U.S. Fixed Broadband Market Outlook: Satellite, FWA, Fiber and Cable Face Changing Competitive Dynamics

Disclaimer: ChatGPT was used extensively to assist with research, technical and market analysis, and drafting of this article. The author independently conducted additional research and incorporated source material and quotations. He reviewed, edited, and fact-checked the final content and is 100% responsible for its accuracy.

Introduction:

The U.S. fixed broadband market is entering a period of significant competitive change as low-Earth-orbit (LEO) satellite broadband expands, cellular fixed wireless access (FWA) growth moderates, fiber-to-the-premises (FTTP) deployment continues, and cable operators face increasing subscriber losses. These developments are changing the competitive relationships among four principal broadband access technologies: satellite, cellular FWA, fiber and hybrid fiber-coaxial (HFC) cable.

A November 19, 2026 report by New Street Research, summarized in Light Reading’s article, Satellite Broadband to Capture Most Market Share from FWA – Report, provides an interesting assessment of these trends. The analysts expect satellite broadband to expand the addressable broadband market while capturing subscribers from existing providers, with cellular FWA experiencing the greatest competitive impact.  “Satellite can do things wireline hasn’t been able to, and the take-up demonstrates there is latent demand out there we haven’t incorporated previously,” states the report authored by New Street Research analysts David Barden, Vikash Harlalka and Ryan Smith.

The report also raises a broader question: as fiber’s traditional source of subscriber growth diminishes and satellite capacity expands, will satellite broadband eventually take a meaningful share of the fiber market as well?

Satellite Broadband: Starlink Drives Market Expansion

LEO satellite broadband, led by SpaceX’s Starlink, is becoming an increasingly significant participant in the U.S. residential broadband market. Unlike traditional geostationary satellite systems, LEO constellations operate much closer to Earth, substantially reducing propagation delay and enabling more competitive broadband services.

New Street Research forecasts that Starlink’s U.S. subscriber base will increase from approximately 3 million at year-end 2025 to approximately 10 million by 2030. Its analysis also forecasts that satellite broadband will capture approximately 7% of the existing broadband market in its terminal state.  “Satellite net adds have accelerated in recent quarters as Starlink has become a more viable product. We expect Satellite net adds to remain at current levels for the foreseeable future,”  the analysts wrote.

The analysts expect satellite net additions to remain around current levels for the foreseeable future. They also believe satellite is expanding the overall addressable market by serving customers whose broadband requirements have not been adequately met by terrestrial providers.

New Street estimates that U.S. residential broadband subscriber growth reached 2.2% year-over-year in Q3 2026. Without satellite broadband additions, growth would have been only 1.6%.

These figures demonstrate that satellite is contributing to overall broadband market expansion, not simply redistributing existing subscribers among competing providers.

Nevertheless, satellite broadband is also a competitive substitute for terrestrial fixed broadband. Starlink can attract customers from cable, cellular FWA and, potentially, fiber providers, depending on local availability, service performance and pricing.

Its competitive position will depend heavily on continued constellation expansion, satellite throughput, spectrum availability, ground infrastructure, terminal costs and the economics of serving increasingly dense geographic markets.

Cellular FWA: Growth Continues but Subscriber Additions May Slow:

Cellular FWA has become an important alternative to wireline broadband, particularly in locations where fiber is unavailable or where customers prefer lower-cost service and simpler installation.  T-Mobile, Verizon and AT&T have been the principal U.S. cellular FWA providers. Their services use 4G LTE and 5G radio access networks to deliver broadband connectivity to fixed residential locations.

According to New Street Research, combined FWA net additions have remained in the 3.6–3.9 million range annually over the past four years. However, the analysts report that additions are beginning to slow at T-Mobile and Verizon, with AT&T also believed to have reached peak net additions.  New Street expects FWA subscriber growth to decelerate further for all three operators.

“FWA net adds have been in the 3.6-3.9MM range for the past 4 years. Both T-Mobile and Verizon adds have started to slow down with the latter showing this more acutely. AT&T continues to grow but we estimate it has reached peak net adds. We expect FWA net adds to continue slowing down for all 3 carriers from here,” states the report.

The belief that FWA is slowing runs counter to the narrative coming from some executives: “…on FWA, look, we’re seeing continued very strong momentum. So deceleration, et cetera, is not where we are,” T-Mobile CEO Srini Gopalan told analysts on the operator’s Q3 earnings call.

This forecast is significant because cellular FWA has benefited from relatively rapid deployment, attractive introductory pricing and the ability to use existing cellular network infrastructure. However, FWA capacity is constrained by available spectrum, radio resources, cell-site density and competing mobile traffic.  As subscriber penetration increases, operators must balance fixed broadband demand against mobile network performance. Additional spectrum, network densification and improved radio efficiency can expand capacity, but they require further investment.

Satellite broadband could therefore become a stronger competitor to cellular FWA, particularly in rural and lower-density markets where terrestrial cellular capacity or coverage is constrained.

It remains important to distinguish subscriber additions from market share. Slower FWA net additions do not necessarily mean that FWA subscribers will migrate directly to satellite; some customers may instead choose fiber or cable, while satellite may also attract previously unserved households.

The evolving U.S. fixed broadband landscape: LEO satellite, cellular fixed wireless access (FWA), fiber-to-the-premises (FTTP), and hybrid fiber-coaxial (HFC) networks compete to deliver residential broadband connectivity.  Image generated by ChatGPT.

Fiber Broadband: Strong Growth Through 2028, Followed by Deceleration:

Fiber broadband remains the strongest long-term fixed broadband access technology in terms of capacity, scalability, symmetrical bandwidth potential and low latency.

U.S. fiber deployments continue to expand, with telecommunications operators, independent fiber providers and municipal or cooperative networks extending FTTP coverage.

New Street Research expects fiber net additions to reach approximately 3 million in 2026, followed by 2.8–2.9 million annually in 2027 and 2028.

However, the analysts expect fiber net additions to decelerate thereafter as the pool of remaining DSL subscribers available for conversion approaches exhaustion.

They forecast that DSL subscriptions will be nearly zero by 2029. This removes an important source of relatively straightforward fiber subscriber growth: migrating customers from an operator’s legacy copper network to its own fiber infrastructure.  As DSL converts to fiber decrease, the analysts also expect fiber net adds to reach a peak of 3 million in 2026 and remain in the 2.8 million-2.9 million range over the next two years “before decelerating thereafter,” according to the report.  “We expect a deceleration in fiber net adds because DSL subs will be nearly zero by 2029 and an easy source of fiber net adds will disappear. Fiber operators will need to fight harder to find new subscribers,” the report said.

New Street says its long-term outlook does not expect satellite “to gain any share at the expense of fiber.”  Consequently, fiber operators will increasingly need to compete for subscribers already served by cable, cellular FWA and other broadband providers.  This represents an important change in the economics of fiber deployment and customer acquisition. Fiber providers will need to demonstrate sufficient advantages in bandwidth, reliability, latency and pricing to persuade customers to switch from existing broadband services.

Will Satellite Broadband  Take Market Share From Fiber based Access?

New Street Research does not currently expect satellite broadband to gain market share at fiber’s expense. Instead, its forecast assigns the greatest competitive impact to cellular FWA.

That is a reasonable base-case assumption, particularly because fiber generally provides substantially higher capacity, lower latency, more consistent performance and better economics per delivered bit in densely populated markets.

However, it would be premature to conclude that satellite will not take market share from fiber.

LEO satellite broadband is not limited to customers who lack terrestrial broadband access. Customers in some suburban and exurban locations may consider satellite a viable alternative even when fiber is available, depending on service pricing, performance, installation costs and individual broadband requirements.

Furthermore, improvements in satellite capacity, spectral efficiency and terminal economics could expand the number of locations where satellite can compete effectively.

The geographic distribution of fiber and satellite capacity will be particularly important. Satellite can provide meaningful competition in lower-density markets where fiber deployment and maintenance are expensive, even if its nationwide share of broadband subscribers remains relatively modest.

The distinction is between aggregate market-share forecasts and actual competitive substitution in individual markets. Satellite can take customers from fiber in selected geographic markets without producing a significant change in nationwide fiber market share.

Accordingly, New Street’s forecast that satellite will primarily take share from FWA should be viewed as a projection rather than an established competitive outcome.

Cable Broadband: DOCSIS Upgrades Amid Subscriber Losses:

Cable operators continue to face competitive pressure from fiber, cellular FWA and satellite broadband.  New Street Research reports that cable continues to lose market share, reflecting the increasing availability of alternative broadband technologies.  Nevertheless, HFC networks remain important broadband infrastructure. Cable operators are investing in DOCSIS upgrades, including DOCSIS 4.0, to increase downstream and upstream capacity and extend the competitive life of their existing coaxial access networks.

DOCSIS 4.0 supports full-duplex and extended-spectrum approaches, with the potential for multi-gigabit downstream and upstream service capabilities.  These upgrades allow cable operators to compete more effectively with fiber without replacing their entire HFC access infrastructure. However, technological improvements alone will not necessarily reverse subscriber losses. Cable operators must also address pricing, customer satisfaction, service reliability and the competitive advantages of fiber and wireless alternatives.

Satellite broadband could add another source of competitive pressure, especially in areas where cable is the principal terrestrial broadband alternative.

Technical and Economic Factors Determine the Competitive Outcome:

The four broadband technologies have different performance characteristics, deployment economics and capacity constraints.

Technology Principal advantages Principal constraints
Fiber (FTTP) High capacity, symmetrical bandwidth, low latency, scalability and reliability High initial construction costs, permitting and deployment time
Cellular FWA Rapid deployment, existing cellular infrastructure, flexible service offerings Spectrum, cell capacity, radio conditions and mobile traffic competition
LEO satellite Broad geographic coverage, rapid availability and service to remote locations Shared satellite capacity, spectrum, terminal costs, latency and congestion
Cable (HFC) Extensive existing footprint, high downstream capacity and DOCSIS upgrade path Shared coaxial capacity, upstream limitations and competition from fiber

An important technical consideration for LEO satellite broadband is that coverage does not necessarily equal unlimited service capacity.

Satellite throughput is shared among users within a service area. Actual performance depends on constellation size, satellite capacity, available spectrum, gateway infrastructure, spot-beam allocation and subscriber density.

Consequently, expanding the number of satellites does not automatically translate into proportional growth in the number of subscribers that can receive consistently high broadband speeds.

A technical assessment of LEO broadband prepared by Cartesian for the Fiber Broadband Association highlights these capacity and performance constraints. Its modeling, based primarily on public information available through November 2024 with minor updates in early 2025, estimates substantial limitations on the number of U.S. locations that Starlink could serve at the FCC’s 100 Mbps download/20 Mbps upload benchmark.

Those estimates should not be treated as current operational measurements or definitive forecasts, particularly given subsequent constellation developments. However, the underlying engineering constraint remains important: satellite capacity must be shared geographically, and concentrated demand can degrade user performance.

Fiber does not face the same last-mile radio spectrum and shared satellite beam constraints. This gives fiber a substantial long-term advantage in high-density markets and locations with significant bandwidth demand.

U.S. Fixed Broadband Market Outlook: 2027–2030:

The competitive outlook can be summarized as follows:

Satellite broadband: Continued expansion, led by Starlink, with increasing competition against terrestrial broadband providers. New Street expects satellite to capture approximately 7% of the existing broadband market in its terminal state, primarily at the expense of cellular FWA.

Cellular FWA: Continued subscriber growth, but with decelerating net additions as operators encounter capacity and market-penetration constraints. Additional spectrum and network investment could extend its growth potential.

Fiber: Strong subscriber additions through 2028, followed by slower growth as DSL conversion opportunities diminish. Fiber operators will increasingly compete for customers already served by other broadband technologies.

Cable/HFC: Continued market-share pressure from fiber, FWA and satellite, despite ongoing DOCSIS upgrades intended to improve network capacity and competitiveness.

The following table summarizes the directional outlook, rather than providing independently verified subscriber forecasts.

Technology 2027–2028 outlook Longer-term direction
Satellite Strong growth Increasing competitive presence
Cellular FWA Growth decelerates Dependent on spectrum and capacity expansion
Fiber Approximately 2.8–2.9M annual net adds Deceleration after 2028
Cable/HFC Continued market-share pressure DOCSIS upgrades support competitiveness
DSL/Copper Continued subscriber decline Nearly zero subscribers projected by 2029

Conclusions:

The U.S. fixed broadband market is becoming increasingly competitive as satellite, cellular FWA, fiber and cable technologies converge in their ability to deliver residential broadband services.

New Street Research’s forecast that satellite will take most of its market share from cellular FWA is plausible, particularly given the expected slowdown in FWA subscriber additions. However, the assumption that satellite will not take market share from fiber deserves further scrutiny.

Fiber should retain a substantial advantage in densely populated markets because of its capacity, low latency, reliability and scalability. But those advantages do not make it immune to competitive substitution in rural, exurban and selected suburban markets.

At the same time, the approaching exhaustion of DSL-to-fiber conversions will require fiber operators to compete more aggressively for subscribers already served by other broadband providers.

Ultimately, the competitive outcome will depend on the economics of delivering broadband capacity to individual geographic markets, rather than simply comparing nationwide subscriber totals.

The most important question is not whether satellite will replace fiber nationally, but how much of the addressable broadband market satellite can economically serve—and which terrestrial broadband providers will lose subscribers as that capability expands.

……………………………………………………………………………………………………………………………………………………………..

References:

https://www.lightreading.com/satellite/satellite-broadband-to-capture-most-market-share-from-fwa-report

https://fiberbroadband.org/resources/technical-assessment-of-leo-satellite-broadband/

https://www2.itif.org/2025-broadband-convergence.pdf

https://www.reuters.com/business/media-telecom/fcc-approves-spacex-plan-deploy-additional-7500-starlink-satellites-2026-01-09/

https://www.sec.gov/Archives/edgar/data/1181412/000162828026052515/earningsreleaseq22608042.htm

New Street Research study: Cable broadband will continue its decline, but total broadband access subscribers will increase

ABI Research joins the chorus: 5G FWA is a competitive alternative to wired broadband access

BroadbandNow Research: Best & Worst States for Broadband Access

U.S. broadband subscriber growth slowed in 1Q-2024 after net adds in 2023

FTTH was ~60% of France fixed broadband market in Q1-2023

 

The AI-RAN Silicon Conundrum: GPUs, General-Purpose CPUs, or Custom ASICs?

Disclaimer:

Google Gemini was extensively used to research this article and generate the comparative table herein. Google Search was used to verify and detail the AI-RAN telco trials.

Introduction:

As mobile network operators prepare to embed artificial intelligence directly into the Radio Access Network (RAN), a critical architectural battle has emerged over the underlying compute silicon. On November 18, 2026, VodafoneThree announced UK lab and field trials to evaluate Nokia’s AI-native anyRAN software running on NVIDIA’s Aerial RAN Computer platform and Red Hat OpenShift. That same day, Vodafone Group opened a European AI-RAN test hub in Málaga, Spain, designed to benchmark GPUs side-by-side against purpose-built telecom ASICs and Intel Xeon 6 (Granite Rapids) x86 processors.

Yet even as operators initiate trials, sharp disagreement persists across the semiconductor and telecom sectors over whether data center GPUs belong in the RAN. Speaking at an Omdia AI-RAN symposium, AMD telco lead Piotr Weglicki warned of an impending “big, big backlash” against deploying power-hungry data center GPUs in mobile networks, advocating instead for high-core-count general-purpose CPUs. “Probably there’s going to be a big, big backlash,” he said when commenting on the use of “power hungry” data center GPUs in the RAN.

Like Intel, AMD continues to prefers a virtual RAN based on its portfolio of CPUs. Most networks today use custom silicon, purpose-built for the RAN to be as energy efficient as possible. While difficult to develop, custom RAN silicon might have to be replaced more often to accommodate new features or upgrade to the next 3GPP Release.  That explains why a general purpose CPU holds some appeal. “What it gives telcos, actually, is access to data-center economics with much bigger R&D budgets with much faster development cycles,” said Weglicki. “What we also see is that the general-purpose compute is absorbing the RAN acceleration. With each of the generations, we bring more CPU performance within the same power envelope, and we get much better power efficiency,” he added.

Meanwhile, Vodafone’s Head of RAN Evolution, Kyriakos Exadaktylos, acknowledged strong interest in AI-RAN algorithms while noting that Vodafone has not yet given a green light to commercial GPU deployments. His hesitation reflects unresolved questions about the performance, power consumption and economics of GPU-based RAN processing compared with conventional CPUs and purpose-built ASICs. Ultimately, commercial deployment will depend on demonstrating measurable network performance gains and a compelling total cost of ownership.

…………………………………………………………………………………………………………………………………………………………………………………….

Comparing AI RAN Implementation Alternatives:

For network architects, the core challenge is clear: determining which silicon foundation—GPUs, general-purpose CPUs, or custom ASICs—can meet strict deterministic Layer 1 (L1) baseband timing deadlines while delivering acceptable power efficiency, workload portability, and total cost of ownership (TCO).

1. GPU-Accelerated AI-RAN: The NVIDIA and Nokia Model:

Backed by NVIDIA’s $1 billion strategic stake in Nokia, the GPU-centric AI-RAN model treats baseband signal processing and AI inference as unified software workloads running on accelerated computing platforms.

    • Architectural Approach: In VodafoneThree’s UK trial architecture, Nokia’s anyRAN software runs atop NVIDIA’s Aerial RAN Computer (such as the MGX 2U Short platform) orchestrated by Red Hat OpenShift. The design targets three operational tiers: AI-for-RAN (using machine learning for channel estimation, link adaptation, and beamforming), AI-and-RAN (sharing GPU resources dynamically between RAN functions and external AI workloads), and AI-on-RAN (hosting low-latency edge AI services such as vision or robotics near end users).
    • The Engineering Trade-Offs: GPUs offer massive parallel compute density and software programmability, which advocates argue can extend hardware lifecycles as AI algorithms evolve. However, real-time RAN processing operates under sub-millisecond timing budgets—often requiring link-adaptation inference decisions within roughly 40 microseconds. Critics contend that deploying high-wattage data center GPUs to execute compact RAN neural models at distributed cell sites creates severe power, thermal, and cost penalties unless workloads are heavily consolidated in centralized Cloud RAN hubs. >

2. General-Purpose CPUs: Pure Software vs. Integrated Accelerators:

To leverage high-volume data center R&D budgets and rapid silicon refresh cycles without GPU power overhead, x86 CPU vendors are pursuing two competing virtual RAN (vRAN) designs.

    • AMD’s Accelerator-Free CPU Strategy: After its earlier Xilinx FPGA-based T2 Telco Accelerator Card faced operator pushback over power draw, AMD pivoted away from discrete hardware accelerators. Instead, AMD relies on high core density—such as its 84-core server processors—to execute compute-intensive L1 tasks, including Forward Error Correction (FEC), entirely in software on general-purpose cores. Samsung Networks (via VP of Network Strategy Alok Shah) has validated multi-cell capacity on AMD’s 84-core platform without external acceleration. AMD argues this eliminates hardware lock-in and lets operators dynamically partition CPU cores across L1–L3 baseband tasks, AI-RAN models, and edge applications. AMD claims to have a big enough CPU core count per chip to handle a software-based FEC on the CPU. A model advertised earlier this year features as many as 84 cores, exceeding the 72 found in the most advanced Granite Rapids product from Intel. “Forward error correction increasingly could be done on software on general-purpose cores, and the consequences of that are huge,” Weglicki said  during the Omdia symposium.
    • Intel’s Integrated Acceleration (Granite Rapids-D): Intel’s Xeon 6 SoC (Granite Rapids-D, scaling up to 72 cores) integrates dedicated silicon acceleration blocks for L1 FEC alongside built-in AI inference instructions and integrated Ethernet. While integrated acceleration improves watts-per-bit efficiency for L1 processing compared to pure-software execution, competitors argue it complicates multi-vendor software portability across different x86 platforms. 

3. Purpose-Built Custom Silicon (ASICs): The Efficiency Benchmark:

Traditional RAN architectures rely on custom application-specific integrated circuits (ASICs) engineered specifically for deterministic baseband signal processing.

    • Architectural Approach: Leading RAN suppliers continue to embed specialized neural network accelerators directly into custom baseband and radio silicon. For example, Ericsson’s Head of RAN Strategy, Gabriel Foglander, has highlighted running AI-native link adaptation directly on custom ASICs to guarantee deterministic, sub-40-microsecond inference latency. Similarly, VodafoneThree recently deployed Ericsson’s AIR 3266 Massive MIMO radios with integrated neural accelerators in London, while Virgin Media O2 demonstrated a 20% spectral efficiency gain in an AI-RAN trial running on existing Nokia-Marvell custom silicon.
    • The Engineering Trade-Offs: Custom ASICs remain unmatched in power efficiency and deterministic L1 performance. Their primary vulnerability is economic: unlike general-purpose CPUs or GPUs, RAN ASICs are confined to a roughly $35 billion global RAN equipment market, limiting R&D amortization and risking premature hardware obsolescence if rapidly evolving AI models require architectural changes.

Comparative Architecture Matrix

Architectural Attribute GPU-Accelerated AI-RAN (NVIDIA Aerial / Nokia) Accelerator-Free x86 CPU (AMD 84-Core) Accelerated x86 CPU (Intel Xeon 6 / Granite Rapids) Purpose-Built RAN ASIC (Ericsson / Marvell)
Layer 1 & FEC Execution Offloaded to GPU parallel tensor/CUDA cores Pure software execution on general-purpose CPU cores

Integrated on-die hardware FEC accelerator + CPU cores Hardwired baseband pipeline + integrated neural blocks
Power & Thermal Profile High; best suited for centralized vRAN / Cloud RAN hubs Moderate; scales with core utilization and software efficiency Moderate-to-low; hardware offload reduces L1 core draw Lowest watts-per-Gbps; optimized for distributed cell sites
Hardware Portability Tied to GPU software stack (e.g., CUDA / Aerial) High; standard x86 software portability without custom offload Moderate; software must target integrated accelerator APIs Low; tightly coupled to vendor-specific silicon
Multi-Tenant AI Sharing High capacity for concurrent RAN and third-party AI inference Flexible core partitioning between L1–L3 and edge AI Supports edge AI inference alongside vRAN workloads Limited to embedded RAN inference models
Primary Deployment Risk Cell-site power/cooling limits and unclear edge AI monetization Power efficiency of software-only FEC at high cell densities Architectural lock-in to vendor-specific acceleration Narrow RAN R&D amortization and slower hardware refresh cycles

Image Generated by ChatGPT

………………………………………………………………………………………………………………………………………………………………………………………………………………………………………………………

AI RAN Telco Trials:

Readers should pay attention to the results of AI RAN trials this year and next for clues. Nine major telecom operators are conducting or have scheduled live network AI-RAN field trials across 2026 and 2027, distinguishing over-the-air commercial testing from early laboratory evaluations.

  • T-Mobile US (United States). Completed the world’s first live AI-RAN call on its commercial 5G network in Bellevue, Washington, in October 2026 with Nokia and NVIDIA, alongside live non-GPU AI beamforming trials with Ericsson.
  • Verizon (United States). Initiating live U.S. site rollouts of the AI-native Ericsson Intelligent Controller (EIC) before the end of 2026.
  • Elisa (Finland). Completed Europe’s first over-the-air 5G Advanced AI-RAN field trial in October 2026 ahead of planned commercial deployments in 2027.
  • Virgin Media O2 (United Kingdom). Completed a live London AI-RAN deployment in late October 2026.
  • VodafoneThree (United Kingdom). Transitioning from late-2026 lab validation into live U.K. mobile site trials with Nokia and NVIDIA, alongside live Ericsson neural-accelerator radio testing in London.
  • BT / EE (United Kingdom). Scheduled live over-the-air network trials in Birmingham for 2027 using Nokia anyRAN and NVIDIA hardware.
  • Deutsche Telekom (Germany). Moving from late-2026 lab measurements into live field trials in 2027 to test AI-based uplink channel estimation.
  • Orange (France). Conducting phased lab-to-live network AI-RAN trials across 2026 and 2027.
  • SoftBank (Japan). Scheduled live commercial network trials in 2027 running Nokia anyRAN on its AITRAS platform and NVIDIA ARC-Pro hardware.
…………………………………………………………………………………………………………………………………………………………………………………………………………………..

Conclusions:

The telecom industry is unlikely to converge on a single “winner-take-all” AI-RAN silicon architecture. Instead, deployments will segment by network topology and power envelope:

  • Rigorous Multi-Platform Benchmarking: Vodafone’s new Málaga test hub—which evaluates NVIDIA Aerial GPUs, Intel Granite Rapids CPUs, and custom ASICs using SPEC, MLPerf, and AI-RAN Alliance metrics—will provide essential empirical data. Engineers should look beyond headline spectral efficiency claims (such as Nokia’s target to double spectral efficiency by 2028) and demand normalized watts-per-Gbps and total cost per bit comparisons.
  • Deterministic Isolation in Shared Compute: For AI-and-RAN multi-tenancy to succeed commercially, vendors must prove that bursty third-party AI inference workloads cannot introduce scheduling jitter or missed deadlines into real-time L1 baseband processing.
  • Topology-Driven Silicon Selection: Custom ASICs with integrated neural accelerators will remain difficult to displace at power-constrained macro cell sites, whereas general-purpose CPUs and GPU-accelerated servers will compete primarily in centralized Cloud RAN and regional edge aggregation nodes where power, cooling, and multi-tenant enterprise AI demand can justify the infrastructure investment.

Bottom Line: The AI-RAN silicon debate is moving from raw compute capability to software portability, real-time performance, and total cost of ownership at the cell site. For general-purpose CPUs and GPUs to replace baseband ASICs at scale, the industry must overcome hurdles regarding hardware abstraction, thermal limits, and workload coexistence.

…………………………………………………………………………………………………………………………………………………………………………………………………………………..

References:

AMD Telco Position: Light Reading, “AMD telco chief says GPU ‘backlash’ is coming in AI-RAN,” November 18, 2026

https://www.lightreading.com/5g/amd-telco-chief-says-gpu-backlash-is-coming-in-ai-ran

Nokia & NVIDIA Trials: Light Reading, “VodafoneThree to test Nokia’s AI-RAN with Nvidia GPUs,” November 18, 2026  https://www.lightreading.com/ai-machine-learning/vodafonethree-to-test-nokia-s-ai-ran-with-nvidia-gpus

NVIDIA Arc Connect:

https://developer.nvidia.com/blog/deploy-ai-ran-at-cell-sites-with-nvidia-arc-compact/

Vodafone AI-RAN Hub: Vodafone Newsroom, “Vodafone builds AI-RAN test hub to improve mobile connectivity and support new AI services,” November 18, 2026 

https://www.vodafone.com/news/newsroom/technology/malaga-lab-ai-ran

Vodafone Strategy: Light Reading, “Vodafone is sold on AI-RAN but not yet on need for GPUs,” November 17, 2026

https://www.lightreading.com/5g/vodafone-is-sold-on-ai-ran-but-not-yet-on-need-for-gpus

Industry Ecosystem: RCR Wireless News, “VodafoneThree joins Nokia/Nvidia test squad as carriers kick tires on AI-RAN,” November 18, 2026

https://rcrwireless.com/20261118/network-infrastructure/vodafonethree-joins-nokia-nvidia-test-squad-as-carriers-kick-tires-on-ai-ran

IEEE ComSoc Technology Blog, “Orange, Nokia, Nvidia, and Intel debate: ASICs vs. GPUs vs. General-Purpose CPUs for RAN Baseband Processing,” April 29, 2026

https://techblog.comsoc.org/2026/04/29/orange-nokia-nvidia-and-intel-asics-vs-gpus-vs-general-purpose-cpus-for-ran-baseband-processing/

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

RAN silicon rethink – from purpose built products & ASICs to general purpose processors or GPUs for vRAN & AI RAN

RAN Silicon Rethink- Part II; vRAN and General-Purpose Compute

Analysis: Nvidia’s $2 billion investment in Marvell; NVLink Fusion ecosystem & RAN vendor silicon strategy

Jensen Huang’s Resignation as CEO Creates Uncertainty for Nvidia and Opportunities for Competing AI Chip Suppliers

AI’s Double-Edged Impact on Telecom Networks: Accelerating Cyber Threats While Improving Network Resilience

Disclaimer: ChatGPT was used in a dialog to help select this important telecom topic and subsequently conduct research that the author used to write this article.

Introduction:

Artificial intelligence (AI) is creating a fundamental paradox for telecommunications networks and other critical infrastructure. The same technology that enables network operators to predict equipment failures, improve service reliability and detect security threats can also help malicious actors discover software vulnerabilities, develop exploits and accelerate cyberattacks.

AI is therefore becoming both a potential source of increased network risk and an important tool for reducing it. The outcome will depend on how quickly telecommunications providers and other critical infrastructure operators can deploy effective AI-enabled defenses, improve operational resilience and cooperate in responding to emerging threats.

AT&T CEO John Stankey recently highlighted this dilemma at Semafor’s Silicon Valley & The World conference. He said AI makes all companies more vulnerable to cyberattacks, at least for now, but rejected a pessimistic outlook.  “At this moment in time, I think everybody’s more vulnerable,” said Stankey. “The meat’s going to be put on the bones by businesses and industry working with each other, working with those that are generating the new technology … and it’s going to require better sharing,” he added.

Stankey believes AI can ultimately make critical infrastructure safer. However, he cautioned that businesses need time, better information sharing and cooperation with AI developers to turn the technology from a liability into an advantage.  His comments reflect an increasingly urgent cybersecurity challenge: AI is accelerating both the discovery of vulnerabilities and the ability to exploit them. He highlighted Anthropic’s Project Glasswing controlled rollout as an example of how businesses can work with AI technology companies in assessing security risks and finding vulnerabilities before new AI models are widely released.

“Now there’s a long list of things that have to be addressed and fixed, many of which are going to stretch out for two years or more. The question, how much risk do you want to take during that period, and how do you mitigate that?” Stankey concluded.

Source: Marcos Alvarado/Alamy Stock Photo
………………………………………………………………………………………………………………………………………………………………………………….

AI Increases the Cyber Threat from Malicious Actors:

AI is changing the economics and speed of cyberattacks. Previously, many sophisticated attacks required substantial human expertise and considerable time to identify targets, investigate network weaknesses and develop working exploits. Advanced AI models can increasingly automate or accelerate portions of that process.

Potential applications by malicious actors include:

  • Automated reconnaissance: Identifying exposed network infrastructure, vulnerable systems, software versions and misconfigured services.
  • Vulnerability discovery: Analyzing large software codebases, firmware and system configurations to identify weaknesses that conventional security tools may overlook.
  • Exploit development: Assisting with the development and testing of exploit code and identifying ways to combine multiple vulnerabilities into an effective attack.
  • Social engineering: Generating convincing phishing messages, impersonation attempts and other targeted communications.
  • Attack orchestration: Using AI agents to automate reconnaissance, target selection and other stages of an attack.

The UK National Cyber Security Centre (NCSC), in its 2026 Annual Review, describes AI as a force multiplier for cyber threats, increasing their scale, speed and efficiency. It identifies reconnaissance, social engineering, vulnerability discovery and exploit development as areas where AI is already changing cyber operations.

Importantly, the NCSC does not conclude that AI has fundamentally changed the objectives of hostile actors or that fully autonomous cyberattacks are commonplace. Rather, AI is expected to accelerate and scale existing attack techniques. Human operators remain important in directing and validating AI-generated results, particularly for sophisticated operations.

The concern is that AI-assisted vulnerability discovery and exploit development could substantially shorten the interval between identifying a weakness and exploiting it. Network operators and software suppliers may have less time to evaluate, patch and remediate vulnerabilities before attackers attempt to use them.

Claude Mythos illustrates the changing threat:

Anthropic’s Claude Mythos Preview provides an important example of the capabilities involved. The model demonstrated advanced abilities to identify software vulnerabilities and construct complex exploit chains. Anthropic subsequently established Project Glasswing, a controlled initiative providing selected technology companies and critical infrastructure organizations with access to the model for defensive security work.

One participant, Comcast, used Claude Mythos Preview to examine 258 business-critical systems encompassing approximately 170 million lines of code. According to Anthropic, the model identified a critical authentication vulnerability in a public-facing platform that existing scanners had missed. The weakness was remediated before evidence of exploitation was observed.

This example demonstrates why AI-enabled vulnerability discovery matters. A weakness that might appear insignificant when examined in isolation can become serious when combined with other software or configuration problems. AI models can help trace these relationships across complex systems.

However, the same capabilities could eventually be used by hostile actors. The defensive advantage provided by controlled access to advanced models is not a permanent guarantee that comparable capabilities will remain unavailable to attackers.

Telecom Operators Use AI to Improve Network Reliability:

The cybersecurity threat is only one side of AI’s impact on telecommunications. Network operators are also applying machine learning, predictive analytics and increasingly agentic AI to improve network operations.

Modern telecommunications networks generate enormous volumes of telemetry, performance measurements, alarms and configuration data. These data originate from routers, switches, optical transport systems, radio access networks, core network equipment, data centers and supporting infrastructure.

Conventional network management systems frequently rely on predefined thresholds, rules and alarm correlation. These methods remain useful, but they can struggle to identify complex failures involving multiple network elements, vendors and interconnected domains.

AI can analyze much larger volumes of operational data, recognize patterns and identify developing problems that might otherwise remain undetected until service performance deteriorates.

1.  Detecting incipient network failures

AI-based network analytics can identify anomalies and early warning signals, including:

  • Unusual traffic patterns, congestion and increasing packet loss or latency.
  • Deteriorating optical link performance and transceiver behavior.
  • Abnormal temperature, power consumption or cooling conditions.
  • Changes in equipment performance that precede hardware failures.
  • Correlated alarms across network elements that may indicate a common underlying problem.

By identifying these conditions earlier, operators can investigate and correct problems before they develop into service-affecting outages.

2. Predictive maintenance

Predictive maintenance applies AI to historical equipment performance, real-time telemetry and failure patterns to estimate the likelihood of future component failures.

For example, an AI system might identify an optical transceiver exhibiting progressive performance degradation, a power supply showing abnormal behavior or a network component approaching a failure condition. Maintenance teams can then schedule inspection or replacement before the component fails.

The potential benefits include fewer unplanned outages, reduced maintenance costs, fewer unnecessary truck rolls and improved network availability.

3.  AI-assisted fault isolation and network operations

AI can also correlate alarms, identify probable root causes and recommend corrective actions across multiple network domains. This can reduce the time engineers spend collecting data and manually investigating complex incidents.  A concrete example is NTT DOCOMO’s commercial deployment of an agentic AI system for mobile network maintenance in February 2026. Developed using AWS services, the system analyzes traffic and alarm data in real time from more than one million network devices, including base stations and core network equipment.

It uses multiple AI agents and graph-modeled network topology to identify anomalies, locate suspected failure points and recommend actions to maintenance engineers. DOCOMO reported that the system reduced response times for complex network failures requiring manual analysis by more than 50%.  This is an important example because the system is not simply generating summaries of network alarms. It is correlating operational data across a large, interconnected network to assist in fault isolation and service restoration.

Nevertheless, DOCOMO’s implementation retains human involvement in evaluating recommended actions. This is an important distinction when AI is being introduced into production networks where an incorrect automated response could cause additional service disruption.

AI Also Strengthens Telecom Cybersecurity Defenses

AI applications in network reliability and AI applications in cybersecurity are different disciplines, but both benefit from the ability to analyze large volumes of data, identify anomalies and recognize relationships that may not be apparent through conventional monitoring.

Telecom operators can use AI-based security systems to:

  • Analyze network traffic and security logs to identify suspicious behavior and potential intrusions.
  • Detect unusual access patterns, compromised accounts and potentially infected network elements.
  • Correlate security events across network infrastructure, cloud environments and enterprise IT systems.
  • Identify vulnerabilities in software, firmware, APIs and network configurations.
  • Prioritize vulnerabilities based on exposure, exploitability and potential service impact.
  • Accelerate security investigations, incident classification and containment.

AI-assisted vulnerability analysis is particularly significant. Conventional vulnerability scanners are useful for identifying known weaknesses and individual coding problems, but may not recognize complex exploit chains involving multiple software components and configuration settings.

As the Comcast example illustrates, advanced AI models can examine relationships across codebases and identify weaknesses that conventional scanning tools may miss.

This creates an opportunity for telecom operators to use advanced AI capabilities to examine their own network software, management systems and supporting infrastructure in ways that more closely resemble how sophisticated attackers might investigate them.

However, identifying a vulnerability is only the beginning. Operators must validate whether it is exploitable in their actual environments, determine its severity, identify affected systems and deploy appropriate remediation without disrupting services.

AI Cybersecurity Arms Race: Can Defenders Keep Pace?

The central question is whether AI-enabled defenses can advance quickly enough to offset the additional capabilities AI provides to attackers.  Both sides can benefit from faster vulnerability discovery, automated analysis and improved recognition of complex patterns. But the operational circumstances are different.  An attacker may need to identify only one exploitable weakness to compromise a target. A telecommunications provider must protect a complex, continuously operating infrastructure comprising thousands or millions of network elements, multiple vendors, legacy systems, cloud-native functions and interconnected management platforms.  Furthermore, operators cannot simply apply every available patch or configuration change immediately. Network changes must be tested, validated and coordinated to avoid introducing service disruptions. In some cases, maintenance windows, equipment compatibility and dependencies among network functions complicate remediation.

This creates several challenges:

  • The vulnerability remediation backlog. AI may identify vulnerabilities faster than operators and equipment suppliers can validate, prioritize and correct them. A larger volume of findings does not automatically translate into improved security.
  • False positives and exploitability. AI-generated vulnerability reports require technical validation. Not every identified weakness is exploitable, and not every theoretical exploit represents an immediate operational threat.
  • The risks of automated remediation. AI systems that recommend or execute network changes must operate within carefully defined permissions and policies. An incorrect configuration change or automated response could cause the very service disruption the operator is trying to prevent.
  • AI system security. AI agents themselves introduce new risks involving access permissions, sensitive operational data, model manipulation, unreliable outputs and potentially unauthorized actions. Operators must secure the AI systems used to protect their networks.
  • Human oversight. AI can accelerate analysis and decision-making, but network operators remain responsible for service availability, security and regulatory compliance. Human validation and well-defined escalation procedures remain essential, particularly for high-impact network changes.

The challenge is therefore not simply to deploy more AI. It is to integrate AI into operational and cybersecurity processes with sufficient accuracy, governance and accountability to improve resilience rather than create additional vulnerabilities.

Cross-Sector Cooperation Is Essential:

Stankey’s call for greater cooperation recognizes that cybersecurity threats do not respect industry boundaries. Telecommunications networks support financial services, healthcare, transportation, energy, government and virtually every other major economic sector. A compromise affecting one part of this interconnected infrastructure can have consequences well beyond the original target.

Effective defenses therefore require more than individual companies improving their own security operations.

Telecommunications providers, equipment manufacturers, cloud service providers, AI developers, cybersecurity firms and government agencies need to share actionable threat intelligence and coordinate responses to vulnerabilities affecting common infrastructure.

AT&T is a founding member of the Alliance for Critical Infrastructure, which brings together companies from communications, energy, financial services, information technology, transportation and water systems. The initiative provides a forum for cooperation among organizations that have a shared interest in protecting critical infrastructure.

As note above, Anthropic’s Project Glasswing is another example of cross-industry cooperation. By providing selected organizations with access to advanced AI models for defensive security research, the initiative seeks to identify and remediate vulnerabilities in critical software before comparable capabilities become more widely available.  Such initiatives are useful, but cooperation must translate into operational results: faster vulnerability disclosure, coordinated patch development, improved threat intelligence sharing and more effective incident response.

Conclusions: AI Can Make Telecom Networks More Resilient, but Not Automatically More Secure:

AI is simultaneously increasing the cyber threat and providing telecommunications providers with new tools to improve network reliability and security.  Malicious actors can use AI to accelerate reconnaissance, discover vulnerabilities, develop exploits and scale existing attack techniques. Telecom operators, meanwhile, can use AI to detect incipient network failures, predict equipment degradation, improve maintenance, identify security anomalies and accelerate vulnerability remediation.  These applications should not be confused. Predicting an optical transceiver failure is fundamentally different from detecting an intrusion or identifying an exploitable software vulnerability. Nevertheless, both demonstrate AI’s ability to extract actionable intelligence from enormous volumes of complex network data.

The DOCOMO deployment illustrates the operational benefits already being achieved through AI-assisted fault isolation. Anthropic’s Project Glasswing and Comcast’s experience demonstrate how advanced AI models can help defenders identify security weaknesses that conventional tools may miss. Neither example, however, establishes that AI can independently guarantee network reliability or cybersecurity.

The ultimate issue is whether defenders can use AI more effectively than attackers can use it to exploit weaknesses. That will require continuous vulnerability assessment, faster remediation, secure AI deployment, rigorous operational controls and cooperation across critical infrastructure sectors.  Repositioning AI from a cybersecurity liability into a defensive advantage will require more than adopting new models and automation tools. It will require telecom operators and their partners to fundamentally improve the speed, coordination and effectiveness of network operations and security.

…………………………………………………………………………………………………………………………………………………………………………………………..

References:

https://www.lightreading.com/ai-machine-learning/at-t-ceo-calls-for-cross-sector-action-to-combat-ai-risks

https://www.semafor.com/hub/semafor-silicon-valley-and-the-world-2026

https://www.ncsc.gov.uk/collection/annual-review-2026/understanding-threat/ai-impact-cyber-threat

https://www.ncsc.gov.uk/sites/default/files/2026-11/ncsc-annual-review-2026.pdf

https://www.anthropic.com/project/glasswing

https://claude.com/resources/articles/how-comcast-and-booz-allen-use-claude-mythos-to-secure-their-codebases

https://dcm-dpweb61.tz.nttdocomo.co.jp/english/info/media_center/pr/2026/0225_01.html

https://www.ericsson.com/en/blog/2026/6/from-data-to-decisions-make-agentic-ai-driven-telecom-operations-a-reality

https://www.nokia.com/blog/from-prediction-to-action-the-next-leap-in-ai-driven-network-care/

Gates warns of “turbulent AI era;” OpenAI calls for collective action on AI cybersecurity

Anthropic’s Project Glasswing aims to reshape IT cybersecurity

SHIELD-6G with AI-native cyber threat intelligence platform to enhance cybersecurity for Europe’s future 6G networks

Cybersecurity threats in telecoms require protection of network infrastructure and availability

Analysis & Implications of the Communications Cybersecurity Information Sharing and Analysis Center (C2 ISAC)

Key Differences Between Network Cybersecurity and Control System Cybersecurity & Why It Matters

Emerging Cybersecurity Risks in Modern Manufacturing Factory Networks

StrandConsult Analysis: European Commission second 5G Cybersecurity Toolbox report

 

Broadcom’s Open Ethernet Fabric: Architectural Evolution, AI Tail Latency, Proprietary Interconnects, Cisco Silicon One

Introduction:

On November 15, 2026, Broadcom showcased its open Ethernet fabric portfolio for artificial intelligence (AI) and high-performance computing at the Super Computing 2026 (SC26) event in Chicago, IL. The silicon included the 102.4-terabit Tomahawk 6 switch, Tomahawk Ultra, Jericho 4, and the 800G Thor Ultra network interface card with enhanced RoCE.  Those components are designed to rival proprietary systems like Nvidia’s InfiniBand and NVLink in AI and HPC clusters. With bandwidth reaching parity across ASICs, competition now centers on tail-latency determinism in collective operations.

Here’s a comparison chart that provides a quick comparison of the individual Broadcom component attributes:

Component Silicon Class Throughput Key Architectural Features Target Topology Role
Tomahawk 6 (BCM78910) Ethernet Switch ASIC 102.4 Tbps 3nm process, 512 × 200G SerDes, dynamic load balancing, UEC telemetry Scale-out spine/leaf fabrics for massive GPU clusters
Tomahawk Ultra Low-Latency Switch ASIC 51.2 Tbps ~250ns latency, Credit-Based Flow Control (CBFC), Link Layer Retry (LLR) Intra-rack scale-up accelerator memory fabrics
Thor Ultra AI Network Interface Card (NIC) 800 Gbps (2 × 400G / 1 × 800G) PCIe Gen6 x16, eRoCE, direct GPU memory placement, packet trimming Host accelerator endpoints (GPUs, TPUs, XPUs)
Jericho 4 Fabric / Core Router ASIC Multi-Tbps Modular Deep HBM packet buffers, virtual output queuing (VOQ) Inter-pod aggregation and multi-datacenter AI federations

Broadcom also co-presented technical sessions at the SC26 Workshop on High-Performance Fabrics for AI and HPC, organized in collaboration with researchers from The Ohio State University. The technical presentations highlighted how emerging specifications—including Scale-Up Ethernet Transport (SUE-T) under the Open Compute Project and transport standards from the Ultra Ethernet Consortium (UEC)—provide deterministic, lossless transmission for collective communication routines across distributed foundation models.

…………………………………………………………………………………………………………………………………………………………………………………

Architectural Evolution: Toward Open Ethernet Fabrics:

Broadcom’s Tomahawk 6 switches and Thor Ultra Ethernet NICs illustrate the growing role of Ethernet in high-performance AI and HPC networking. Rather than relying solely on conventional RoCEv2 behavior, this architecture combines higher switching capacity with packet-level multipath out-of-order data placement, selective retransmission, and programmable congestion control. These capabilities address key transport and congestion-management challenges in large accelerator clusters.

Thor Ultra’s enhanced RDMA over Converged Ethernet, or eRoCE [1.], allows supported RDMA traffic to reach accelerator memory even when packets arrive out of order. This capability enables packet spraying across multiple network paths, improving load distribution without requiring every packet in a transfer to follow the same route. Ordering requirements still apply to some operations, including sends and atomics; out-of-order placement is not universal across all RDMA traffic.

Note 1. eRoCE is a network protocol that allows servers to use Remote Direct Memory Access (RDMA) directly over standard Ethernet networks. It lets applications read and write data in the memory of a remote computer with ultra-low latency and minimal CPU involvement.

…………………………………………………………………………………………………………………………………………………………………………………………….

The architecture also supports selective retransmission and packet trimming. Selective retransmission improves recovery efficiency by resending missing packets rather than unnecessarily retransmitting data already received. Packet trimming provides a congestion signal by removing a packet’s payload while preserving header information, helping endpoints detect loss and initiate recovery. These mechanisms support more responsive congestion handling, but should not be characterized as providing instantaneous recovery or eliminating every source of network congestion.

This direction is consistent with the Ultra Ethernet Consortium’s goal of developing an open, interoperable communications architecture for AI and HPC. The consortium emphasizes multipathing, congestion management, scalability, and compatibility with widely deployed APIs. Support for individual UEC features, however, should be distinguished from demonstrated interoperability across a complete multivendor deployment.

Hot Chips 2026 Broadcom Thor Ultra Ethernet NIC -Ethernet Networking for AI and HPC

…………………………………………………………………………………………………………………………………………………………………

Solving AI Tail Latency: Broadcom vs. Proprietary Fabrics:

To overcome the congestion and recovery bottlenecks of legacy RoCEv2 (such as ECMP hash collisions and PFC pause storms), Broadcom’s SC26 announced architecture splits intelligence between the Thor Ultra NIC and Tomahawk switches. This integrates per-packet spraying, hardware-level out-of-order GPU reassembly, packet trimming with hardware SACK, and Link-Layer Retry (LLR) with credit-based flow control (CBFC) to achieve ~250ns scale-up latency [2.].

Note 2. Scale-up latency: For an Ethernet switch, scale-up latency refers to the time required for a small control, memory, or synchronization packet to traverse a single-hop network domain (such as a rack or pod) and reach a target accelerator (XPU/GPU). However, Ethernet switch latency is not equivalent to end-to-end network latency, and neither necessarily represents collective-operation completion time across a GPU cluster. Therefore, Tomahawk Ultra’s approximately 250 ns switch latency should not, by itself, be interpreted as evidence of performance comparable to proprietary accelerator interconnects for AI collective operations.

………………………………………………………………………………………………………………………………………………………………..

Whereas proprietary fabrics like Nvidia InfiniBand and NVLink rely on single-vendor, closed-loop credit control and specialized memory semantics to prevent packet drops, Broadcom achieves equivalent tail-latency determinism over standard IEEE 802.3 Ethernet frames. By spraying packets across all available links and allowing the Thor Ultra NIC to reassemble out-of-order payloads directly into GPU memory via dma-buffer, the fabric eliminates ECMP elephant-flow collisions [3.] without incurring CPU reordering overhead. Simultaneously, switch-level packet trimming and link-local LLR/CBFC recover transient 200G SerDes bit errors and queue spikes in nanoseconds, matching proprietary collective completion times while avoiding vendor lock-in.

Note 3.  ECMP elephant-flow collisions occur when Equal-Cost Multi-Path (ECMP) routing hash algorithms accidentally assign two or more massive, high-bandwidth data transfers—known as elephant flows—to the same physical network path.

…………………………………………………………………………………………………………………………………………………………………

Implications for Research and Cloud Infrastructure:

Tomahawk 6 provides 102.4 Tbps of switching capacity on a single ASIC. Broadcom has described architectures using this capacity to connect approximately 128,000 accelerators through two switching tiers. Depending on port speeds, topology, and oversubscription requirements, higher switch radix can reduce switching stages and the associated optics, cabling, latency, and power overhead.

For university computing centers, national laboratories, and cloud operators, the broader opportunity is an Ethernet-based infrastructure strategy supported by a wider supplier ecosystem. The UEC explicitly identifies manufacturer choice and interoperability as intended benefits, alongside performance, scalability, and total cost of ownership. These goals make open Ethernet relevant to organizations evaluating both technical requirements and long-term procurement flexibility.

For application and systems teams, the objective is to expose advanced network capabilities through compatible software interfaces—not to imply that high-performance RDMA traffic must traverse the conventional Linux networking data path. The UEC’s stated emphasis on compatibility with widely deployed APIs provides a more precise basis for discussing software portability and ecosystem continuity.

The strongest architectural conclusion is that Ethernet is evolving beyond higher link speeds alone: endpoint transport capabilities and fabric-level congestion management are becoming central to its suitability for large-scale AI and HPC workloads. Broadcom’s published features support that conclusion, while workload-performance and energy-efficiency comparisons still require clearly defined test conditions.

Broadcom Tomahawk vs. Cisco Silicon One:

Broadcom’s disaggregated strategy—utilizing specialized ASICs like Tomahawk Ultra, Tomahawk 6, and Jericho 4 paired with Thor Ultra NICs—contrasts with Cisco’s unified, P4-programmable Silicon One architecture (such as the G300 and G302 series). While Broadcom offloads congestion management and reassembly to the endpoint NIC and keeps switch pipelines lean, Cisco relies on large on-die shared buffers and switch-driven Intelligent Packet Flow to minimize collective completion times without specialized endpoint dependencies.

In practice, this creates a clear architectural trade-off for tail-latency control across heterogeneous clusters. Broadcom’s endpoint-offload design requires pairing Tomahawk switches with UEC-compliant NICs like Thor Ultra to execute hardware out-of-order reassembly and packet trimming, delivering maximum scale-up determinism when both ends of the wire are aligned.

Conversely, Cisco’s Intelligent Packet Flow absorbs bursty all-reduce micro-bursts directly within Silicon One’s deep, fully shared on-die packet buffers while dynamically steering flows via real-time fabric telemetry, enabling predictable tail latency even when connected to standard, off-the-shelf third-party RoCEv2 NICs.

……………………………………………………………………………………………………………………………………………………………………

https://casrai.org/news/broadcom-sc26-tomahawk-6-thor-ultra-open-ethernet-ai-2026

https://investors.broadcom.com/news-releases/news-release-details/broadcom-ships-tomahawk-ultra-reimagining-ethernet-switch-hpc

https://www.servethehome.com/broadcom-tomahawk-ultra-launch-for-scale-up-ethernet/

https://www.broadcom.com/company/news/product-releases/63361

https://www.servethehome.com/broadcom-jericho4-51-2tbps-ai-router-chip-now-shipping-with-3-2tbps-hyperports/

https://www.broadcom.com/company/news/articles/ai-infrastructure/broadcom-brings-open-high-performance-ai-networking-to-sc26

https://www.broadcom.com/company/news/articles/ai-infrastructure/ocp-global-summit-2026-broadcom-advances-open-ethernet-scalable-ai

https://www.servethehome.com/broadcom-thor-ultra-ethernet-nic-at-hot-chips-2026/

https://www.cisco.com/c/en/us/solutions/collateral/silicon-one/silicon-one-g300-ds.html

https://investor.cisco.com/news/news-details/2026/Cisco-Advances-Scalability-and-Openness-at-OCP-2026/default.aspx

https://blogs.cisco.com/datacenter/ai-network-performance-with-cisco-intelligent-packet-flow

Ethernet Advances Towards 400 Gb/s-per-Lane Signaling for AI Networks: IEEE P802.3dv, Copper and Optical Interconnects

1.6T Ethernet Packet Loss Problem Solved: PHY-Layer Buffer Exhaustion with MAC-Layer Credit-Based Flow Control

2026 OCP Summit: Full-Stack AI Networking Silicon from Nvidia/Cisco and Broadcom; Ethernet vs InfiniBand Interconnects

The AI Infrastructure Build-Out: A $10 Trillion Bet on Compute, Power, and Networks

Jensen Huang’s Resignation as CEO Creates Uncertainty for Nvidia and Opportunities for Competing AI Chip Suppliers

Introduction:

Jensen Huang’s abrupt resignation as Nvidia’s President and CEO on November 12, 2026, creates uncertainty for the world’s leading AI semiconductor company and potentially far-reaching consequences for the artificial intelligence (AI) industry. Huang co-founded Nvidia in 1993 and spent 33 years transforming it from a graphics processor supplier into the dominant provider of accelerated computing platforms powering today’s AI infrastructure buildout.

Huang was much more than Nvidia’s CEO. He was the AI industry’s chief cheerleader, dealmaker and ecosystem architect. Through his technological vision, industry relationships, investments and strategic partnerships, he helped establish Nvidia at the center of an extensive network of AI developers, cloud service providers, semiconductor companies and data center operators.

“My management style is what made it possible for the company to grow,” Huang told Bloomberg News’ Emily Chang in a recent interview. “I wanted an organization that was able to adapt to change, both in inventing change for itself and the rest of the industry, as well as reacting to change that’s coming toward us.”

Nvidia’s board appointed longtime CFO Colette Kress as interim CEO while beginning a search for a permanent successor. Although the company says its operations, customer commitments and product roadmap will continue without interruption, the sudden departure of the executive who personally drove much of its strategy raises questions about leadership continuity, customer confidence and the future direction of its AI ecosystem.

Huang’s exit could reverberate across the global tech sector, well past Nvidia’s position as the primary silicon supplier for the AI era. As an active strategic dealmaker, Huang has leveraged the company’s resources to strengthen critical supply chain bottlenecks and back next-generation AI startups. Furthermore, he remains one of the sector’s most visible evangelists, championing accelerated computing and enterprise AI adoption while countering alarmist views on existential safety risks.  The most important potential consequence is competitive: uncertainty surrounding Nvidia could encourage customers to accelerate diversification of their AI accelerator suppliers, creating opportunities for AMD, Google’s Tensor Processing Units (TPUs), Broadcom, Cerebras and Huawei’s Ascend AI processors in China.

Nvidia’s Leadership and Ecosystem Advantage:

Nvidia’s competitive position is based on considerably more than the performance of its GPUs. The company has developed an integrated computing platform combining:

  • GPU accelerators: High-performance processors designed for parallel computing workloads, including AI model training and inference.
  • NVLink and NVSwitch: High-bandwidth interconnect technologies that allow multiple GPUs to operate as tightly coupled computing systems.
  • InfiniBand and Spectrum-X Ethernet: Networking platforms connecting large numbers of GPUs across AI data centers via the company’s acquisition of Mellanox (announced March 11, 2019; completed on April 27, 2020).
  • CUDA software ecosystem: A programming platform, libraries and development tools that support a large community of AI developers and applications.
  • Complete AI computing systems: Integrated combinations of GPUs, CPUs, networking, storage and software designed for large-scale AI infrastructure.

This combination of hardware, networking and software has created substantial customer dependence on Nvidia’s technology. Replacing Nvidia GPUs is not simply a matter of purchasing competing accelerator chips. Customers must also consider software compatibility, application optimization, networking architecture, performance, power consumption and the cost of migrating existing workloads.

Huang was instrumental in building this integrated ecosystem and cultivating the relationships that helped make Nvidia the preferred AI computing platform for many of the industry’s largest customers.  His departure raises critical questions about whether the AI industry will continue its unprecedented circular financing deals.

…………………………………………………………………………………………………………………………………………………………………………………………………………

Will Nvidia’s Circular AI Financing Continue Without Jensen Huang?

The diagram above shows how Nvidia sits at the center of a complex web of circular investments, hardware purchases and services relationships. Broadcom is another important participant, but the central question is whether Nvidia’s financing strategy will retain its momentum without Huang, who personally championed the initiative.  Nvidia announced in August 2026 that it was partnering with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR to establish financing platforms intended to mobilize more than $500 billion in third-party capital for AI infrastructure. Huang described the initiative as a way to turn AI computing infrastructure into an investable asset class.  However, Bloomberg subsequently reported that the initiative had not produced financing deals after months of activity, with some customers and financial partners seeking stronger financial backing from Nvidia. The uncertainty surrounding Huang’s departure therefore raises questions about the initiative’s future.

Will Nvidia’s new leadership maintain Huang’s aggressive approach to financing the AI ecosystem, including investments and financial commitments that help stimulate demand for Nvidia’s own products? Or will the company become more conservative in allocating capital and assuming financial exposure?

The implications extend beyond Nvidia. As the Bloomberg graphic illustrates, AI infrastructure investment increasingly involves interconnected relationships among chip suppliers, cloud providers, AI developers and financial institutions. If Nvidia reduces or restructures its financial commitments, the consequences could affect AI infrastructure deployment, customer purchasing plans and competing semiconductor suppliers.  Ultimately, the question is whether the AI industry’s closely interconnected investments and financing model can sustain its momentum without the executive who was its most prominent champion and dealmaker.

………………………………………………………………………………………………………………………………………………………………………………………………

Potential Opportunities for Competing AI Chip Suppliers and Platforms:

Jensen’s exit also raises the question if Nvidia can maintain the same pace of technological innovation, strategic deal making and ecosystem development under different leadership?   A sudden leadership transition at a company as strategically important as Nvidia could prompt some customers to reassess their dependence on its products. Large cloud providers and AI developers have strong incentives to maintain alternative accelerator platforms, reduce supplier concentration and negotiate more favorable computing costs.  Several AI chip competitors warrant particular attention.

AMD: Direct Competition in AI GPUs

Advanced Micro Devices (AMD) is Nvidia’s most direct competitor in the market for general-purpose AI GPU accelerators.

AMD’s Instinct accelerator family, supported by its ROCm software ecosystem, offers an alternative platform for AI training and inference. AMD has been working to improve accelerator performance, memory capacity, interconnect capabilities and software compatibility to compete more effectively with Nvidia.

The principal opportunity for AMD is to attract customers seeking a second source of AI accelerators or greater negotiating leverage in purchasing decisions.

However, AMD must demonstrate competitive performance, reliable software support and the ability to deliver complete systems at scale. Nvidia’s established CUDA ecosystem remains a significant competitive advantage.

Google: Tensor Processing Units

Google’s Tensor Processing Units are custom-designed application-specific integrated circuits (ASICs) developed to accelerate machine learning workloads.

Unlike general-purpose GPUs, TPUs are designed around Google’s particular AI computing requirements and software environment. They are deployed within Google Cloud and support Google’s own AI models and services, as well as external customers using its cloud infrastructure.  Google’s TPU platform provides an alternative to Nvidia GPUs for selected workloads, particularly where customers can optimize their software and applications for Google’s architecture.  Google could benefit if customers increasingly seek alternatives to Nvidia’s GPU platform, although its opportunity is primarily tied to its own cloud and AI ecosystem rather than unrestricted sales of standalone accelerators.

Broadcom: Custom AI Accelerators and Networking Silicon

Broadcom is arguably one of Nvidia’s most important competitors in the broader AI semiconductor market, although it competes differently from AMD.

Rather than selling general-purpose GPUs, Broadcom specializes in designing custom AI accelerators (ASICs) for hyperscale customers. Its XPU platform combines custom compute silicon, high-speed SerDes, advanced packaging, memory interfaces and networking technologies.

Broadcom’s partnerships with major AI infrastructure customers are particularly significant:

  • Google: Broadcom has been a key technology partner in Google’s custom TPU development.
  • Meta: Broadcom is collaborating on Meta’s MTIA (Meta Training and Inference Accelerator) custom AI silicon, including a multi-generation deployment roadmap.
  • OpenAI: Broadcom and OpenAI announced a collaboration in June 2026 to develop custom AI accelerators optimized for large language model inference, with planned deployment at gigawatt scale.

Broadcom’s competitive significance extends beyond individual chips. Its Ethernet networking and connectivity technologies also support the construction of large-scale AI computing clusters.

As hyperscalers increasingly develop their own AI accelerators to reduce dependence on Nvidia GPUs, Broadcom is positioned to benefit from the shift toward custom silicon.

Cerebras Systems: Wafer-Scale AI Computing

Cerebras Systems takes a substantially different architectural approach to AI acceleration.

Its Wafer-Scale Engine (WSE) integrates an exceptionally large number of processing cores, memory and interconnects onto a single silicon wafer. This architecture is designed to reduce communication bottlenecks associated with distributing AI workloads across large numbers of conventional processors.

Cerebras offers integrated AI computing systems rather than simply selling accelerator cards. Its technology targets large-scale AI model training and inference, with particular emphasis on high-throughput, low-latency inference.

Cerebras has attracted significant attention through commercial relationships with major AI developers, including OpenAI.

Although Cerebras does not yet have Nvidia’s breadth of software ecosystem, installed base or general-purpose GPU market presence, its wafer-scale architecture represents a credible alternative for selected AI workloads.

Its potential competitive advantage is architectural differentiation rather than direct replication of Nvidia’s GPU platform.

Intel: Gaudi AI Accelerators and Data Center Computing

Intel is another potential competitor, although its position in AI accelerators is less established than AMD’s or Broadcom’s.

Intel’s Gaudi 3 accelerator was designed for generative AI training and inference, featuring high-bandwidth memory and integrated networking capabilities. Intel has promoted Gaudi as an alternative to Nvidia GPUs, emphasizing performance, power efficiency and total cost of ownership.

Intel also brings an extensive data center processor business and established relationships with enterprise customers and server manufacturers.

However, Intel has faced challenges establishing Gaudi as a widely adopted alternative to Nvidia’s dominant GPU platform. Its ability to compete will depend on future accelerator products, software development, customer adoption and execution of its AI semiconductor roadmap.

Intel should therefore be considered a potential challenger rather than an established competitor on the same footing as AMD or Broadcom.

Huawei: China’s Domestic AI Accelerator Alternative

Huawei’s Ascend AI processor family is an important alternative for Chinese AI developers and data center operators.  U.S. export restrictions have already encouraged China’s domestic semiconductor industry to develop alternatives to advanced Nvidia accelerators. Huawei is positioned to benefit from this trend through its Ascend AI processors, computing systems and associated software ecosystem. Recent reports indicate that lawmakers in Congress are actively pushing for even stricter export curbs on Nvidia’s China-bound chips (such as the H20 and H200) and have summoned Huang to testify.  This impending regulatory tightening would further reinforce Nvidia’s China headwinds and boost sales of Huawei’s Ascend chips in China.

While Huang’s departure adds another element of uncertainty to Nvidia’s long-term position in China. Huawei’s opportunity is driven primarily by China’s semiconductor self-sufficiency objectives, domestic procurement policies and U.S. export controls—not simply Nvidia’s leadership transition.  Huawei must still address performance, manufacturing capacity, software compatibility and the ability to support large-scale AI deployments.

………………………………………………………………………………………………………………………………………………………………………………

In addition to the above AI chip suppliers, Amazon, Microsoft and Meta hyperscalers are developing their own custom AI accelerators, while Marvell is an important custom-silicon and connectivity supplier.

………………………………………………………………………………………………………………………………………………………………………………

Major AI chip companies and competitive market positioning:

The following table distinguishes companies selling merchant AI accelerators from those developing custom silicon primarily for their own cloud or AI infrastructure.

Company AI chip / platform Niche and technological strengths Market presence
Nvidia Blackwell, Rubin GPUs; Grace CPUs GPU acceleration, CUDA software, NVLink, InfiniBand, Spectrum-X Ethernet; integrated AI factories
Dominant market leader

Broadest commercial AI accelerator ecosystem

AMD Instinct MI350/MI400; ROCm General-purpose AI GPUs, HBM, chiplet architecture, open software ecosystem
Major challenger

Competing for hyperscaler, enterprise and HPC deployments

Broadcom Custom XPUs, networking ASICs Custom AI accelerators for hyperscalers; high-speed SerDes, Ethernet switching and connectivity
Major competitor

Important supplier to hyperscalers developing custom AI silicon

Google TPU Custom ASICs optimized for AI training and inference; tightly integrated with Google Cloud and AI software
Major alternative

Significant internal deployment and external cloud availability

Huawei Ascend processors; Atlas systems AI accelerators, integrated computing systems and domestic Chinese semiconductor ecosystem
Major China-market player

Important alternative amid U.S. export restrictions

Intel Gaudi; Xeon AI accelerators, CPUs, memory and data-center platforms
Challenger

Established semiconductor presence, but limited AI accelerator traction relative to Nvidia

Cerebras Systems Wafer-Scale Engine (WSE) Wafer-scale integration, massive on-chip parallelism, high-bandwidth memory and specialized AI inference systems
Specialist competitor. Differentiated architecture for large AI workloads. Full-year 2026 core revenue outlook of $940 million to $960 million.
Amazon Trainium, Inferentia Custom ASICs for AI training and inference; integration with AWS infrastructure
Significant captive platform

Expanding AWS alternative to Nvidia

Microsoft Maia Custom AI accelerator architecture for Azure workloads
Emerging platform

Strategic internal development and deployment

Meta MTIA Custom accelerators designed for Meta’s recommendation, ranking and generative AI workloads
Emerging platform

Large internal deployment opportunity

Marvell Custom AI ASICs, optical DSPs, interconnect silicon Custom compute silicon, high-speed connectivity and optical networking for AI data centers
Important infrastructure competitor

Benefits from hyperscaler custom silicon investment

Source: Frontier Tech Intelligence, ChatGPT

Editor’s Note: Not every company listed above is a direct merchant-chip competitor to Nvidia. Some are building captive platforms, while others supply the technology that enables those platforms.

………………………………………………………………………………………………………………………………………………………………………………………………

Top AI Chip Company Beneficiaries & Competitive Positioning:

Supplier Core AI Platform Why They Gain Share Now Primary Bottleneck / Risk
Broadcom Custom XPUs, Tomahawk 6 / Jericho switches Holds 70%+ share of custom AI ASIC design with a $73B+ backlog across Google, Meta, OpenAI, and Anthropic. Captures both custom compute and the Ethernet networking fabric required for non-Nvidia clusters. Custom ASICs require 12–24 month design cycles and are largely limited to hyperscale budgets.
AMD Instinct MI350 / MI400 series, Helios racks, ROCm 7 The only merchant general-purpose GPU alternative capable of drop-in cluster deployment. Offers 30–50% lower TCO and higher HBM capacity per chip for inference workloads. ROCm software maturity and scale-up interconnect bandwidth still trail CUDA and NVLink in frontier training.
Google (TPU) Ironwood (TPU v7) / Trillium (TPU v6) Most mature non-Nvidia production ecosystem. Captures external AI developers through Google Cloud who want immediate compute capacity without waiting on Nvidia’s post-Huang roadmap. Captive to Google Cloud; not sold as standalone merchant silicon for on-premises data centers.
Marvell Custom AI ASICs, 1.6T optical DSPs Co-designs AWS Trainium and Microsoft Maia. Benefits directly as AWS and Azure shift internal inference workloads onto captive silicon to reduce Nvidia exposure. Smaller customer base than Broadcom; remains vulnerable if hyperscalers bring physical design fully in-house.
Cerebras Systems Wafer-Scale Engine (WSE-3) Eliminates off-chip interconnect bottlenecks for ultra-low-latency inference. Recently raised its 2026 core revenue outlook to $940M–$960M following major enterprise/cloud wins like Microsoft. Niche architectural footprint; lacks the general-purpose software ecosystem needed to replace broad GPU fleets.

Source: Google Gemini

………………………………………………………………………………………………………………………………………………………………………………………………..

Will Nvidia Customers Diversify Their AI Accelerator Suppliers?

The key question is whether Nvidia’s leadership transition will materially change customer purchasing decisions.  Nvidia’s board has expressed confidence in the company’s existing management team and technology roadmap. Colette Kress is an experienced executive who has served as CFO since 2013, and the company has emphasized continuity in its operations and customer commitments.  Nevertheless, the loss of Huang’s leadership creates uncertainty that competitors could exploit.

AI infrastructure investments involve long planning cycles, substantial capital expenditures and commitments to particular hardware and software platforms. Customers making multibillion-dollar investments need confidence that their suppliers will maintain technological leadership, product availability, software support and long-term strategic relationships.  Consequently, even if Nvidia’s engineering organization and product roadmap remain intact, some customers could accelerate qualification and deployment of competing accelerators.

This would not necessarily mean abandoning Nvidia. A more likely initial response would be supplier diversification, with AMD GPUs, Google TPUs and other specialized accelerators deployed alongside Nvidia systems where technically and economically appropriate.  The transition toward multi-vendor AI computing could also increase interest in software portability, open programming frameworks and more interoperable data center architectures.

………………………………………………………………………………………………………………………………………………………………………………………………

Conclusions:

Jensen Huang’s resignation does not eliminate Nvidia’s substantial technological advantages, established customer relationships or extensive CUDA software ecosystem. The company retains a strong engineering organization, an experienced executive team and a broad portfolio of AI computing products.

However, Huang’s departure removes the individual who was arguably the most influential executive in the AI infrastructure industry. His combination of technological vision, deal making, industry relationships and strategic leadership was central to Nvidia’s rise and its position at the center of the global AI ecosystem.  The immediate issue is whether Nvidia’s customers and strategic partners will maintain confidence in the company’s future direction. If uncertainty persists, competitors could benefit as hyperscalers, AI developers and enterprise customers seek greater supplier diversity.

AMD has the clearest opportunity to gain external GPU accelerator business, Google can expand its TPU-based AI computing services, and Huawei has an opportunity to strengthen its position in China’s domestic AI accelerator market.

Nvidia’s leadership transition faces immediate geopolitical and financial tests that could accelerate customer diversification. In Washington, lawmakers are pushing new defense-bill export controls—such as the AI Overwatch, Chip Security, and Remote Access Security acts—to tighten China chip curbs and close overseas cloud-rental loopholes while summoning Jensen Huang to testify in December. Meanwhile, interim CEO Colette Kress faces her first major test during Nvidia’s November 17 Q3 FY27 earnings call, where customers and investors will watch closely for reassurance on Vera Rubin schedules and ongoing ecosystem financing commitments.

The ultimate competitive test will not be whether Nvidia can continue selling GPUs without Jensen Huang. It will be whether the company can retain customer confidence, sustain its technology and software leadership, and preserve the strategic ecosystem that he spent 33 years building.

……………………………………………………………………………………………………………………………………………………………………………………………..

Addendum: Two competing approaches to AI-RAN implementation:

1. Nvidia and Nokia: GPU-centric RAN

  • Nvidia and Nokia are promoting GPUs as the computing engine for AI-RAN.

  • Nokia is designing future RAN products around Nvidia GPUs, including execution of demanding Layer 1 (Physical layer) processing functions.

  • The objective is to use accelerated computing for both conventional RAN processing and AI workloads, potentially improving spectral efficiency and enabling shared infrastructure.

  • This represents a departure from Nokia’s previous reliance on Marvell custom silicon for Physical layer processing.

2. Ericsson: Custom ASICs and CPU-based vRAN

  • Ericsson favors its own purpose-built ASICs for demanding Layer 1 RAN processing.

  • Its separate virtual RAN strategy uses Intel x86 CPUs for most Layer 1 functions, with hardware acceleration for forward error correction (FEC).

  • Ericsson is also exploring software portability to AMD and Arm-based CPU platforms.

  • Ericsson argues that its custom silicon offers cost, performance and technological independence advantages over adopting Nvidia GPUs throughout the RAN.
3. Marvell: An important part of the competitive picture

Marvell deserves particular attention. It was Nokia’s custom silicon partner, and its technology remains relevant to existing Nokia RAN deployments. Even Nokia’s AI-RAN trials with Virgin Media O2 reportedly used Marvell Layer 1 silicon rather than Nvidia GPUs.

Attribution: ChatGPT was used extensively for research, technical analysis, and drafting, including an assessment of competing AI-RAN architectures, Nvidia’s GPU-centric approach with Nokia, and alternative custom-silicon and CPU-based RAN strategies pursued by Ericsson and its technology partners. The author reviewed, revised, and fact-checked all the material and is 100% responsible for its content.

………………………………………………………………………………………………………………………………………………………………………………………………………………………….

References:

https://nvidianews.nvidia.com/news/nvidia-announces-ceo-transition

https://www.bloomberg.com/news/features/2026-11-09/nvidia-ceo-jensen-huang-s-affair-with-alleged-spy-raises-succession-concerns [Paywall]

https://www.sec.gov/Archives/edgar/data/1045810/000104581026000082/nvda-20261111.htm

https://www.reuters.com/business/view-nvidia-ceo-jensen-huang-steps-down-2026-11-12/

https://www.gartner.com/en/newsroom/press-releases/2026-07-16-gartner-identifies-the-companies-to-beat-in-the-ai-semiconductor-vendor-race

https://www.amd.com/en/products/accelerators/instinct/mi350.html

https://rocm.docs.amd.com/en/latest/reference/gpu-arch/mi350.html

https://www.oecd.org/en/publications/competition-in-artificial-intelligence-infrastructure_623d1874-en/full-report/component-5.html

https://cloud.google.com/tpu/docs/intro-to-tpu

https://www.broadcom.com/products/ai

https://www.intel.com/content/www/us/en/products/details/processors/ai-accelerators/gaudi.html

https://www.cerebras.ai/chip

https://www.cerebras.ai/blog

https://www.huawei.com/en/technology-insights/industry-insights/outlook/5g

Critical Network Outages: Why Isn’t Redundancy Working? Part II.

Introduction:

Part 1. of this two part series examined the November 13th FAA telecommunications outage at length & briefly described the November 12th Microsoft global cloud disruption. Although the incidents involved different systems and have not been shown to share a common cause, both raise a fundamental engineering question:

Why did existing infrastructure and recovery arrangements fail to maintain essential services?

The answer may involve more than insufficient backup capacity. Network resilience depends on the independence of redundant facilities, the ability to detect and isolate failures, automatic restoration, and realistic testing under adverse conditions.

Redundancy Is Not the Same as Resilience:

A network can have multiple connections and still contain single points of failure. Two fiber routes may share a conduit, bridge crossing, central office or power source. Two data centers may depend on the same control plane, identity service, software release or configuration system.  These are examples of common-mode failures: one event or underlying dependency disables resources that were assumed to be independent.

True route diversity requires physically separate paths, diverse points of presence and, where appropriate, different supporting infrastructure. Cloud resilience similarly requires more than distributing servers across locations. Critical dependencies and management systems must also be considered.

The key question is not how many redundant components an operator owns, but whether those components can continue providing service when the primary system or a shared dependency fails.

A Telecommunications Case Study – Comcast’s Santa Clara, CA Outage:

A July 2024 IEEE Techblog article documented a prolonged Comcast outage in Santa Clara, CA, associated with a physical fiber cut and reported vandalism. Service was unavailable for approximately 17 hours. That incident, which negatively impacted this author, raised a basic question: why wasn’t traffic rerouted automatically around the damaged fiber?

Telecommunications networks have several established protection and restoration approaches. These include 1+1 protection, 1 protection, self-healing rings and mesh-network restoration. Depending on network design, these mechanisms can switch traffic to an alternate path when a link or facility fails.  However, protection mechanisms are effective only when alternate paths are genuinely diverse, sufficient capacity is available, and switching and restoration procedures work as intended. A physically diverse route that shares a vulnerable conduit or aggregation point may provide little real protection.

–>That Comcast outage illustrates why network operators should document not only their redundancy architecture but also the actual restoration performance experienced during outages.

……………………………………………………………………………………………………………………………………………………………………………..

Image Generated by ChatGPT

………………………………………………………………………………………………………………………………………………………………….

Resiliency Recommendation for Network Operators:

Telecommunications providers, cloud operators and organizations supporting critical infrastructure should address several priorities:

  1. Eliminate common-mode failures. Identify shared physical routes, power systems, facilities, software dependencies and control-plane components that could disable supposedly redundant resources.
  2. Provide independent alternate paths. Ensure that backup connectivity is physically and logically diverse, adequately provisioned and available during peak demand.
  3. Test automatic failover. Verify that traffic switches successfully under realistic fault conditions, including fiber cuts, equipment failures, power loss, configuration errors and loss of control-plane functions.
  4. Test recovery under load. A backup system that works in isolation may fail when it suddenly receives the traffic of an entire failed region. Capacity planning must account for this scenario.
  5. Improve fault isolation and graceful degradation. A localized failure should not cascade across unrelated services. Architectures should limit failure domains and preserve essential functions when full service cannot be maintained.
  6. Exercise disaster-recovery procedures. Regularly test recovery plans, including manual intervention, alternate communications, restoration sequencing and coordination among network, cloud and customer operations teams.

Accountability and Operational Transparency:

For critical infrastructure, operators should provide meaningful post-incident reports explaining the failure, its scope, the role of backup systems, why automatic recovery was insufficient, and what corrective actions will prevent recurrence.

Regulators and public-sector agencies should also evaluate resilience based on demonstrated performance—not simply on the existence of documented backup plans or redundant equipment.

The FAA outage is particularly consequential because communications failures can affect the safe operation of the national airspace system. The Microsoft disruption demonstrates how dependence on centralized cloud services can affect organizations across industries and geographic regions.

Conclusions:

Redundancy is a design feature.  It’s an operational result which must be proven to prevent the next network outage from being catastrophic.  Multiple servers, fiber routes, data centers and backup systems provide little assurance unless they are sufficiently independent, have adequate capacity and can maintain essential services when primary resources fail.

The ultimate test of a resilient network is not whether it has backup infrastructure, but whether that infrastructure actually works when needed. Operators must demonstrate this through rigorous failure testing, realistic disaster-recovery exercises, transparent post-incident analysis and measurable improvements in service continuity.

The FAA and Microsoft incidents—and the earlier Comcast outage—should prompt a broader reassessment of whether critical network and cloud infrastructure is as resilient in operation as its architecture claims.

End Quote:

“When people hear ‘copper to fiber,’ upgrade as the backstory to the FAA ground stop, they picture that ‘copper’ as Ethernet cable. But Ethernet over copper doesn’t go past 100 meters, so it never leaves the building. A copper line connecting a remote site over miles is old telecom infrastructure, likely a T1 or something similar, and the equipment on either end is just as old. The risk was never the old copper sitting there working. It’s the moment you connect modern gear to a system that’s been frozen in time since the 1980s.”
— Thomas Doumas, VP of Product, NetAlly

…………………………………………………………………………………………………………………………………………………………………………………

References:

https://www.cisa.gov/sites/default/files/publications/07202017_10_Keys_to_Public_Safety_Network_Resiliency_010418_FINAL508C.pdf

https://www.itu.int/rec/dologin_pub.asp?id=T-REC-L.Sup35-201706-I%21%21PDF-E&lang=e&type=items

https://www.itu.int/rec/dologin_pub.asp?id=T-REC-L.Sup25-201604-I%21%21PDF-E&lang=s&type=items

IETF RFC 9522 – Overview and Principles of Internet Traffic Engineering (2024)
https://www.rfc-editor.org/rfc/rfc9522.html

IETF RFC 9855 – Topology Independent Fast Reroute Using Segment Routing (October 2025)
https://www.rfc-editor.org/rfc/rfc9855.html

NIST SP 800-34 Rev. 1 – Contingency Planning Guide for Federal Information Systems
https://csrc.nist.gov/pubs/sp/800/34/r1/upd1/final

IETF – Problem Statement for Network Resilience (November 2, 2026)
https://datatracker.ietf.org/doc/draft-zhao-opsawg-network-resilience-ps/

Critical Network Outages: Why Isn’t Redundancy Working? Part I.

Telecom network outages: causes, effects, and remedies for telecom providers & IT enterprise

Comcast frequent, intermittent internet outages + long outage in Santa Clara, CA with no auto-recovery!

Ookla: AI platform reliability decreases as outages surge

 

Critical Network Outages: Why Isn’t Redundancy Working? Part I.

Introduction:

Today’s Federal Aviation Administration (FAA) nationwide U.S. flight departure halt, attributed to an unspecified “telecommunications problem,” Microsoft’s global cloud disruption yesterday, and other prolonged telecommunications outages highlight serious questions about the resilience of critical network infrastructure, the effectiveness of automatic failover and service restoration, and the adequacy of disaster-recovery capabilities.

These incidents involved different systems and should not be assumed to share a common technical cause. However, both illustrate the consequences when essential network services become unavailable and recovery is neither immediate nor seamless.

FAA – Telecommunications Failure Disrupts the U.S. National Airspace System:

On November 13, 2026, the FAA ordered a nationwide ground stop for all departing flights following a telecommunications outage affecting air traffic control operations. Transportation Secretary Sean Duffy said the communications issue was impacting all 22 U.S. Air Route Traffic Control Centers (ARTCCs), and FAA technicians were investigating. The FAA’s preliminary statement described the problem as a temporary telecommunications outage and said the ground stop was imposed for safety reasons.

The 22 ARTCCs are a critical part of the U.S. National Airspace System. Unlike airport control towers, which manage aircraft movements on runways and in the immediate vicinity of airports, ARTCCs manage aircraft during the en route portion of their flights, primarily at higher altitudes between airports. They coordinate aircraft separation, traffic flows, handoffs between control sectors and the safe movement of aircraft through the national airspace system.  A communications failure affecting these centers can have consequences well beyond the geographic location where the original problem occurred. ARTCCs depend on reliable communications with aircraft, other control facilities and supporting systems to manage traffic safely. If controllers cannot reliably communicate with aircraft or coordinate traffic handoffs, the FAA may have to reduce traffic volumes or suspend departures to maintain safe operating conditions.

The disruption was substantial. According to FlightAware, more than 1,000 flights were canceled and nearly 10,000 delayed. Reuters reported that 4,042 flights were airborne and destined for U.S. airports when the FAA issued the ground stop. The disruption also affected international operations, forcing cancellations of some flights originating in Canada and Europe.

FAA Flight Control Center

…………………………………………………………………………………………………………………………………………..

The nationwide ground stop prevented additional aircraft from departing while the FAA investigated the problem and worked to restore communications. Thousands of aircraft already airborne were allowed to land, while flights still on the ground could not take off. Inbound international flights were to be handled on a case-by-case basis. The FAA was not ordering aircraft already in the air to land immediately; rather, it was restricting new departures to avoid adding traffic to an airspace system experiencing serious communications problems.

The FAA’s action was only the second nationwide ground stop of departing flights since the September 11, 2001 attacks. The previous occurrence was in January 2023, when the FAA halted domestic departures for less than two hours following a failure of its NOTAM (Notice to Air Missions) system. That incident involved a different system and failure scenario and should not be confused with the November 2026 telecommunications outage.

The Critical Question: Why Didn’t Redundancy Prevent a Nationwide Disruption?

The central engineering issue is not simply why the telecommunications system failed, but why the failure resulted in a nationwide ground stop. The incident raises several important questions:

  • What telecommunications infrastructure and services were affected, and how did the problem propagate across the ARTCC network?
  • Were independent primary and alternate communications routes available to the affected centers?
  • Did the outage involve a common network, switching facility, software platform, power source or other shared dependency?
  • What backup communications systems and failover procedures were in place, and why were they unable to maintain sufficient service to avoid a nationwide suspension of departures?
  • Were alternate facilities or operating procedures available to isolate the failure and maintain at least limited air traffic operations?

These questions cannot be answered conclusively until the FAA provides a detailed technical explanation. As of the initial reports, the precise cause had not been established publicly. Homeland Security Secretary Markwayne Mullin said there was no indication at that time of a malicious attack, while emphasizing that the cause remained under investigation. The possibility of a problem involving an update to the private communications network supporting FAA operations was raised in subsequent reporting, but should be treated as preliminary until confirmed by the FAA and its contractors.

FAA Modernization: Acknowledged Problems, but Resilience Remains Essential

The November 13 incident occurred as the FAA is undertaking a major modernization of its aging air traffic control infrastructure. The Brand New Air Traffic Control System (BNATCS) initiative includes replacing legacy telecommunications infrastructure, modernizing voice switches and radio systems, and upgrading other critical air traffic control technologies.

A September 2026 Government Accountability Office (GAO) report, Air Traffic Control Systems: Ambitious New Modernization Effort Needs to Improve Cost and Schedule Planning (GAO-26-107992), describes both the scale of this effort and the challenges facing the FAA. By May 2026, the agency had replaced 2,560 of 5,170 aging copper communications wires with high-speed fiber-optic cables—approximately half of the planned replacements. The report also identifies significant cost-estimation and project-scheduling weaknesses that could delay modernization and prolong reliance on outdated systems.

Modernization is clearly necessary, but replacing aging copper lines with fiber and upgrading equipment are not sufficient by themselves to guarantee network resilience. The architecture must also provide physically and logically diverse communications paths, independent backup facilities, automatic failover, effective fault isolation and tested disaster-recovery procedures.

The FAA’s previously implemented backup arrangements for its NOTAM system should not be confused with backup communications for the ARTCC network. They address different systems and failure scenarios. The November 13 incident requires a separate examination of the communications architecture, route diversity, failover procedures and operational readiness supporting all 22 centers.

Ultimately, the question is whether the FAA’s communications infrastructure was engineered to withstand the failure of individual links, facilities or shared network components—and, if so, why those protections were insufficient to prevent a nationwide ground stop.

As of the initial reports, the precise technical cause had not been established publicly. Homeland Security Secretary Markwayne Mullin said there was no indication at that time of a malicious attack, while emphasizing that the cause was still under investigation. Subsequent reporting raised the possibility of a problem involving an update to the private communications network supporting FAA operations, but that explanation should be treated as preliminary until confirmed by the FAA and its contractors.

The FAA has been undertaking a major modernization of its aging air traffic control infrastructure, including replacing older copper communications lines with fiber and upgrading radio systems. The November 13 incident underscores why modernization must address not just equipment age and capacity, but also network architecture, physical and logical route diversity, automatic failover, fault isolation and disaster recovery.

The FAA has been working to modernize aging air traffic control infrastructure. A September 2026 Government Accountability Office report, Air Traffic Control Systems: Ambitious New Modernization Effort Needs to Improve Cost and Schedule Planning (GAO-26-107992), describes the scale of the modernization challenge and the risks of continued reliance on outdated systems.

The FAA’s previously implemented backup arrangements for its NOTAM system should not be confused with backup communications for the regional air traffic control centers. They address different systems and failure scenarios. The November 13th incident requires a separate examination of communications architecture, route diversity, failover procedures and operational readiness.

Microsoft: A Global Cloud-Service Disruption:

A day earlier, on November 12, Microsoft experienced a widespread disruption affecting Microsoft 365 services, including Exchange Online, Outlook and other applications dependent on its cloud infrastructure.  Microsoft reported that a significant number of servers had become unexpectedly unavailable. The remaining infrastructure experienced increased load, complicating recovery as affected servers were returned to service. The disruption continued across regions and affected businesses, government organizations and other customers.  This raises a different but equally important question: how could a cloud platform operating at Microsoft’s scale lose enough server capacity to cause prolonged disruption across multiple services and regions?

Cloud infrastructure is designed to distribute workloads across servers, availability zones and geographic regions. But distributing workloads does not automatically guarantee resilience. Shared software, configuration changes, dependencies, authentication services, traffic-management systems or common operational procedures can create failure domains that extend well beyond an individual server or data center.  Microsoft’s incident also demonstrates that restoring infrastructure is not necessarily the same as restoring service. Recovery must account for traffic redistribution, overloaded surviving systems, dependency failures and the possibility that returning capacity to service may itself destabilize the environment.

The Common Question: Where Was the Effective Backup?

The FAA and Microsoft incidents are not evidence of the same underlying failure. They do, however, expose a common issue: the difference between having redundant infrastructure and maintaining uninterrupted service when something fails.  A network may have multiple links, servers, data centers or geographic locations and still be vulnerable if those resources share a common failure point. Backup capacity is useful only if it is sufficiently independent, operationally available and capable of taking over under real-world failure conditions.

…………………………………………………………………………………………………………………………………………………………………………………………….

Part 2 will examine why redundancy can fail, the engineering principles required for genuine network resilience, and what telecommunications providers, cloud operators and critical-infrastructure organizations should do to improve service continuity.

…………………………………………………………………………………………………………………………………………………………………………………………….

References:

https://www.faa.gov/newsroom/statements/general-statements 

https://www.gao.gov/products/gao-26-107992

https://www.washingtonpost.com/transportation/2026/11/13/faa-grounds-flights-all-us-airports-citing-communications-issue/ [paywall]

https://www.reuters.com/world/us/faa-issues-national-ground-stop-due-extraordinary-circumstance-reasons-safety-2026-11-13/ [paywall]

Telecom network outages: causes, effects, and remedies for telecom providers & IT enterprise

Ookla: AI platform reliability decreases as outages surge

 

MWC26 Doha Recap: Strong Turnout; GSMA-ITU Letter of Intent; Open Questions Remain

Introduction:

MWC26 Doha closed on Nov 10th with over 10,000 attendees from 100 countries and territories, 195 speakers, about 100 exhibitors, sponsors and partners, and 42 partnerships and announcements signed in the Announcement Lounge. It was also the first MWC to run alongside an ITU Plenipotentiary Conference (PP-26).  Held every four years, ITU PP-26 serves as the highest decision-making body to determine ITU’s strategic and financial plans, elect its leadership, and set its overall direction. Together, the two events united industry leaders, policymakers and innovators to explore critical digital issues transforming society.

Sianne Ryder, CEO, Events and Industry Services, GSMA Ltd, said:

“This week at MWC26 Doha, policymakers and regulators sat side by side with industry leaders and innovators to debate, connect and shape MENA’s digital future. Co-locating with the opening days of ITU PP-26 created something unique: a place where those setting the global digital agenda could engage directly with the operators, technology companies and investors bringing it to life. The result was sharper conversations, stronger connections and a clearer path from policy to progress. We look forward to seeing the partnerships forged in Doha deliver impact across the region and beyond. Thank you to everyone who joined us – and especially to our hosts at MCIT.”

Photo Credit: GSMA

The GSMA’s closing press release is full of impressive statistics and signed agreements. This IEEE Techblog post separates what was announced from what was demonstrated, and asks what evidence is still missing.

………………………………………………………………………………………………………………………………………………………………………………………………………………………………………..

Open, trusted AI for telecoms”: a non-binding letter and unproven models

The GSMA-ITU Letter of Intent

The most widely publicized conference deliverable is the GSMA-ITU letter of intent on “trusted, open AI for telecoms.”  According to RCR Wireless, it is non-binding, and neither organization has disclosed funding, staffing, deadlines, governance or licensing terms.  RCR reports that the framework builds on the GSMA’s Open Telco AI program. It covers telco-grade AI for network operations and standards, telecom-specific data assets, and training, deploying and evaluating the OTel open model family. RCR frames the underlying motivation as reducing operators’ dependence on a few large AI providers.

The letter also formalizes existing ties. The ITU already hosts AI challenges developed with the GSMA, ETSI, IEEE and others, and the letter sits alongside, but separate from, the ITU’s OTAI initiative on AI agent interoperability. How the two efforts will be coordinated is not stated.

The first workstream covers telco-grade AI for network operations, infrastructure and standards. The second covers telecom-specific data assets, and the third covers training, deploying and evaluating the OTel open model family. They also plan to align technical standards, open data interoperability and evaluation frameworks. The stated aim is “secure, transparent and high-precision AI across global networks.”  The wrap-up MWC26 Doha press release gives no performance data, datasets or timelines for this GSMA-ITU collaboration.

GSMA also announced a collaboration with Khalifa University and AMD on radio frequency foundation models.  Built around the GSMA’s Open Telco AI program, it reaffirms that network operators shouldn’t depend on a handful of large providers for AI intelligence.

Open Telco AI models: what the evidence shows

The models behind the initiative are where measurable claims belong. For OTel, RCR notes that OTel 2.0 passed 30 million downloads by October, though downloads don’t measure deployment. It also reports that AT&T’s roughly 90% benchmark figure is vendor-reported, the OTel paper describes the resource as English-only and mainly text-based, and its main evaluations use unseen portions of its own corpus rather than fully independent benchmarks. “Open-weight” means the weights can be downloaded and adapted, which is not the same as open source. GSMA and ITU sign letter of intent on telecom AI +2

RFGPT. The most technical announcement at the show was an operator evaluation initiative for RFGPT from Khalifa University, GSMA and AMD. RFGPT is an RF foundation model that combines wireless signal understanding with natural-language interaction, hosted on AMD compute with API access. Operators can request a two-week trial. The six evaluation areas include identifying technologies such as Wi-Fi, Bluetooth, LTE, 5G NR, UMTS and DVB-S2, recognizing modulation schemes, analyzing wideband multi-signal environments, assessing coexistence and interference, estimating WLAN and NR user density, and interpreting 5G NR parameters such as SSB patterns and CSI-RS and SRS configurations.

Khalifa University, GSMA and AMD Collaborate to Bring Radio Frequency Foundation Models to Telecom Operators  – Newsroom +2

The release describes capabilities and potential uses. It gives no accuracy figures, training-data description, signal-to-noise conditions, or comparison with conventional RF analysis tools. Useful questions for the evaluation include how it compares with established signal classifiers, how often it gives confident wrong answers, and what operator data leaves the network when a hosted API is used.

“Open” and “trusted,” in a geopolitical context

The AI collaboration is unfolding amid a policy dispute. RCR notes that Open Telco AI’s participants include China Telecom, China Unicom and China Mobile, and that Huawei is a sponsor and data contributor. Huawei was also the event’s headline sponsor and gave a keynote titled “The Infrastructure of Intelligence.”

Meanwhile, PP-26 continues. The conference runs about three weeks, with delegates from ITU’s 194 Member States, so MWC overlapped only its opening. A CircleID analysis describes a US proposal on ITU Resolution 214 that would replace “artificial intelligence” with “super intelligence,” remove references to shared concerns, and prohibit ITU work supporting AI governance, regulation and ethics. The outcome isn’t known yet.  Separately, Malaysia’s communications minister described AI as creating a “compute divide” between developed and developing nations.  “Trusted” will be easier to assess once governance, licensing and supply-chain terms are published.

ITU meeting in Qatar to address digital opportunities and challenges – Developing Telecoms +2

Agentic AI: governance language, no deliverables yet

The Agentic AI Summit description says it would examine the balance between autonomy, trust, safety and human oversight. That describes a topic rather than an outcome. For network operations, autonomy claims are easier to evaluate with a stated autonomy level, the actions agents may take, how policy is enforced, and failure behavior. I did not find those details in the materials reviewed.

6G: what “already underway” means

The Future Networks Summit description says the 6G transition is already underway through intelligent, programmable and interoperable 5G networks. Network APIs, automation, edge, private networks and satellite integration are real and progressing, but they are 5G-Advanced work. Readers should look for which specific features, and in which 3GPP release, are meant when “6G” is used.

Satellite and NTN: organizer statistics

The Satellite & NTN Summit description says 2026 is “go time” for D2D, the Middle East has 11 D2D partnerships between satellite companies and telcos, and 500-600 million people have no coverage. It also cites consolidation moves, including Amazon/Globalstar/Viasat and Space42/Equatys. [verify each; the page gives no sources] A partnership is not yet a commercial service. Spectrum, device support and regulatory approval will determine which ones become products. (For the U.S. side of this story, see my earlier posts on SpaceX’s 800 MHz spectrum and Starlink Mobile.)

One concrete demo

Ooredoo Qatar, Ericsson, Eonetics, Hikvision and GSMA Fusion linked a tethered drone in Melbourne with the GSMA Pavilion in Doha, more than 12,000 km away. The distance makes a good headline, but long-haul IP connectivity is not a property of 5G itself. The release gives no latency, throughput, uplink performance, drone-to-core architecture or setup time, so the demo shows feasibility rather than performance.

Independent coverage

Post-event coverage so far largely restates the GSMA release. Examples include Fast Company Middle East, IndexBox and TelecomTV, which labels its item “Via GSMA.” Huawei’s account is a vendor release. Reporting on measured performance, stage discussion and attendee reaction was limited in what I could find.

Evidence we’d like to see

  • The full text of the GSMA-ITU letter, with scope, governance and licensing.
  • Independent benchmarks for OTel and RFGPT against conventional tools.
  • Latency and throughput data from the emergency-response demo.
  • Operator evidence of agentic AI in production, including guardrails.
  • The 3GPP features behind “6G is already underway.”
  • The final outcome on ITU Resolution 214 at PP-26.

Much of the post-event coverage repeats the GSMA’s numbers. One Gulf Business piece carries “brand-view” in its URL path, and the Huawei item is a vendor announcement. Independent reporting, such as measured performance, attendee skepticism and what was said on stage, was scarce in what I could find.

End Note:

GSMA, ITU and IEEE will host a webinar on November 18 introducing the AI Agents for RAN Challenge, which uses the Giulia simulator.

……………………………………………………………………………………………………………………………………………………..

References:

GSMA MWC26 Doha welcomed over 10,000 attendees, as world-first ITU co-location puts MENA at the centre of global connectivity policy and innovation 

https://www.mwcdoha.com/agenda

https://rcrwireless.com/20261112/ai/gsma-itu-letter-of-intent-telecom-ai

https://www.developingtelecoms.com/telecom-business/telecom-regulation/21107-itu-meeting-in-qatar-to-address-digital-opportunities-and-challenges.html

Khalifa University, GSMA and AMD Collaborate to Bring Radio Frequency Foundation Models to Telecom Operators 

 

Analysis: Taara Demonstrates 14-km Free-Space Optical Link with AT&T and Nokia

Introduction:

Taara, an independent private company spun out of X, Google’s Moonshot Factory, and a developer of high-speed, high-capacity wireless optical communications, announced an expansion of its commercial activities targeting enterprise and federal government customers in the U.S.  IEEE Techblog has written about the company previously (see References below).

Following a successful deployment with AT&T at the world’s largest soccer event at at MetLife Stadium, NJ, Taara demonstrated its Lightbridge Free Space Optics (FSO) system at Texas Tech University. The demonstration illustrated how narrow, invisible beams of light can deliver high-capacity connectivity where rapid deployment, security and operational flexibility are important for mission-critical applications.

Conducted in collaboration with AT&T and Nokia Federal Solutions, the demonstration carried data across a 14-km path between a university rooftop and a mobile cell site, known as a cell on wheels (COW). The demonstration was presented to enterprise, state and federal government stakeholders as part of Texas Tech University’s Critical Infrastructure Telecom Ecosystem Incubator.

The significance is that high-capacity optical connectivity can be established without installing fiber along the entire route or obtaining licensed radio-frequency spectrum. However, free-space optical communications have distinctive propagation and availability limitations that must be considered when evaluating their use in telecommunications infrastructure.

Free-Space Optical (FSO) Communications Explained:

Free-space optical communications transmit data using highly focused beams of invisible near-infrared light through the atmosphere rather than through a physical optical fiber.  The basic architecture consists of two optical terminals, each containing an optical transmitter and receiver. The transmitting terminal converts electrical data into modulated optical signals, which are directed toward the receiving terminal. A photodetector at the receiving end converts the optical signal back into electrical data.  Unlike conventional fiber-optic communications, the optical signal travels through an unguided atmospheric path. This eliminates the need to install physical cable between the two endpoints while retaining the potential for very high data rates and low latency.

Taara Lightbridge supports bidirectional transmission of up to 20 Gbps over distances of up to 20 km, according to the company’s published specifications. Its lightweight terminals can be installed on rooftops, towers and other elevated structures with a clear line of sight.

Image Credit: Communications Physics article “Towards high-capacity and robust free-space optical communications,” by Jian Wang, Jinfeng Liu & Chengkun Cai

……………………………………………………………………………………………………………………………………………………………………………..

FSO technology is particularly attractive for:

  • Wireless backhaul and fronthaul connectivity for cellular networks.
  • Temporary high-capacity connections for major events and emergency deployments.
  • Enterprise and campus network extensions.
  • Connectivity between data centers and other facilities.
  • Physically diverse network paths that reduce dependence on a single fiber route.

AT&T’s 80,000-Spectator Deployment at MetLife Stadium in New Jersey:

An important commercial example involved AT&T’s deployment of two Taara Lightbridge links at MetLife Stadium in New Jersey. The installation connected more than 80,000 spectators during a major international soccer event.  According to Taara, the links were installed in a matter of hours and delivered up to 20 Gbps during the event.

Importantly, the deployment combined the optical links with a millimeter-wave radio backup. Taara reported 99.999% availability for the combined hybrid connection.  This distinction is important. The five-nines availability figure applies to the combined optical and millimeter-wave arrangement, not to the standalone optical link. The deployment illustrates the value of combining different transmission technologies to improve network resilience.  AT&T is also evaluating Taara links for broader network applications.

Advantages of FSO Compared with Fiber and Radio:

FSO offers several potentially important advantages for telecommunications operators and enterprise network providers.

Rapid deployment: Unlike fiber construction, FSO does not require trenching, cable installation, or lengthy rights-of-way negotiations along the transmission path. This can be particularly valuable for temporary installations, disaster recovery and locations where fiber construction is difficult or expensive.

High capacity: Optical wavelengths provide substantial bandwidth potential. Taara specifies up to 20 Gbps bidirectional throughput for Lightbridge, making it relevant to high-capacity wireless backhaul applications.

No licensed RF spectrum: Because the transmission uses optical rather than radio-frequency signals, FSO does not require licensed RF spectrum. This can be advantageous in congested wireless environments.

Low latency: Optical propagation through air is close to the speed of light in vacuum, avoiding the additional propagation delay associated with optical fiber. Actual end-to-end latency also depends on terminal electronics and network equipment.

Physical network diversity: An FSO link can provide an alternative physical path alongside fiber infrastructure. This can reduce dependence on a single cable route, provided that the optical path itself is sufficiently protected and does not share the same physical vulnerabilities.

FSO’s Risks and Technical Limitations:

Despite its considerable potential, FSO is not a universal replacement for fiber or conventional wireless transmission. Its principal weakness is that the optical signal must propagate through an uncontrolled atmospheric environment.  Taara’s own specifications acknowledge 90% average availability at 10 km for its Lightbridge system. That is a substantial limitation for a standalone carrier connection.  Several engineering challenges affect link reliability and availability:

1. Atmospheric attenuation and weather

Fog is one of the most serious challenges for terrestrial FSO. Water droplets scatter and attenuate near-infrared optical signals, potentially reducing received optical power below the level required for reliable communications.

Heavy rain, snow, dust and other atmospheric conditions can also degrade performance. Atmospheric turbulence can cause fluctuations in signal intensity, beam distortion and variations in received optical power.

Taara’s own technical FAQ acknowledges these limitations. Its published averages indicate approximately 95% annual availability at 5 km and 90% at 10 km for Lightbridge links. The company also reports that fog causes significantly greater attenuation than is typically experienced by millimeter-wave systems.

These figures demonstrate why advertised maximum transmission distance should not be confused with guaranteed service availability.

2. Line-of-sight requirements and physical obstruction

FSO requires a clear optical path between the two terminals. Buildings, trees, construction equipment, vehicles and other physical obstructions can interrupt the connection.

Unlike radio systems that may support non-line-of-sight propagation through reflection or diffraction, a highly directional optical link generally cannot maintain its connection when the direct path is blocked.

Consequently, careful site surveys, rooftop placement and protection of the optical path are essential.

3. Pointing, acquisition and tracking

Because the optical beam is highly directional, the transmitting and receiving terminals must remain accurately aligned.

Wind-induced movement of towers, rooftop structures, mounting equipment and other mechanical disturbances can cause pointing errors. Building movement and thermal expansion can also affect alignment.

Automatic acquisition and tracking mechanisms are therefore important components of practical FSO systems. Maintaining alignment over long distances is a significant engineering challenge.

4. Availability and network redundancy

FSO should not automatically be treated as a fiber replacement for applications requiring continuous connectivity.

Operators need to engineer around atmospheric outages and physical obstructions through redundant paths, adaptive transmission rates, backup radio links or alternative fiber routes.

Taara’s AT&T stadium deployment provides a useful example: combining FSO with millimeter-wave radio backup allowed the network to achieve the vendor-reported five-nines availability figure.

For mission-critical applications, the relevant engineering metric is not simply peak throughput. It is sustained throughput and service availability under actual operating conditions.

5. Security considerations

The narrow, highly directional optical beam can make interception more difficult than with broadly propagated wireless signals. FSO also avoids the licensed RF spectrum and radio-frequency interference issues associated with conventional wireless links.

However, optical transmission should not be equated with inherently secure communications. Physical access to the optical path, endpoint compromise, deliberate obstruction and attacks on network equipment remain relevant security considerations.

Encryption, authentication, access controls and network monitoring are still necessary.

Taara’s Next Technology Development: Integrated Photonics and Optical Phased Arrays:

Beyond its deployed Lightbridge technology, Taara is developing a more compact integrated-photonics platform called Taara Photonics.

Introduced in February 2026, the platform moves important optical communication and beam-control functions into a small silicon-based photonic module. Its central technology is an optical phased array (OPA) containing more than 1,000 miniature, software-controlled light emitters.

By electronically controlling the phase of the emitted light, an OPA can steer and track optical beams without relying on conventional moving mechanical components.

Taara’s first product based on this platform, Taara Beam, is specified for up to 25 Gbps bidirectional throughput over distances of up to 10 km.  The approach could eventually reduce terminal size, complexity and deployment costs while enabling more flexible optical wireless network architectures.

However, integrated photonics and electronic beam steering do not eliminate the fundamental atmospheric propagation challenges of FSO. Fog attenuation, turbulence, physical obstructions and link availability must still be addressed.

It is also important to distinguish this newer platform from the Lightbridge equipment used in the 14-km Texas demonstration. The Texas demonstration does not establish the operational performance or weather resilience of Taara Beam.

Wireless Telecom Industry Implications:

Taara’s demonstrations illustrate the potential role of wireless optical communications as another component of high-capacity telecommunications infrastructure.

FSO is particularly attractive where fiber construction is prohibitively expensive, deployment schedules are demanding, or an additional physically diverse connection is needed. It can complement fiber and millimeter-wave radio rather than compete with either technology in every application.

The technology could also become relevant to data-center interconnection, where increasing traffic volumes and the need for additional network capacity are creating demand for alternative connectivity solutions. However, practical deployment will depend on distance, atmospheric conditions, site geometry, redundancy requirements and total cost of ownership. The most important consideration for network operators is whether FSO can deliver sufficient sustained capacity and availability at a competitive cost under the specific environmental conditions of each deployment.

The company’s hybrid approach is probably the correct commercial direction, but it also suggests that FSO’s role is to complement fiber and RF, not replace them.

Conclusions:

Taara’s 14-km demonstration with AT&T and Nokia illustrates the growing potential of free-space optical communications to provide high-capacity connectivity without the construction costs, delays and physical constraints of fiber deployment. Taara appears to have made meaningful progress in equipment design, photonics integration, link planning and network redundancy. Its deployments deserve serious attention, rather than dismissal based on the failures of the earlier FSO industry from 1998-2002.

FSO is particularly attractive for wireless backhaul, temporary high-capacity links, enterprise connectivity and network-route diversity. However, impressive peak data rates and transmission distances do not establish carrier-grade reliability. Atmospheric attenuation—especially from fog—line-of-sight obstructions and precise beam-alignment requirements remain significant engineering challenges.

Consequently, FSO is best deployed with appropriate redundancy, including fiber or millimeter-wave backup where uninterrupted service is required. Taara’s newer integrated-photonics and optical phased-array development could make optical wireless terminals smaller, less mechanically complex and easier to deploy. But it does not eliminate the fundamental limitations of atmospheric optical transmission.

We believe that FSO is a promising complementary technology for expanding high-capacity network connectivity, not a universal replacement for fiber. Its commercial success will depend on demonstrating sustained throughput, predictable availability and competitive total cost of ownership under real-world operating conditions.

……………………………………………………………………………………………………………………….

References:

https://www.taaraconnect.com/post/taara-expands-in-texas-after-connecting-80-000-fans-with-at-t-at-the-worlds-biggest-soccer-stage

https://x.company/blog/posts/taara-graduation/

https://www.taaraconnect.com/post/introducing-taara-photonics-and-beam

https://support.taaraconnect.com/en/faq/design-specs

https://www.nature.com/articles/s42005-026-02947-3

Taara Lightbridge Pro: an ultra reliable wireless optical communications system for 5G mobile backhaul

Google X spin-out Taara and Digicomm International partner to offer high speed wireless communications

2021 Optical Fiber Conference (OFC) Highlights and Links to Videos

 

Broadband Forum Advances Subscriber Session Steering for Dynamic BNG Resource Optimization

Attribution:

Ken Pyle, Broadband Forum Marketing Director, provided the author with references which he used to write this article.  ChatGPT helped with organizing the piece AFTER the author completed his own diligent research.

Introduction:

As broadband networks evolve toward disaggregated, cloud-native architectures, service providers need more flexible ways to manage subscriber sessions and allocate network resources. Traditional broadband architectures typically associate subscribers with specific Broadband Network Gateways (BNGs), limiting operators’ ability to dynamically balance traffic, optimize resource utilization, or respond to changing service requirements.

The Broadband Forum (BBF) is addressing this challenge through Subscriber Session Steering (3S), an architecture that enables service providers to dynamically select and manage the network gateways serving individual subscribers or groups of subscribers. The objective is to optimize network performance, improve resource utilization, reduce energy consumption, and support differentiated broadband services without interrupting active subscriber sessions.

At its November 2–5, 2026 Fall Member Meeting in Denver, CO the Broadband Forum’s Network Architecture Work Area approved WT-474, Subscriber Session Steering, and advanced it to Final Ballot. Publication as Technical Report TR-474 is expected before year-end 2026.

Subscriber Session Steering Explained:

Subscriber Session Steering provides dynamic, policy-driven selection and management of subscriber sessions across available service gateways. These include BNGs and Access Gateway Functions (AGFs).

Unlike static subscriber-to-gateway assignments, 3S enables individual subscribers or entire subscriber groups to be moved between network resources based on traffic load, service requirements, Quality of Service (QoS), network availability, maintenance needs, and energy consumption.

The architecture is designed to perform these changes without interrupting ongoing subscriber sessions.

Two important functional elements are:

  • User Plane Selection Function (UPSF): Determines the appropriate BNG user-plane instance for a subscriber or subscriber group based on network conditions and service-provider policies. It provides the selection and coordination function for subscriber-session placement.
  • Traffic Steering Function (TSF): Implements the traffic redirection needed to move subscriber sessions and associated traffic flows to the selected user-plane instance. The TSF can be deployed at different network locations, including within an Optical Line Terminal (OLT).

Together, these functions enable dynamic binding of subscriber sessions to available BNG user planes. This is particularly relevant to Control and User Plane Separation (CUPS), which allows BNG control-plane functions to manage multiple independently deployed user-plane instances.

The architecture also supports cloud-native BNG implementations in which user-plane functions can be deployed, scaled, consolidated, or deactivated as network demand changes.

2024 Broadband Forum Innovation Demonstration: Multi-Vendor Subscriber Session Steering:

The technical feasibility and operational benefits of 3S were demonstrated during the Broadband Forum’s November 2024 Innovation Demonstrations, involving three vendors: Adtran, BISDN, and OutSys.

The demonstration combined disaggregated BNG architecture, cloud-native user planes, traffic steering, and Quality of Experience (QoE) monitoring to dynamically optimize network resource utilization.

Key contributions included:

  • BISDN: Supplied BISDN Linux, an open-source network operating system running on the aggregation switch, and the UPSF based on the Broadband Forum’s Open Broadband Subscriber Session Steering (OB-STEER) project. Its BNG micro-user planes were cloud-native and deployable in Kubernetes-based x86 infrastructure.
  • Adtran: Integrated the TSF into its SDX 6330 PON OLT. Under UPSF direction, the TSF dynamically redirected subscriber sessions and associated traffic flows between BNG user-plane instances. Locating traffic steering in the OLT enables redirection closer to the access network, including when an OLT uplink or aggregation switch fails.
  • OutSys: Provided Quality of Experience Delivered (QED) probes and reflectors, based on Broadband Forum TR-452, to monitor service quality across BNG user planes. Its QED Analyzer, integrated into Nokia’s Altiplano framework, used network measurements and subscriber counts to support closed-loop resource optimization.

The demonstration illustrated an energy-saving scenario. When traffic and subscriber counts on a particular BNG user plane fell below a predefined threshold, and measured service quality remained satisfactory, the automation system moved active subscribers to another user plane. The underutilized instance could then be shut down.

This approach reduces the number of running BNG micro-user planes and their associated energy consumption in x86-based Kubernetes clusters while maintaining acceptable subscriber service quality.

The demonstration also showed the resilience benefits of OLT-based steering. If an uplink or aggregation switch fails, subscriber traffic can be redirected through another OLT uplink toward a different network path.

Importantly, the demonstration brought together components from three vendors using Broadband Forum standards and Open Broadband software, illustrating the potential for multi-vendor implementations.

WT-474: The New Subscriber Session Steering Technical Report:

The Broadband Forum’s WT-474 project formalizes the 3S architecture and its operational requirements.  According to the Forum’s September 2026 technical overview, TR-474 will define:

  • Subscriber Session Steering use cases.
  • The overall 3S architecture and functional requirements.
  • Operational workflows for dynamic subscriber-session selection and movement.
  • Informational models describing the entities and relationships involved.
  • Data models needed to support implementation and management of the architecture.

The specification addresses the selection, balancing, and management of subscriber-session mappings to service gateways. It is intended to support dynamic network optimization based on access conditions, service requirements, and operational events.

An important aspect is the ability to steer individual subscribers or groups of subscribers without interrupting active sessions. This supports service-aware network resource allocation rather than treating all broadband subscribers identically.

The forthcoming TR-474 [1.] is also designed to work independently or in conjunction with other Broadband Forum disaggregation and cloud specifications.

The related Open Broadband OB-STEER project provides reference software implementations of key architectural components, including UPSF-related functions. Its purpose is to support rapid prototyping, evaluation of specification changes, and potential interoperability testing as commercial implementations emerge.

Note 1. At the Fall 2026 BBF Member Meeting, WT-474 was approved and sent out for Final Ballot. TR-474 is expected to be approved before the end of 2026.  Once approved, it will be sent to ITU-T for consideration as a new recommendation.

Broadband Industry Significance:

Subscriber Session Steering extends broadband network resource management beyond conventional load balancing. Its significance spans several areas:

Energy efficiency: Operators can consolidate subscriber sessions onto fewer active BNG user-plane instances during periods of lower demand, allowing underutilized resources to be deactivated. This can reduce power consumption in cloud-native broadband infrastructure.

QoS and differentiated services: Subscribers can be assigned to network gateways and paths that better match their service requirements. High-value customers and latency-sensitive applications, including gaming and cloud services, can potentially be directed toward higher-capability network resources.

Wholesale and open-access broadband: 3S enables a Broadband Service Provider’s subscribers to be moved to different portions of a wholesale network. This flexibility can support more efficient use of shared infrastructure and new wholesale service models.

Network resilience and maintenance: Subscriber sessions can be redirected in response to equipment failures, network congestion, or planned maintenance. Dynamic placement also provides opportunities for predictive maintenance and more flexible network upgrades.

Cloud-native and disaggregated BNGs: By supporting dynamic subscriber placement across independently deployed user planes, 3S complements CUPS and cloud-native BNG architectures. Operators gain greater flexibility in scaling network functions and utilizing distributed computing resources.

Multi-vendor interoperability: Standardized architecture, requirements, workflows, and information and data models provide a common basis for implementations from different suppliers. The 2024 demonstration involving Adtran, BISDN, and OutSys provided an early example of how these components can work together.

…………………………………………………………………………………………………………………………………………………………………………………………………………………………..

From Broadband Forum Marketing Director Ken Pyle:

“Another important point is that 3S enables an access provider to automate and adapt the path to wholesale handoffs while preserving a clear operational boundary: subscriber steering is managed locally, while the wholesale partner controls service termination beyond the handoff.”

……………………………………………………………………………………………………………………………………………………………………………………………

Conclusions:

The Broadband Forum’s Subscriber Session Steering initiative addresses an important requirement for the next generation of broadband networks: dynamically matching subscriber sessions to available network resources according to service needs, network conditions, and operational objectives.

By formalizing the 3S architecture in TR-474, the Broadband Forum is establishing a framework for more flexible BNG resource allocation, energy-efficient operations, improved resilience, and differentiated broadband services.

The 2024 multi-vendor demonstration showed how UPSF, OLT-based traffic steering, cloud-native BNG user planes, and QoE monitoring can work together to achieve these objectives. The forthcoming TR-474 will provide the technical requirements, architecture, workflows, and models needed to advance this approach toward interoperable industry implementations.

…………………………………………………………………………………………………………………………………………………………………………

References:

https://www.broadband-forum.org/events/fall-2026-member-meeting-roundup

https://www.broadband-forum.org/news-and-insights/bb-insights/steer-smarter-deliver-better-3s-overview

https://www.broadband-forum.org/video-insights/subscriber-session-steering-for-energy-savings-more-broadband-forum-2024-innovation-demonstrations

https://broadband-forum.atlassian.net/wiki/spaces/BBF/pages/46649496/Open+Broadband+-%2BSubscriber+Session+Steering+OB-STEER

Broadband Forum new work areas to enable broadband services & apps

Broadband Forum SDN/NFV Work Area launches four new projects at Bangkok Q4-2023 meeting

Broadband Forum and LAN Laboratory Expand Certification Program to include XGS-PON

Broadband Forum’s vBG network spec targeted at SD-WANs; led by ONUG

Inside TM Forum’s Catalyst project “Living Networks – Phase III”

Page 1 of 376
1 2 3 … 376