AI in Telecom
Network X Americas: AT&T and Comcast reveal huge AI impact on network operations
Echoing a recent Cisco report, telecom leaders at the Network X Americas conference (held in Irving, TX last week) noted that AI is fundamentally shifting traffic patterns while having a very positive impact on network operations. With billions of connected sensors and devices (like autonomous vehicles generating 20GB of data per day), operators are forced to prioritize uplink capacity and low latency over traditional consumer downlink traffic.
AT&T’s network CTO, Yigal Elbaz, cited the robo-taxi as a bellwether for how AI is affecting network traffic. Each Waymo vehicle generates about 20 gigabytes of data per day, roughly 30 times the amount a typical mobile user consumes. Most of that traffic flows from the car to the cloud. “Every other week,” Elbaz noted, “a new flavor of a frontier AI model drops on us.”
“We already have about 700,000 changes on a daily basis in our network made by AI,” said Elbaz, noting that AT&T has built a proprietary foundation AI model because standard large language models (LLMs) don’t understand KPIs, network alarms or fiber deployment specifics. He cited a 20-25% cost reduction and 12-15% better results than general-purpose models.
In his keynote speech, Comcast EVP and Chief Network Officer Elad Nafshi described 200 edge compute centers capable of self-healing 77% of network events. He touted AI chipsets close enough to customers’ homes to pinpoint outside plant faults with 99.2% precision, and a partnership with Nvidia to push that edge platform further.
Nafshi highlighted the gap in network provider promises vs delivery with a hypothetical small-business use case example. A pizza shop operator, could materially change workflow and productivity if the service provider delivered an AI-enabled concierge—built on a task-optimized small language model—to manage order intake and customer interaction. In that scenario, the network evolves from a passive access pipe into an application-aware platform that augments business operations. The concept is credible from a technical standpoint, but remains largely theoretical until operators can effectively reach and educate SMB customers who still perceive connectivity as a fixed monthly expense.
Both AT&T and Comcast Israeli executives said this was more than modernization and discussed the changes in what a network does. The network is now a platform, not a pipe. Today’s network learns, adapts and increasingly acts on behalf of its customers. But I can’t help but wonder if the customers know… or if that network value will ever trickle down to the customers who need it most.
In a keynote panel session titled, ” Convergence in action – Competing, scaling and winning in the AI-driven connectivity market,” Josh Goodell, AT&T’s VP of Broadband and Converged Product Development, framed the company’s objective as becoming “the greatest simplifier of our customers’ lives” while instilling “connectivity confidence.” That positioning is notable for a sector that has historically under-communicated its value proposition beyond basic service metrics.
The broader industry narrative appears to be shifting. Historically, go-to-market strategies emphasized throughput benchmarks and promotional pricing. As Omdia’s Ruth Brown (panel session moderator) observed, packaging has been largely defensive, optimized around billing constructs rather than differentiated user experience. The emerging model instead centers on networks that operate contextually and autonomously—delivering value in ways that are largely invisible to the end user.
Derek Peterson, CTO of Boingo Wireless, articulated a parallel issue in venue networks, describing the “stadium problem.” Operators dimension infrastructure for peak ingress and then underutilize that capacity once users are inside the venue. The architectural question is no longer solely about capacity provisioning, but about service-layer innovation on top of that capacity. At Petco Park, Boingo leveraged existing network assets to enable pre-entry commerce, driving incremental revenue before fans pass through the gates. The infrastructure was not the constraint; the limiting factor was identifying and executing on higher-order use cases.
A similar disconnect persists in the industry’s framing of the digital divide. AT&T’s John Stankey and others have suggested the gap is nearing closure, citing expanded fiber footprints and fixed wireless access. While coverage metrics have improved, the divide has never been purely a function of infrastructure availability. Adoption is equally constrained by affordability and, critically, by perceived value. If connectivity continues to be positioned as a commoditized utility, the most economically vulnerable segments—those with the greatest need for digital enablement—remain the least likely to engage.
This is particularly relevant in an AI-driven economy. The users and small enterprises that could benefit most from intelligent, network-delivered services are often those least exposed to the evolving capabilities of the platform. The industry risks over-indexing on measurable deployment milestones while under-communicating the functional value of next-generation networks.
The Network X keynotes underscored that the technical roadmap is largely in place. Network operators are advancing toward networks capable of real-time traffic learning, proactive cybersecurity at the edge, and highly personalized in-home connectivity experiences. These capabilities represent a more compelling value proposition than traditional service tier comparisons.
However, the central challenge remains go-to-market execution. The industry has demonstrated that it can architect and deploy these capabilities at scale. It has yet to establish a clear, effective framework for articulating that value to end users and enterprises in a way that drives adoption.
As a final observation, the broader telecom ecosystem—illustrated by developments such as autonomous vehicle platforms—already depends on AI-enabled, highly distributed network intelligence. While the underlying infrastructure is incrementally aligning with these requirements, the industry dialogue around its broader economic and societal implications remains underdeveloped.
References:
Cisco report: Agentic AI to reshape WAN traffic, AI inference will be ~25% of total traffic by 2035
Will the wave of AI generated user-to/from-network traffic increase spectacularly as Cisco and Nokia predict?
Telecom operators investing in Agentic AI while Self Organizing Network AI market set for rapid growth
Analysis: Cisco, HPE/Juniper, and Nvidia network equipment for AI data centers
Cisco CEO sees great potential in AI data center connectivity, silicon, optics, and optical systems
The Financial Trap of Autonomous Networks: Scaling Agentic AI in the Telecom Core
Ericsson integrates Agentic AI into its NetCloud platform for self healing and autonomous 5G private networks
STL Partners webinar: Agentic AI needed for RAN autonomy & efficiency
Nokia to showcase agentic AI network slicing; Ericsson partners with Ookla to measure 5G network slicing performance
Agentic AI and the Future of Communications for Autonomous Vehicles (V2X)
Telecom data centers must be redesigned for the AI era with rack scale architectures, enhanced power & cooling requirements
Is the “far edge” a bridge to far to cross for AI inferencing? What about “Distributed AI Grids”?
T-Mobile US announces new broadband wireless and fiber targets, 5G-A with agentic AI and live voice call translation
Intel and AI chip startup SambaNova partner; SN50 AI inferencing chip max speed said to be 5X faster than competitive AI chips
CES 2025: Intel announces edge compute processors with AI inferencing capabilities
Cisco report: Agentic AI to reshape WAN traffic, AI inference will be ~25% of total traffic by 2035
Executive Summary:
Consumer-driven AI traffic [1.] currently represents a marginal share of aggregate Internet traffic. However, accelerating adoption of agentic AI is expected to materially reshape traffic composition over the next decade. In its “AI Impact on Wide Area Networks” report, Cisco projects that AI will emerge as the dominant driver of network traffic growth. As consumer AI adoption approaches “near-universal usage,” AI and agentic AI are forecast to increase consumer-driven network traffic by approximately 6.6× by the mid-2030s (see chart below).
Cisco estimates that this AI expansion will account for roughly 63% of incremental traffic growth relative to non-AI scenarios. The study focuses specifically on WAN implications, rather than data center or GPU infrastructure, and provides guidance on network design and capacity planning. Methodologically, the report integrates real-world traffic observations (via Cisco Crosswork Assurance User Experience), third-party industry datasets, and controlled laboratory evaluations of AI agents to characterize how AI-generated traffic diverges from conventional web traffic patterns.
Token-consumption data shows nearly 10x year-over-year growth, while in some service provider measurements Cisco is seeing ~4x growth in just eight months. Sustained growth at these rates means AI traffic will become a meaningful component of overall network traffic by 2035.
Note 1. Consumer AI traffic has a few defining technical traits: it is still dominated by short text-based exchanges, but it is becoming more stateful, more upstream-heavy, and more latency-sensitive as users move from simple prompts to agentic workflows and multimodal interactions. Today’s consumer AI traffic is still overwhelmingly text-oriented, which is one reason the aggregate bandwidth impact remains modest despite rapid adoption. Comcast’s network observation is a useful real-world proxy: 97.1% of AI traffic was text-based, while images accounted for 2.6% and video only 0.3%. The key technical implication is that current traffic volumes are often limited more by conversation frequency and session behavior than by very large payloads, though that changes quickly as users adopt image, audio, and video generation.
Although AI inference traffic is currently “negligible” relative to dominant categories such as video streaming, Cisco projects it will comprise approximately 25% of total network traffic by 2035 (see chart below). At that point, AI traffic is expected to represent a “meaningful component” of overall network load. Importantly, AI-generated traffic exhibits distinct characteristics: inference flows are approximately twice the duration of typical web transactions, demonstrate higher upstream bandwidth demand, and operate at “software speed” rather than human interaction rates.
The emergence of AI agents as “power users” further amplifies these dynamics. Cisco notes that agent-executed tasks can generate up to 450% more traffic per task compared to human-driven interactions. This shift is expected to drive operator adoption of “flow-aware network and security systems” as traffic patterns become increasingly machine-driven and less predictable.
Cisco’s broader framing is that AI traffic “isn’t just adding traffic,” but is changing the shape of traffic, with inference flows running about twice as long as typical web transactions and, in some cases, generating up to 450% more traffic per task when an agent executes the workload. AI inference sessions tend to hold resources longer, create more sustained flows, and push operators to think in terms of flow-aware behavior rather than only peak-throughput sizing. Cisco also notes that about 9% of AI inference flows carry more upstream than downstream traffic, versus about 0.5% for typical web traffic, which is a meaningful shift for access and broadband networks. Cisco reports that approximately 9% of AI inference flows are upstream-dominant, compared to roughly 0.5% for traditional web traffic, with this divergence expected to widen alongside increased agentic AI utilization. In parallel, latency sensitivity is anticipated to become a more critical performance parameter for AI-driven applications.
Latency and symmetry:
AI traffic is also more sensitive to latency than many ordinary consumer web transactions because the user experience is often conversational and interactive, with the expectation of near-immediate turn-taking. Cisco describes AI inference as operating at “software speed” rather than human speed, which means small delays can be more noticeable and operationally important. At the same time, upstream demand becomes more significant because prompts, context, attachments, and agent-generated actions can increase return-path traffic, especially as multimodal inputs and agentic tool use expand.
Multimodal growth:
The biggest step-up in technical impact comes when consumer AI shifts from text-only prompting to multimodal generation and agent-driven workflows. In those cases, each task can involve multiple model calls, retrieval steps, tool invocations, and richer media payloads, which expands both flow count and bytes per session. Cisco’s study suggests that this is why AI traffic will increasingly require “flow-aware network and security systems,” because the traffic profile is not just larger, but structurally different from conventional browsing.
Infrastructure Implications:
Telecom infrastructure is becoming “increasingly intertwined with hyperscale infrastructure, not because operators are leading AI investment, but because they are becoming part of the ecosystem that supports it,” analyst firm MTN Consulting said in an April 27th research note. “Demand for optical transport, data-center interconnect, and edge infrastructure is rising as telecom networks carry growing volumes of cloud and AI-driven traffic,” the firm said.
“AI network traffic is already reshaping infrastructure needs. What we are seeing is clear: AI isn’t just adding traffic. It’s changing the shape of traffic,” Javier Antich, principal product management engineer in the CTO office of Cisco’s provider connectivity group, and Gurudatt Shenoy, SVP, product management, provider connectivity, explained in this blog post.
These shifts are beginning to influence access network evolution. Fiber networks already provide relatively symmetric throughput and low latency, while cable operators are advancing similar capabilities through DOCSIS upgrades. Mid-split and high-split architectures increase upstream spectrum allocation, enabling more balanced capacity profiles. Concurrently, Tier 1 operators such as Comcast and Charter Communications are introducing low-latency enhancements within DOCSIS networks.
Operational data reflects early-stage impacts. Comcast Chief Network Officer Elad Nafshi noted at the Cable Next-Gen event in March that approximately 97.1% of AI traffic on Comcast’s network remains text-based, with images accounting for 2.6% and video just 0.3%, indicating that bandwidth-intensive multimodal AI traffic has yet to scale materially.
Network design impact:
For broadband and access networks, the immediate engineering issues are upstream traffic capacity, queue behavior, and latency consistency rather than raw total throughput alone. Symmetry upgrades (such as DOCSIS mid-split and high-split for MSOs), along with low-latency capabilities, are relevant because consumer AI creates more return-path pressure and more time-sensitive sessions. In other words, the challenge is not simply to carry more bytes; it is to carry more interactive sessions with predictable performance, especially as multimodal and agentic usage scales.
………………………………………………………………………………………………………………………………………………………………………………………………………….
References:
Will the wave of AI generated user-to/from-network traffic increase spectacularly as Cisco and Nokia predict?
Telecom operators investing in Agentic AI while Self Organizing Network AI market set for rapid growth
Analysis: Cisco, HPE/Juniper, and Nvidia network equipment for AI data centers
Cisco CEO sees great potential in AI data center connectivity, silicon, optics, and optical systems
The Financial Trap of Autonomous Networks: Scaling Agentic AI in the Telecom Core
Ericsson integrates Agentic AI into its NetCloud platform for self healing and autonomous 5G private networks
STL Partners webinar: Agentic AI needed for RAN autonomy & efficiency
Nokia to showcase agentic AI network slicing; Ericsson partners with Ookla to measure 5G network slicing performance
Agentic AI and the Future of Communications for Autonomous Vehicles (V2X)
Telecom data centers must be redesigned for the AI era with rack scale architectures, enhanced power & cooling requirements
Is the “far edge” a bridge to far to cross for AI inferencing? What about “Distributed AI Grids”?
T-Mobile US announces new broadband wireless and fiber targets, 5G-A with agentic AI and live voice call translation
Intel and AI chip startup SambaNova partner; SN50 AI inferencing chip max speed said to be 5X faster than competitive AI chips
CES 2025: Intel announces edge compute processors with AI inferencing capabilities
Inside Nokia’s new AI Networking Innovation Lab
- Silicon & Compute: Collaborating with AMD to optimize enterprise AI workloads alongside Nokia data center switches.
- Testing & Infrastructure: Partnering with Keysight Technologies to emulate workloads across Ultra Ethernet Consortium (UEC) and RoCEv2 transports.
- Hardware & Servers: Integrating high-performance platforms from Lenovo and Supermicro.
- Data Storage & Cloud: Working with Weka and cloud builders like Nscale to eliminate storage bottlenecks during heavy computational training.
Nokia’s AI Networking Innovation Lab is built upon three fundamental pillars: Technology Innovation, Ecosystem Collaboration, and Validation. Image credit: Nokia
………………………………………………………………………………………………………………….
Technology Innovation: The lab provides a dedicated space for AI partners to experiment with next-gen solutions across the entire networking stack – driving emerging standards forward with pioneering approaches to new protocols, switching silicon, congestion control, real-time telemetry, and automation.
“Partnering with Nokia in the AI Networking Innovation Lab has enabled us to benchmark and optimize AI networks under real-world conditions…Together, we are helping accelerate AI network adoption by giving operators and hyperscalers the validated insights needed for confident, large-scale deployment.”
Ecosystem Collaboration: True progress depends on a strong ecosystem of technology providers – silicon manufacturers, GPU developers, system, storage and test vendors, and cloud platforms – that work together to create highly-compatible AI-ready solutions. This facilitates joint testing for interoperability, improves integration, and ensures roadmaps are aligned across different hardware, software, and orchestration layers.
Travis Karr, Corporate Vice President, HPC and Sovereign AI at AMD believes customer collaboration and an open ecosystem are fundamental to accelerating AI innovation:
“By co-developing solutions with partners, such as Nokia in their AI networking innovation lab, we ensure our AMD enterprise AI solutions are tested with Nokia data center switches on real-world workloads and network demands. An open, standards-driven approach empowers customers to integrate seamlessly across heterogeneous environments, avoiding lock-in and fostering industry-wide advancement in AI.”
Validation: This positions the lab as the testing ground for Nokia Validated Designs, where customers and partners rigorously validate multi-vendor data center architectures under authentic AI training and inference workloads. By testing failure scenarios, congestion behavior, and operational automation, the lab turns NVDs into proven, deployable solutions — enabling predictable performance, faster deployment, and reduced operational complexity and risk for organizations navigating the AI era.
Arno van Huyssteen, Vice President of Global Telecommunications for Nscale:
“Nokia is a strategic networking partner for Nscale as we build towards AI Grid, and the engineering rigour behind their Validated Designs reflects the kind of innovation needed to enable next-generation AI infrastructure. The depth of hardware, software and failure testing behind those blueprints is what will give operators the confidence to deploy complex AI environments faster, with fewer integration risks and less operational disruption. We’re excited to collaborate in the AI Networking Innovation Lab to help push the boundaries of AI-native networking and validate the next generation of solutions before they reach production.”
A primary focal point inside the lab is managing data center congestion. Unlike traditional cloud traffic, back-end AI networks feature high-density data synchronization across massive GPU clusters. The lab uses advanced automation, AIOps, and lossless Ethernet solutions—such as the Nokia 7220 IXR-H6 switches—to handle these intense uplink and synchronization demands safely.
The AI Networking Innovation Lab supports Nokia’s broader strategy to accelerate the next era of AI-driven connectivity. As demand for AI infrastructure continues to grow, data center networking has become one of the most critical foundations of the global AI ecosystem. Through this investment, Nokia is strengthening its capabilities in AI and cloud infrastructure while advancing its vision of AI-native networking.
Rudy Hoebeke, Vice President of Software Product Management at Nokia:
“The launch of Nokia’s AI Networking Innovation Lab marks a major milestone in our commitment to drive the next era of AI-native connectivity. As the industry continues to evolve with solutions like scale-across and AI-Grid, this lab is poised to accelerate AI networking technology that will not only support but optimize these emerging industry offerings. This center gives our customers and partners early access to new technologies, deeper collaboration with the world’s leading AI ecosystem players, and the confidence that their networks are validated under more realistic AI conditions. By accelerating innovation and reducing deployment risks, we’re enabling the industry to deliver faster, more reliable, and more sustainable AI experiences to people and businesses everywhere.”
………………………………………………………………………………………………………………………
References:
Analysis: Nokia’s strong growth in Optical Networks and AI network infrastructure
Orange, Nokia, Nvidia, and Intel debate: ASICs vs. GPUs vs. General-Purpose CPUs for RAN Baseband Processing
Nokia’s AI Applications Study: “Physical AI” may require RAN redesign to support high‑volume, low‑latency uplink traffic
Australia’s NBN and Nokia demonstrate multi-generation optical technologies concurrently over existing FTTP infrastructure
Nokia to showcase agentic AI network slicing; Ericsson partners with Ookla to measure 5G network slicing performance
Tampnet to expand 5G offshore connectivity in the Gulf of Mexico using Nokia AirScale 5G radios
Dell’Oro: Analysis of the Nokia-NVIDIA-partnership on AI RAN
Orange, Nokia, Nvidia, and Intel debate: ASICs vs. GPUs vs. General-Purpose CPUs for RAN Baseband Processing
For Orange CTO Laurent Leboucher, the main attraction of AI today lies in its potential to improve the efficiency of 5G radio access networks (RANs). That helps explain Orange’s recent collaboration with Nokia and Nvidia. Orange already deploys Nokia’s purpose-built 5G network equipment and software at mobile sites in France and other markets. Until recently, it had little obvious need for Nvidia, the U.S. chip making king best known for the graphics processing units (GPUs) used to train large language models. But Nokia and Nvidia became closely aligned last October, when Nvidia took a 3% stake in Nokia as part of a $1 billion investment. Nokia is now developing AI RAN software designed to run on GPUs.
Leboucher’s interest is driven in part by concerns over the cost of custom silicon — the application-specific integrated circuits (ASICs) used in purpose-built 5G networks. “It creates an opportunity to bring a general-purpose chipset instead of an ASIC implementation,” he told Light Reading at last week’s FutureNet World event in London. “I think we could, at some point, benefit from the economies of scale of new chipsets. That could be Nvidia.”
The rationale is much easier to understand than arguments about 5G for autonomous vehicles. Chip manufacturing is already expensive, and both Nokia and Ericsson expect component costs to rise further this year amid relentless AI demand. At the same time, the RAN market remains relatively small and has contracted. According to market research firm Omdia, telco spending fell from $45 billion in 2022 to $35 billion last year and is expected to stay at that level. In that context, it is increasingly difficult to justify designing high-cost chips with limited reuse outside telecom.
Image Credit: Orange
Last year, Nvidia spent about $18.5 billion on research and development, generated nearly $216 billion in revenue, and reported a gross margin of more than 70%. Its financial strength is not in question. If telecom operators can use its GPUs for RAN software, they may face less pressure to secure the long-term economics of 5G and 6G development. That alone could be enough to support the case for Nvidia. The counterarguments are cost and power consumption. By design, custom silicon is optimized for a specific workload and will always outperform a more general-purpose processor at that task. An Nvidia GPU in the RAN could therefore be seen as excessive — like using a crop duster to water a hanging basket.
Leboucher, believes that Nokia and Nvidia are developing something far more compact than a typical data-center deployment. “It is not a Blackwell GPU,” he said, referring to Nvidia’s current hyperscaler-class product line. “I have an understanding it’s something which is a little bit smaller.” One of the first GPU-based products is expected to come on a card that Orange can insert into an existing Nokia AirScale chassis.
He is also interested in replacing traditional RAN algorithms with AI to improve spectral efficiency and overall performance. Through trials with Nokia and Nvidia, Orange wants to determine whether a GPU is actually required to capture the full benefit. “We can completely rethink the way we are doing algorithms today, using AI for the radio Layer 1,” he said, referring to the most compute-intensive part of the RAN software stack. Some of the “AI-RAN” narrative still sounds “a little bit like science fiction,” Leboucher admitted. “But I think there are some very interesting ideas behind that. We want to understand where we are.”
This is not the first time the industry has debated a shift from ASICs to general-purpose processors for RAN equipment. Alongside its purpose-built 5G portfolio, Ericsson already offers cloud RAN products based on Intel CPUs. Samsung is now focused on Intel-based virtual RAN and has recently predicted the end of purpose-built 5G. Even so, cloud and virtual RAN still account for only a small share of live 5G deployments. Huawei and Ericsson, the two largest RAN vendors, remain committed to custom silicon development.
Nvidia’s entry into the market has clearly given Leboucher and his team more to evaluate as RAN technology becomes more sophisticated. “We are introducing new requirements for radio networks, typically for beamforming, and we have to consider the need for quite powerful chipsets,” he said. “Whether the best way to keep going is using ASICs or a general-purpose architecture – I think this is a good time to ask the question. Before, it was too early.”
The answer could shape Orange’s next major RAN decisions. The operator is preparing for what Leboucher describes as a “refresh” of RAN equipment across several countries ahead of the expected 6G launch in 2030. For the first time, he said, Orange will include cloud RAN as a “major option” in its request for proposal.
The concern around Intel as an alternative to Nvidia is its still-fragile financial position. Before December, Intel had been trying to spin off its network and edge group (NEX), which develops RAN chips. Those plans were later shelved, but the company’s net loss widened to about $4.3 billion in the most recent first quarter, from $887 million a year earlier, while revenue rose only 7% year over year to $13.6 billion. Cristina Rodriguez, who had led NEX, left this month to join Coherent, and Intel has not yet named a successor. “The shares jumped 28% in after-hours trading, taking Intel firmly into meme-stock territory,” said Radio Free Mobile analyst Richard Windsor in a blog published after results came out on April 23. “I say meme-stock because there is no other way to describe it when the shares are on a 2026 PER [price-to-earnings ratio] of 137x, and its technology looks obsolete.”
Orange places significant value on separating hardware from software, allowing the same RAN software to run across multiple hardware platforms. Ericsson and Samsung both say the virtual RAN software they have built for Intel CPUs could, with relatively modest changes, be ported to AMD silicon using the same x86 architecture or to Arm-based CPUs.
By contrast, Layer 1 code written for Nvidia GPUs and the CUDA software stack would not be portable to other platforms, according to Ericsson. “I think the main challenge we see with that is we are trying very hard to keep our stack portable, to give hardware options,” Michael Begley, Ericsson’s head of RAN compute, told Light Reading at MWC Barcelona this year. “If you go all in on one, it’s great, but you’re all in on one, and you can’t offer those other options to the operators or the ecosystem.”
Leboucher acknowledges that risk. “The risk of lock-in exists, definitely,” he said. “We really want to stay open. At the same time, we know that benefiting from a very, very large-scale general-purpose architecture should improve the TCO [total cost of ownership]. At the end of the day, it will be a trade-off. But we would welcome an architecture where we have the capacity at some point to decide to swap if we need to swap.”
Nokia’s hope is that much of the Layer 1 software written for Nvidia GPUs will eventually be deployable on other GPU platforms. But Nvidia’s near-monopoly in that segment leaves the industry with few alternatives for now. There is also optimism inside Nokia that GPU-based code could later be adapted for capable CPUs, although Ericsson’s comments suggest that would be much harder. For telecom executives, the choices made over the next couple of years may be pivotal as 6G approaches.
………………………………………………………………………………………………………………………………………………………
References:
https://www.lightreading.com/5g/orange-weighs-nvidia-against-intel-for-5g-chips-ahead-of-new-rfp
RAN Silicon Rethink- Part II; vRAN and General-Purpose Compute
RAN silicon rethink – from purpose built products & ASICs to general purpose processors or GPUs for vRAN & AI RAN
Analysis: Nokia and Marvell partnership to develop 5G RAN silicon technology + other Nokia moves
Analysis: Nvidia’s $2 billion investment in Marvell; NVLink Fusion ecosystem & RAN vendor silicon strategy
Ericsson goes with custom silicon (rather than Nvidia GPUs) for AI RAN
Marvell shrinking share of the RAN custom silicon market & acquisition of XConn Technologies for AI data center connectivity
Custom AI Chips: Powering the next wave of Intelligent Computing
OpenAI and Broadcom in $10B deal to make custom AI chips
Will Google Cloud’s AI and data analytics revenue +TPU IP licensing income offset huge AI CAPEX to produce a decent ROI?
Big Tech AI spending binge results in massive job cuts!
Big Tech AI spending binge results in massive job cuts!
Executive Summary:
The tech industry is undergoing a massive structural realignment. Hyperscalers, Software as a Service (SaaS) vendors, and telecom network and equipment providers are aggressively slashing workforces to reallocate capital toward massive AI infrastructure investments. Alphabet, Meta, Amazon, and Microsoft are projected to spend a collective $674 billion in 2026—over double their 2024 levels. Most of that spending is AI related.
From the referenced WSJ article:
“Tech companies are in effect playing a game of chicken with each other on capital-spending plans. They are shelling out as much as they can—more than their rivals, they hope—on AI chips and data centers that could put them in the lead in a race they feel they can’t afford to lose. That in turn is heightening competition over who can use AI to help do more with a lot less, freeing up money to spend on expensive chips.”
Hyperscalers, such as Microsoft and Meta Platforms (Meta), are the latest to their significantly reduce their workforces to scale AI-driven operations. Meta is reportedly reducing its headcount by approximately 8,000, while Microsoft has initiated a “voluntary retirement program” (aka a buyout) targeting 7% of its U.S. workforce—a strategic move to trim payroll before resorting to involuntary layoffs.
This trend is industry-wide: Oracle and Snap have executed significant reductions, while Block announced plans to cut 40% of its staff (over 4,000 employees). March 2026 represented a two-year peak in tech industry contraction, with Layoffs.fyi reporting 45,800 tech job reductions.
The AI Transformation Narrative vs. Financial Reality:
Executive leadership is framing these cuts as a strategic pivot toward an AI-native future where automated workflows replace legacy human-centric processes. While CEOs like Block’s Jack Dorsey insist these decisions aren’t driven by distress, a “game of chicken” is unfolding in capital planning.
Companies are locked in an escalating race to secure AI silicon (GPUs), High Bandwidth Memory (HBM) and expand Data Center footprints, creating a massive drain on liquidity. This heightens the pressure to achieve “doing more with less”—using AI to automate internal functions and free up the capital necessary for expensive infrastructure. However, in many cases, these cuts are simply corrective measures for pandemic-era overhiring or efforts to normalize efficiency metrics:
- Oracle: Annual revenue per employee remains significantly below industry leaders like Microsoft.
- Snap: Headcount remains 65% above pre-COVID levels despite consistent operating losses.
Strategic Risks and “Off-Balance-Sheet” Engineering:
While slashing headcounts improves Revenue Per Employee (RPE)—a key KPI for Wall Street—it introduces significant long-term risks:
- Talent Attrition & Brain Drain: Aggressive layoffs degrade morale and may drive elite engineering talent toward startups, potentially creating new competitors.
- Governance & Safety: Reducing human oversight during AI deployment could lead to safety and business model integration failures.
- Regulatory & Public Backlash: The “AI as a job killer” narrative is fueling community opposition to massive data center builds, complicating infrastructure rollouts.
The CAPEX Burden:
The financial strain is becoming evident even for “Deep Pocket” firms. Alphabet, Meta, Amazon, and Microsoft are projected to spend $674 billion in CAPEX this year—more than double their 2022 spend.
- Amazon is projected to be cash-flow negative this year.
- Meta’s CAPEX is set to exceed 50% of its annual revenue, with its debt-to-equity ratio climbing to 39% (up from 8% five years ago).
- Some firms are reportedly utilizing “off-balance-sheet financial wizardry” to maintain their AI compute growth without alarming debt markets.
Verdict of the Market?
Markets are sending mixed signals. While analysts are obsessed with efficiency metrics (questions about efficiency on earnings calls have tripled in two years), they are becoming “skittish” regarding unbridled spending. Tesla (TSLA), for instance, saw a 4% stock dip after raising its spending target to $25 billion.
Ultimately, tech giants—who already average $2M in annual revenue per employee—are betting that further workforce reductions will juice efficiency and fund the AI arms race. The trade-off remains whether these “leaner” organizations can maintain the innovation and safety standards required to lead the next technological cycle.
The telecom sector is particularly vulnerable, as AI-native “zero-touch” operations begin to replace legacy roles permanently.
- Network Operators:BT has announced plans to replace up to 10,000 roles with AI by 2030, specifically targeting network management and customer service.
- Network Equipment Vendors: Equipment giants Ericsson and Nokia have collectively shed over 36,000 roles in recent years, pivoting from traditional hardware to AI-optimized software and networking.
- Integrators:Accenture and IBM are utilizing AI to automate junior-level coding and back-office HR tasks, signaling that AI reskilling is now a prerequisite for workforce retention.
Strategic Outlook – Monetization and the “RPE” Battle:
For both MNOs and tech giants, the coming years are about monetization. Investors have shifted from cheering bold AI visions to demanding tangible results, with a heavy focus on Revenue Per Employee (RPE)—a metric that workforce reductions are designed to “juice.”
That “Great Realignment” is a high-stakes gamble, in this author’s opinion. The firms that successfully bridge the gap between massive infrastructure investments and scalable, profitable AI-native services will lead the next generation of global technology. Those that fail to balance efficiency with talent retention may find themselves outpaced by leaner, AI-native startups born from the very talent they have released.
References:
https://www.wsj.com/tech/ai/the-ai-splurge-is-costing-big-tech-its-workforce-34a88e68
AI spending boom accelerates: Big tech to invest an aggregate of $400 billion in 2025; much more in 2026!
AI infrastructure spending boom: a path towards AGI or speculative bubble?
Gartner: AI spending >$2 trillion in 2026 driven by hyperscalers data center investments
AI spending is surging; companies accelerate AI adoption, but job cuts loom large
Big tech spending on AI data centers and infrastructure vs the fiber optic buildout during the dot-com boom (& bust)
Will billions of dollars big tech is spending on Gen AI data centers produce a decent ROI?
Canalys & Gartner: AI investments drive growth in cloud infrastructure spending
STL Partners webinar: Agentic AI needed for RAN autonomy & efficiency
Yesterday, a STL Partners webinar titled “Turning autonomy into margin: Agentic AI and the autonomous RAN,” suggested agentic AI is the missing layer that can turn RAN autonomy from a technical goal into a direct profit margin booster. It argues that operators should prioritize autonomy use cases by business impact, not just by how much automation coverage they add, and that the right roadmap can move autonomy from an engineering KPI to a commercial advantage.
The central message was that autonomy only matters if it improves economics (see poll results below). The webinar revealed that network operators need a dual-axis framework that combines the usual autonomous-network maturity view with a value-creation lens, so they can focus on the capabilities that scale into measurable business outcomes.
Agentic AI is presented as the practical enabler for moving beyond human-in-the-loop operations. In this framing, agents help orchestrate tasks, make decisions, and coordinate network actions in ways that support more closed-loop automation than traditional workflows can deliver.
The results of an “actuality” poll relating to RAN autonomy revealed that controlling costs and reliability were most important, with the enablement of new revenue growth through APIs and sensing only scoring 10.87% of respondents. Similarly, results for an “aspirations” poll for RAN autonomy were also fairly evenly spread between reducing costs and optimizing the customer experience, with just 13.21% citing new revenue growth.
Source: STL Partners
Terje Jensen, SVP, global business security officer and head of network and cloud technology strategy at Telenor, said that he had expected to see network operators’ aspirations shift more clearly towards improving customer experience and even revenue generation, not just efficiency.
Darwin Janz, strategic technology planner at SaskTel, also thought network operators’ ambitions would be higher, but he noted that they still struggle to identify concrete, monetizable use cases. Without that, there’s a real risk of building technical solutions in search of a problem, rather than starting from clear enterprise needs and value, Darwin noted. “We really need to see those use cases and enterprise customer needs,” he added.
……………………………………………………………………………………………………………………….
The webinar was built around four practical questions:
- Which use cases create real commercial impact?
- How to shift from autonomy as an engineering metric to a margin driver?
- Where agentic does AI add value today?
- What data, orchestration, and organizational foundations are needed to scale beyond pilots.
For network operators, the implication is that autonomous RAN strategy should be tied to P&L outcomes such as lower operating cost, better resource utilization, and faster optimization cycles. The webinar’s message is that autonomy becomes strategically important only when it is deployed in a way that compounds across the network and business.
…………………………………………………………………………………………………………………..
References:
The Financial Trap of Autonomous Networks: Scaling Agentic AI in the Telecom Core
Nokia to showcase agentic AI network slicing; Ericsson partners with Ookla to measure 5G network slicing performance
T-Mobile US announces new broadband wireless and fiber targets, 5G-A with agentic AI and live voice call translation
Telecom operators investing in Agentic AI while Self Organizing Network AI market set for rapid growth
Nokia’s AI Applications Study: “Physical AI” may require RAN redesign to support high‑volume, low‑latency uplink traffic
According to Nokia, AI-generated traffic in most mobile networks is at an early stage, with application maturity and adoption by consumers and enterprises only at the start of a broader AI super cycle. The Finland based company analyzed more than 50 AI applications and came to three conclusions: higher uplink traffic, overall data growth and increasing sensitivity to delay in conversational services such as chat and voice. Also, the mobile network industry is moving toward “AI-RAN” or “6G-native” structures that embed AI into the network, transforming radio sites into “robotic” nodes capable of edge inference and handling these new demands.
–>Do those findings require a structural change in Radio Access Network (RAN) design? Let’s take a fresh look…..
Mobile networks traditionally support a heterogeneous mix of traffic, ranging from high-throughput video streaming to low-bandwidth, delay-tolerant messaging. Network operators typically address escalating capacity demands through infrastructure expansion and overprovisioning, relying on best-effort delivery—a model that has proven remarkably resilient. However, capacity alone is insufficient for new use cases.
The transition from circuit-switched voice to packet-switched (voice/video/data) IP traffic requires a redesign to accommodate variable packet sizes instead of predictable, continuous voice patterns. The proliferation of Internet of Things (IoT) devices introduced requirements for massive machine-type communications (mMTC), driving the development of LTE-M and NB-IoT to optimize for deep indoor penetration and power efficiency. Conversely, consumer web-based services and video streaming scale seamlessly by adding RAN and core capacity. Existing AI applications, such as generative AI chatbots, follow this model, making current RAN architectures adequate for the present load.
A paradigm shift is emerging with Physical AI [1.], which enables machines like autonomous vehicles and robots to interact with the environment in real time. Unlike traditional video streaming, these applications cannot leverage buffering to absorb network jitter. In Physical AI, high-definition video frames and sensor data must arrive within stringent time-to-live (TTL) constraints to remain actionable. This shifts the focus from average throughput to consistent low latency. Maintaining this strict QoS, particularly in the uplink, requires abandoning best-effort, overprovisioned models in favor of guaranteed scheduling, which necessitates substantial reserved capacity or specialized AI-RAN functionalities.
Note 1. Physical AI combines sensors, perception, decision-making, and actuators so machines can understand their environment and take physical (real world) action. Physical AI is used by robots, vehicles, drones, industrial machines, and smart infrastructure that generate and consume real-time sensor, video, and control traffic. These systems need tight coupling between low latency, high reliability, and continuous feedback loops because decisions in software immediately affect physical motion or control. Physical AI is different from typical generative AI because the output is not text or images; it is real-world action. That makes network performance critical, especially for uplink-heavy, latency-sensitive traffic where delays can affect safety, control accuracy, and operational efficiency.
“Physical AI introduces the possibility that large-volume uplink video with strict latency requirements. It will become a meaningful part of mobile traffic, creating both a design challenge and a monetization opportunity,” says Harish Viswanathan, Head of the Radio Systems Research Group at Nokia.
Image Credit: Techslang
Delivering uplink video with sub‑20 ms end-to-end latency can require provisioning three to four times the average uplink capacity. While this level of redundancy is manageable for low-bandwidth services such as voice or control signaling, it becomes prohibitively expensive when supporting high-throughput video streams.
As device densities increase, the required headroom for reserved capacity grows disproportionately, significantly constraining network scalability and driving up cost per bit. This makes Physical AI traffic—characterized by real-time sensor and video inputs for machine analysis—fundamentally different from conventional services, and unsuited to existing best‑effort transport models. From a Nokia blog post:
“Physical AI will rely on low latency videos to enable real-time control. While the machines or robots will perform most functions locally, there will be situations where they need to rely on more powerful models or human operators to provide remote control via the network. For example, driverless taxis may require remote assistance in unexpected scenarios; service robots may need guidance in complex environments; drones may depend on real‑time video analysis at the point of delivery; and field workers using AR may require timely visual instructions. In all these cases, the network must deliver fresh video information with low and predictable latency.”
To address these challenges, telecom operators are expected to adopt a multi‑layer approach encompassing network architecture, traffic management, and service monetization.
At the Application layer, not all traffic requires identical latency treatment. When video or sensor data is processed by AI rather than consumed by humans, only semantically relevant information may need immediate uplink transmission. This emerging paradigm, known as semantic communication, allows for significant data reduction while preserving information integrity within latency‑critical loops.
Within the network domain, established mechanisms such as Quality of Service (QoS) and network slicing remain essential. QoS enables prioritization of specific traffic classes, while slicing supports logically isolated virtual networks with guaranteed service-level attributes—latency, jitter, bandwidth, and reliability.
At the service and business model level, supporting low-latency, bandwidth-intensive applications reshapes network economics. Operators must evolve beyond best‑effort pricing structures toward differentiated service tiers or performance-based charging models aligned with enterprise and industrial use cases.
For the RAN, Physical AI underscores the need for greater programmability and elasticity. Future RAN designs will depend on dynamic resource allocation, real-time traffic classification, and AI-driven orchestration to balance throughput, latency, and reliability at scale.
As Physical AI deployments expand—from autonomous mobility to precision manufacturing and tele‑robotics—managing high‑volume, low‑latency uplink traffic will become a defining capability for next‑generation network strategy and differentiation. Unlike conventional mobile data, Physical AI cannot rely on buffering to manage traffic spikes. The requirement for continuous video and sensor data to arrive within strict time limits to inform real-time actions makes traditional “best-effort” network approaches inefficient and costly.
- Uplink-Centric Demand: Physical AI shifts the network requirement from downlink-heavy (human consumption) to uplink-heavy (machine-generated) traffic.
- Strict Latency & Throughput: Maintaining consistent low latency (e.g., around 20 milliseconds) for high-volume video uploads can require 3x to 4x more capacity than average, making overprovisioning unsustainable.
- Need for Programmable Architectures: To support this, RAN must move toward more flexible, AI-native architectures that prioritize critical data and provide deterministic, rather than best-effort, performance.
- Semantic Communication: To reduce data volume while maintaining performance, the RAN will need to adopt semantic communication—transmitting only the essential data needed for the AI to make decisions.
………………………………………………………………………………………………………………………………………………………..
References:
https://www.nokia.com/asset/215147/
https://www.nokia.com/blog/physical-ai-redefining-ran-and-telco-monetization/
https://telcomagazine.com/news/nokia-report-points-to-ai-driven-shift-in-mobile-traffic
Arm Holdings unveils “Physical AI” business unit to focus on robotics and automotive
Is the “far edge” a bridge to far to cross for AI inferencing? What about “Distributed AI Grids”?
The Financial Trap of Autonomous Networks: Scaling Agentic AI in the Telecom Core
Ericsson and Intel collaborate to accelerate AI-Native 6G; other AI-Native 6G advancements at MWC 2026
NVIDIA and global telecom leaders to build 6G on open and secure AI-native platforms + Linux Foundation launches OCUDU
Comparing AI Native mode in 6G (IMT 2030) vs AI Overlay/Add-On status in 5G (IMT 2020)
AI-RAN Reality Check: hype vs hesitation, shaky business case, no specific definition, no standards?
Is the “far edge” a bridge to far to cross for AI inferencing? What about “Distributed AI Grids”?
How Far is the Far Edge?
As major telcos size up distributed edge sites for a possible AI inferencing model, they’re trying to determine how far out the right place is in their networks to invest in AI computing capacity. According to Light Reading, the “far edge” is a divisive option for inferencing. According to Omdia, owned by Informa, the Far edge includes: radio access network (RAN) cell sites, aggregation hubs, exchange offices, optical line terminal (OLT) nodes, and Tier 2 metro hubs.
Many telcos are struggling to define how far is the edge from customer premises and how to serve various use cases with compute and intelligence? It seems that 5G SA core with network slicing would be mandatory to support multiple unique use cases, each with different QoS requirements.
According to Omdia’s Telco Edge Computing Survey last year, just 15% of telcos ranked network far edge as the top location for where most AI inferencing will take place, while even less (11%) said the network near edge would be the main spot (which includes central offices, headend sites and large telco data centers). The results showed AI inferencing is expected to be handled mostly on the end devices themselves and at the enterprise edge (e.g., offices, campus or manufacturing sites).
Kerem Arsal, Omdia senior principal analyst for telco enterprise and whoIe sale, predicted in a research note that this year will see telcos split into camps of “believers” and “doubters” of the far edge.
Image Credit: Sphere
…………………………………………………………………………………………………………………………………………………………………………………………………………………..
AT&T VP Yigal Elbaz, speaking at the recent New Street Research and BCG Global Connectivity Leaders Conference, expressed a cautious view on AI compute at the “far edge,” questioning how far the edge truly needs to extend to serve specific use cases effectively. He said the following (Source: Light Reading)
“The proliferation of compute and high-performing compute across the nation, in all metros is just happening, with a software layer on top of this [and] with the tools that developers need. So, I am not sure that there’s much value in extending that compute all the way to the far edge just to save another millisecond or two milliseconds of latency.”
“AT&T’s fiber and wireless networks can provide the “deterministic experience” needed between any new use cases and help them to “intelligently connect to the right model that they use, the context or the infrastructure that they need because that’s going to be heavily distributed across the US.”
“There’s no doubt that that AI is going to be embedded into wireless networks, and we’re going to call it AI-native and combine the physical space with the intelligence of the network. This is all true,” said Elbaz.
………………………………………………………………………………………………………………………………………………………………………………………………………………………..
Distributed AI Grids:
- Ethernet with RDMA (RoCE): The foundation is built on Nvidia Spectrum-X Ethernet, which utilizes RDMA over Converged Ethernet (RoCE). This allows for direct memory access between edge GPUs (e.g., Nvidia RTX PRO 6000 Blackwell Server Edition) and the network core, bypassing CPU overhead to achieve near-line-rate performance.
- Scale-Across Networking: Using Nvidia Spectrum-XGS, the architecture extends standard RoCE to scale across geographically distributed sites. This creates a unified “AI Factory Grid” where remote edge nodes function as a single, programmable compute substrate.
- Silicon One Routing: Cisco’s Silicon One-based routing is utilized for AI-optimized traffic management, providing the high-speed, high-density throughput required for token-intensive inference workloads.
- Zero Trust & Secure Pathways: The interconnect includes a Zero Trust security layer embedded directly into the fabric. It utilizes localized traffic breakout and policy-enforced pathways to ensure that sensitive IoT and video data (such as public safety feeds) remain within the customer’s secure domain at the network edge.
- Orchestration Control Plane: A workload-aware control plane manages these protocols to intelligently route tasks based on real-time KPIs (latency, cost-per-token, and data sovereignty), ensuring that “mission-critical” inference happens at the optimal node.
- Proprietary Software Lock-in: Integrating network functions into a proprietary ecosystem (like Nvidia’s CUDA or AI Aerial) can create a “subscription trap,” where software is inseparable from specific hardware, making it nearly impossible to swap vendors without a total architectural overhaul.
- Data Fragmentation: Deploying AI across a distributed grid often leads to fragmented data sets across legacy and new multi-vendor platforms, which can result in inaccurate AI models and increased operational complexity.
- Standardization Lag: While industry bodies like the GSMA are pushing for Open Telco AI standards, the rapid deployment of proprietary AI systems often outpaces these frameworks, leading to entrenched, incompatible systems that require significantly more resources to reconcile later.
- Integration with Legacy Systems: Modern “agentic AI” and AI-native stacks often struggle to orchestrate processes across siloed legacy infrastructure, creating rigid operational environments that prevent the seamless flow of data needed for automated network troubleshooting.
Bottom Line: While the AI Grid may offer a more viable roadmap than AI-RAN, there is insufficient industry discourse regarding the strategic risks of a global, geographically distributed computing platform—as Nvidia defines it—reliant on a single-vendor hardware stack. Although Nvidia currently maintains undisputed market dominance, historical precedents such as Intel serve as a cautionary tale; long-term dominance is never guaranteed, and even market leaders face potential obsolescence. Furthermore, Nvidia’s practice of providing capital injections to entities that subsequently re-invest those funds back into Nvidia’s own ecosystem raises significant concerns regarding market sustainability and long-term financial health.
References:
https://www.lightreading.com/ai-machine-learning/at-t-cto-casts-doubt-on-ai-compute-at-the-far-edge
https://www.lightreading.com/5g/nvidia-lines-up-ai-grid-as-orange-cto-echoes-the-ai-ran-doubts
Edge AI Computing Explained: Key Concepts and Industry Use Cases
Will “AI at the Edge” transform telecom or be yet another telco monetization failure?
Analysis: Edge AI and Qualcomm’s AI Program for Innovators 2026 – APAC for startups to lead in AI innovation
Private 5G networks move to include automation, autonomous systems, edge computing & AI operations
Nvidia AI-RAN survey results; AI inferencing as a reinvention of edge computing?
Nvidia’s networking solutions give it an edge over competitive AI chip makers
Using AI, DeepSig Advances Open, Intelligent Baseband RAN Architectures
Using advanced AI techniques, DeepSig has reportedly managed to eliminate a mobile network’s pilot signal, thereby removing signaling overhead without degrading overall performance. Founded in 2016, the U.S.-based startup occupies a leading position at the intersection of artificial intelligence (AI) and the radio access network (RAN), developing data-driven models that could supplant traditional, human-engineered signal processing algorithms.
This work has become especially relevant as the telecom industry moves toward open and software-defined RAN architectures. DeepSig is now a visible contributor to OCUDU (Open Centralized Unit Distributed Unit), an open-source initiative announced by the Linux Foundation in collaboration with the U.S. Department of Defense and its FutureG ecosystem partners to accelerate open CU/DU development for 5G and early 6G systems. OCUDU is intended to establish a carrier-grade reference platform for baseband software, with support for AI-based algorithms and solutions embedded in the RAN compute stack.
As AI becomes a central theme across the telecom ecosystem, DeepSig has rapidly moved from relative obscurity to prominence through collaborations with major industry and government stakeholders. Most recently, the company emerged as a key contributor to OCUDU—the Open Central Unit Distributed Unit initiative announced by the Linux Foundation and the U.S. Department of Defense (DoD) ahead of MWC Barcelona 2026. The program’s goal is to introduce open-source software elements into the RAN baseband domain, an area historically dominated by proprietary offerings from Ericsson, Nokia, and Samsung. By lowering barriers to entry, OCUDU aims to foster innovation and enable smaller players like DeepSig to participate more freely in the U.S. baseband ecosystem.
Image Credit: DeepSig
DeepSig was identified, alongside Ireland-based Software Radio Systems (SRS), as one of two startups selected to deliver OCUDU’s initial software stack. “The National Spectrum Consortium had an RFQ for developing an open-source stack,” explained Jim Shea, DeepSig’s CEO. “SRS already had a capable baseline, but it needed to be elevated to carrier-grade—adding new features and strengthening reliability,” he added.
Meanwhile, major vendors Ericsson and Nokia were named “premier members” of the new OCUDU Ecosystem Foundation. While both could, in principle, leverage the platform to integrate third-party components into their baseband systems, industry observers remain skeptical that these incumbents will fully embrace open-source alternatives over their established proprietary stacks. In comments at MWC, Nokia CEO Justin Hotard characterized OCUDU as a welcome ecosystem evolution to accelerate innovation but clarified that “not everything necessarily needs to be open source.”
Driven in part by DoD interests, OCUDU reflects broader U.S. government ambitions to ensure that 5G and future 6G networks remain open to domestic innovation, particularly for defense and mission-critical use cases. For vendors like Ericsson and Nokia—who view defense markets as increasingly strategic—this alignment could bring both opportunity and complexity.
DeepSig’s trajectory extends beyond OCUDU. The company’s technology originated from research by Tim O’Shea, now CTO, during his tenure at Virginia Tech, where he explored deep learning’s application to wireless signal processing. “You can apply deep learning to enhance the way communication systems operate by replacing many of the traditional algorithms,” said Jim Shea. While these methods do not circumvent theoretical limits such as Shannon’s Law, small efficiency gains can yield substantial operational and economic benefits for cost-sensitive mobile operators.
As DeepSig and peers continue to redefine how intelligence is integrated into the RAN, their work signals a shift toward AI-native architectures—where machine learning, rather than handcrafted algorithms, becomes the foundation for next-generation network optimization.
References:
https://www.lightreading.com/5g/small-deepsig-is-at-heart-of-ai-ran-challenge-to-ericsson-nokia
Accelerating 5G vRAN, AI-RAN, and 6G on OCUDU, “the Linux of RAN”
AI-RAN Reality Check: hype vs hesitation, shaky business case, no specific definition, no standards?
Ericsson goes with custom silicon (rather than Nvidia GPUs) for AI RAN
Dell’Oro: RAN Market Stabilized in 2025 with 1% CAG forecast over next 5 years; Opinion on AI RAN, 5G Advanced, 6G RAN/Core risks
Dell’Oro: Analysis of the Nokia-NVIDIA-partnership on AI RAN
RAN silicon rethink – from purpose built products & ASICs to general purpose processors or GPUs for vRAN & AI RAN
Dell’Oro: AI RAN to account for 1/3 of RAN market by 2029; AI RAN Alliance membership increases but few telcos have joined
Dell’Oro: AI RAN to account for 1/3 of RAN market by 2029; AI RAN Alliance membership increases but few telcos have joined
InterDigital led consortium to advance wireless spectrum coexistence & sharing
Telecom sessions at Nvidia’s 2025 AI developers GTC: March 17–21 in San Jose, CA
Sources: AI is Getting Smarter, but Hallucinations Are Getting Worse
AI-RAN Reality Check: hype vs hesitation, shaky business case, no specific definition, no standards?
Introduction:
The narrative surrounding “AI-RAN” — a term thrust into the spotlight by Nvidia — may have left many believing that boatloads of GPUs are already powering baseband compute in RAN equipment across the world’s seven million mobile sites. In truth, the reality is far more nascent.
Among major RAN vendors, Nokia stands alone in adapting baseband software for GPU acceleration. Yet even Nokia does not anticipate commercial readiness until late 2026, as confirmed by its Chief Technology Officer, Pallavi Mahajan, during the company’s MWC press conference earlier this year. For now, no operator has announced a commercial deployment — despite the buzz around trials.
Early Movers, Limited Momentum:
Much of the current AI-RAN activity centers on two operators: T-Mobile US and Japan’s SoftBank. At MWC, T-Mobile’s Executive Vice President of Innovation and ex-CTO, John Saw, acknowledged the limited availability of deployable solutions, quipping that he hoped Nokia would deliver an AI-RAN product within the year. Nokia CEO Justin Hotard quickly assured him that such a milestone was indeed on track.
Still, the debut of a GPU-based RAN stack does not imply an imminent large-scale rollout. Without tangible network performance or cost advantages over existing virtualized or disaggregated RAN approaches, operators are unlikely to move past controlled trials.
SoftBank, while often positioned as an AI-RAN pioneer, remains cautious. As Ryuji Wakikawa, Vice President of its Advanced Technology Division, outlined last year, the operator aims to deploy only a handful of AI-RAN sites over the next fiscal cycle. Transitioning from testing to carrying live commercial traffic, he emphasized, demands a significant maturity leap in quality and feature completeness.
Beyond Hype: Limited Commercial Engagement:
Elsewhere, Indonesia’s Indosat Ooredoo Hutchison (IOH) was heralded in 2025 as the first operator in Southeast Asia pursuing AI-RAN. More than a year later, authoritative sources indicate IOH’s work remains confined to its research facility in Surabaya, with no near-term plans for GPU investment at cell sites until measurable value is demonstrated.
The challenge for Nokia — and for GPU-backed AI-RAN broadly — is convincing operators that general-purpose accelerators offer sufficient performance or efficiency gains for most RAN workloads. T-Mobile and SoftBank continue evaluating both Nokia and Ericsson, whose AI-RAN philosophies diverge sharply. Nokia is developing GPU-based baseband software, while Ericsson maintains its focus on custom silicon and CPU architectures.
Divergent Architectures and Use Cases:
Ericsson contends that no core RAN performance enhancements intrinsically require GPUs. Its ongoing collaboration with Nvidia leverages the latter’s Grace CPU technology rather than its GPU portfolio, reserving GPU acceleration only for compute-intensive functions like forward error correction (FEC).
If Ericsson’s premise holds, GPUs in the RAN become justifiable only when supporting AI inference workloads. Even then, inference at every radio site remains improbable. A more incremental strategy — deploying GPUs selectively at edge locations where AI workloads justify their cost — may prove more practical.
This modular approach aligns with existing virtual RAN deployments based on Intel CPUs, which already include native FEC acceleration. “It is an off-the-shelf card that you can slide right into an HPE or Dell or Supermicro server,” said Alok Shah, the vice president of network strategy for Samsung Networks. “That gets you the edge functionality you are looking for.”
Rethinking the Economic Case for AI RAN:
Initially, Nvidia positioned GPUs for AI-RAN as viable only if broadly utilized for AI inference across the RAN. Following its strategic alignment with Nokia, however, the company has softened its stance — now suggesting that appropriately sized, power-efficient GPUs could make sense even when dedicated solely to baseband computation.
For now, the global RAN landscape remains far from GPU-saturated. AI-RAN remains an exploratory frontier — one testing not only the technical feasibility of GPUs at the edge, but also the economic/business case rationale for re-architecting a trillion-dollar telecom infrastructure around them.
The AI models suitable for RAN environments must be compact and efficient, far slimmer than those that drive data center-scale AI. There’s no room for the massive, parameter-heavy neural networks that justify a GPU’s cost or energy appetite. In that light, a GPU looks less like a breakthrough and more like a mismatch — a chainsaw brought to a task better handled with a sharp pair of scissors.
Evaluating the Case for AI-RAN Acceleration:
The central question is whether GPUs can deliver meaningful benefits over custom silicon or conventional CPUs for RAN compute. Ericsson’s engineers argue that AI models deployed at the RAN must remain relatively lightweight, with far fewer parameters than those used in large-scale data centers. Excessive model complexity could introduce signaling delays unacceptable in real-time radio environments. In this context, deploying a GPU for such workloads might seem disproportionate — a high-powered tool for a low-demand task.
The most compelling defense of GPU-based RAN acceleration came from Ronnie Vasishta, Nvidia’s Senior Vice President for Telecom, who told Light Reading last summer, “The world is developing on Nvidia.” His point underscores a shift in semiconductor economics: the cost and risk of building dedicated silicon for a mature and shrinking RAN market make general-purpose processors — supported by large-volume ecosystems — increasingly attractive alternatives.
Intel’s difficulties further illustrate this dynamic. Despite $53 billion in revenue during 2025, the former microprocessor king barely broke even despite $53 billion in revenue, following a $19 billion loss the previous year. A major restructuring cut its headcount by nearly 24,000, and its planned spinoff of the Network and Edge division — serving telecom infrastructure customers — was ultimately abandoned in December. Nvidia, the world’s most valuable company, may be eager to step into that space — but the economic logic seems upside down. Wireless network operators are looking to reduce costs, not import data center economics into the RAN.
Ecosystem or Echo Chamber?
Nvidia’s Aerial platform and CUDA-based software ecosystem do present a compelling story: open infrastructure, modular APIs, and space for smaller developers to innovate alongside giants like Nokia. On paper, it’s an alluring image of democratized RAN software. In practice, it ties the industry even more tightly to a vertically integrated, proprietary ecosystem.
Nokia appears comfortable with that trade-off. Nokia CTO Pallavi Mahajan’s recent blog post framed AI-RAN as a vehicle for “software speed and innovation.” He added, “Nokia’s AI-RAN initiative begins with a simple observation: AI is changing not only how networks are operated, but also the nature of the traffic they carry. AI workloads have already reached massive scale, with mobile devices accounting for more than half of AI interactions. Large language model interactions introduce richer uplink flows and burstier patterns as devices continuously send context to models.”
Indeed, that me be true someday. But for now, most wireless network operators need stable, cost-efficient networks, not AI-driven complexity or GPU-level power draw.
Image Credit: Nokia
Conclusions:
The uncomfortable truth is that AI-RAN feels more like a vendor-driven experiment than an operator-driven demand. Until someone proves that GPUs in the RAN deliver a measurable payoff — in performance, cost, or operational simplicity — the whole concept risks joining the long list of telecom “game-changers” that never made it past the trial stage. The hype cycle is predictable; the economics are not. Unless that equation changes, the real intelligence may be knowing when not to deploy AI RAN.
………………………………………………………………………………………………………………
In a Substack post today, Sebastian Barros writes: What Does AI-RAN Even Mean?
Despite the crazy hype, there is no definition for AI-RAN. Today it is at best a vibe, a dangerous reality for an industry that moves on strict standards that are currently completely absent.
The AI RAN hype is crazy right now. But despite the endless talk and vendor announcements, there is no actual technical definition of what it even means. As wild as it sounds for an industry built on strict 3GPP and O-RAN standards (those are specs- not standards), AI RAN is currently just a vendor interpretation designed to move hardware. Moreover, telecom has been using AI in the RAN before it was even cool. In fact, we were among the first industries to use neural networks in signal processing back in the 80s.
The problem is that treating AI-RAN as a marketing narrative rather than a rigid standard actively stalls progress. When the definition of AI-RAN is as different as night and day depending on which OEM you ask, it becomes impossible for any Telco to accurately model TCO or make solid CAPEX decisions.
Editor Notes:
- ITU-R’s IMT-2030 framework (ITU-R Recommendation M.2160-0 for IMT-2030) calls for an AI-native new air interface and AI-enhanced radio networks, but does not mention Nokia’s AI RAN.
- 3GPP Release 18 and later have study/work items on AI/ML for RAN functions such as energy saving, load balancing, mobility optimization, and AI/ML on the RAN air interface, but again no specifics have been discussed let alone agreed upon.
- 3GPP Release 19 continues and expands this work, with reporting that it builds on Release 18’s normative work and adds new AI/ML-based use cases for NG-RAN. In other words, 3GPP does have AI-RAN-related specs in progress and some normative features, but they are distributed across multiple RAN work items rather than packaged as one standalone “AI RAN” specification.
- AI RAN Alliance “is dedicated to driving the enhancement of RAN performance and capability with AI.” However, they’ve not yet produced any implementable specifications for AI RAN. Yet there are only four carriers that are “executive members“: Vodafone, T-Mobile, and SK Telecom, and Softbank (which is a conglomerate).
In Japan, NTT Docomo holds the largest cellular market share, with KDDI second, followed by SoftBank and the rapidly expanding Rakuten Mobile.
References:
https://www.lightreading.com/5g/ai-ran-lots-of-talk-little-action-no-guarantees
https://www.nokia.com/blog/ai-ran-bringing-software-speed-innovation-into-the-radio-network/
Ericsson goes with custom silicon (rather than Nvidia GPUs) for AI RAN
Dell’Oro: RAN Market Stabilized in 2025 with 1% CAG forecast over next 5 years; Opinion on AI RAN, 5G Advanced, 6G RAN/Core risks
Dell’Oro: Analysis of the Nokia-NVIDIA-partnership on AI RAN
RAN silicon rethink – from purpose built products & ASICs to general purpose processors or GPUs for vRAN & AI RAN
Dell’Oro: AI RAN to account for 1/3 of RAN market by 2029; AI RAN Alliance membership increases but few telcos have joined
Dell’Oro: AI RAN to account for 1/3 of RAN market by 2029; AI RAN Alliance membership increases but few telcos have joined
