 4, 2026. World Development Report 2026: The Promise of Artificial Intelligence. https://ppp.worldbank.org/library/world-development-report-2026-promise-artificial-intelligence
[85] Theodore S. Rappaport, David Lee/Ernst Weber Professor, NYU Tandon School of Engineering and Founding Director, NYU WIRELESS. NYU Tandon News. Forget 5G, Let’s Talk About 6G. https://engineering.nyu.edu/news/forget-5g-lets-talk-about-6g
[86] Jakob Hoydis, Fayçal Aït Aoudia, Alvaro Valcarce and Harish Viswanathan (Nokia Bell Labs). December 2020 / IEEE Communications Magazine 2021. Toward a 6G AI-Native Air Interface, arXiv:2012.08285. https://arxiv.org/abs/2012.08285
[87] Jeffrey G. Andrews, Director of 6G@UT, The University of Texas at Austin (ITU Journal webinar). December 6, 2023. Unlocking New Capacity in 6G Cellular Systems via Site-Specific ML-Aided Design. https://www.itu.int/en/journal/j-fet/webinars/20231206/Pages/default.aspx
