2026 TPI Aspen Forum: challenges and risks of scaling AI, managing power infrastructure and permitting
-
- Siting Constraints: Speakers noted that power availability has become the primary bottleneck for AI data center expansion, dictating where new infrastructure can realistically be built.
- Economic Tension: Severe debates surfaced around funding. Grid upgrades are hitting friction due to the politics of utility rate increases—specifically over whether everyday consumers or massive tech firms should shoulder the multi-billion-dollar costs.
-
- Resource Competition: Internet Service Providers (ISPs) at the forum expressed mounting concerns that the sheer scale of the AI data center buildout is worsening supply chain costs and causing significant permitting delays for standard broadband networks.
- Following the recent sale of its residential fiber business to AT&T, Lumen Technologies is facing permitting issues as it looks to expand its network to support billion-dollar deals with hyperscalers and enable a wide range of AI use cases. After exiting 2025 with about 17 million fiber miles, Lumen is projected to expand that to 58 million fiber miles when it exits 2031, explained Melissa Mann, Lumen’s chief public policy officer.
- “It’s not just an engineering question. It’s really a policy question and our ability to meet these demands,” Mann said, noting that it’s not clear whether Lumen will be able to obtain all the permits required to build as quickly as the hyperscalers want it to. “If we’re actually going to do this and double our fiber capacity across the industry, we’ve got to fix permitting,” Mann added.
- Giulia McHenry, SVP for public policy at AT&T, said the network operator has seen a 15% increase in overall data traffic since 2023, though not all is AI-related. “But we are ensuring that we’re ready for AI to cross our networks,” she said.
- Mann noted that up to 50% of Internet traffic on Lumen’s network is being driven by autonomous AI agents. Noting that delivering service at low latencies is becoming table stakes, not a special feature, she added, “Latency is no longer a preference. There’s a floor on latency for many of these [AI] use cases and applications.”
-
- Upstream Focus: Federal Trade Commission (FTC) Chairman Andrew Ferguson noted during his fireside chat that the most significant competitive and antitrust risks in artificial intelligence do not lie among competing AI models themselves, but rather upstream in the control of data, chips, and power infrastructure.
- Mann said 90% of Lumen’s customers now use more than one AI provider and more than one cloud provider. The ability to give them more control was a primary driver of Lumen’s recent acquisition of Alkira, a company that enables partners to orchestrate and move their data to different clouds and AI providers via a single pane of glass. In practice, that means that if an enterprise sees energy prices spike in Virginia, it can shift workloads to another region where energy costs are lower and where ample capacity is available, she explained.
4. Supply chain issues:
-
- Supply chain costs are “skyrocketing,” said AT&T’s McHenry, noting that a large data center might use as much fiber as the company lays down in a year.
- Supply chain constraints, particularly on memory, also impact broadband customer premises equipment (CPE), said Mark Walker, VP of technology policy at CableLabs. “As we are building those additional network miles and upgrading our networks, that increase in memory costs flows directly through to the capital costs and the ability to deliver services,” Walker said.
- “If broadband service providers are forced to pass along those hidden costs without measurably improving the service, customers will become frustrated,” added Harold Feld, SVP at Public Knowledge, a consumer advocacy group.
- Customer Premises Equipment (CPE) makers face mounting operational challenges due to global memory chip shortages. Also, the FCC ban on new foreign-produced WiFi routers, forces hardware developers to navigate complex recertification workflows to secure conditional regulatory approvals for redesigned router models. Manufacturers must re-apply for compliance clearances following any major component substitutions.
………………………………………………………………………………………………………………………………………………………………………
References:
Suggestions from IEEE Techblog Team and Qwoted “Experts” – How to Revitalize IEEE
In preparation for an IEEE Town Hall Meeting, 2pm-5pm Sept 26th at SCU library, I put out a request to the IEEE Techblog Team and to Qwoted “experts” to offer suggestions on how to revitalize IEEE. Event notice will be posted as soon as the participants are finalized.
Here are the suggestions from the Team members:
Alan’s Request to Qwoted “Experts”:
I’m looking at how IEEE can better serve its members through seminars, workshops, and short courses in newer technologies like AI, Cloud Native IT, cloud network topologies and architectures, and multi-cloud computing.
I’m looking to talk to people who can speak to what IEEE could be doing differently. whether that’s the format of training, the specific technologies being prioritized, or how these organizations approach continuing education more broadly for members whose careers are shifting toward software and cloud-based skill sets. Relevant suggestions from contributors will be consolidated into an article to be posted at the IEEE Techblog and IEEE Region 6 Newsletter.
Key takeaways will be discussed at an IEEE Town Hall meeting on September 26th 2pm-5pm at Santa Clara University organized by the IEEE Techblog Editorial Team and the IEEE Region 6 Director Joseph Wei. There will be two panel sessions and ample time for audience Q & A.
Recommended experts:
-Continuing education professionals at other tech nonprofits or professional associations
-Cloud architecture or multi-cloud computing specialists
-AI and cloud-native IT trainers or curriculum developers
-Engineers or IT professionals who have had to reskill from hardware-focused to software-focused roles
…………………………………………………………………………………………………………………………………………………………..
Responses from Qwoted Experts:
Srinivas Chippagiri, Salesforce:
Where I think IEEE could do things differently:
– Teach the transition, not just the tools. The genuinely hard part of reskilling wasn’t learning a specific technology like Kubernetes or Terraform. It was rewiring the underlying mental model: moving from “I own this box and its state” to designing for horizontal scale, statelessness, eventual consistency, and graceful failure. Most short courses teach the tool and skip the paradigm shift, which is exactly where hardware-background engineers get stuck. IEEE could differentiate by explicitly bridging that gap.
– Prioritize multi-cloud fluency over single-vendor certification. Real enterprise work now spans AWS and Azure and OCI, often at once. Training that locks members into one provider’s certification track leaves them half-equipped and vulnerable to vendor lock-in in their own careers. IEEE is vendor-neutral by nature — that’s a structural advantage over AWS/Azure/Google’s own training, and it should lean into teaching cloud concepts and cross-cloud architecture rather than one ecosystem.
– Go project-based, not lecture-based. Hardware and systems engineers learn by building and breaking things. A seminar or a slide deck doesn’t build cloud intuition; standing up a real multi-region deployment, watching it fail, and debugging it does. Hands-on labs against live cloud infrastructure will move members further than a lecture series on the same topic.
– Sequence the curriculum for career-shifters specifically. Someone coming from telecom or embedded doesn’t need the same on-ramp as a new grad. IEEE could design tracks that assume deep systems fundamentals but zero cloud exposure which is a very common and underserved profile among its long-tenured members.
– Prioritize the durable layer over the hype layer. AI is moving fast, but the skills that survive are cloud-native architecture, distributed systems reasoning, and cost/observability discipline. I’d weight the curriculum toward those foundations, with AI tooling taught on top of them, rather than chasing whatever’s trending that quarter.
…………………………………………………………………………………………………………..
Emily Hartstone, Hartstone LLC:
Filling a specific gap in the reskilling landscape:
There is now an entire training ecosystem for building with AI, prompt engineering, agent frameworks, RAG pipelines, and almost nothing teaching engineers how to govern what those systems are permitted to do once deployed. That gap matters for IEEE members specifically, because engineers reskilling from hardware into cloud and AI roles are the people who will be asked to sign off on giving autonomous agents access to production systems. This month’s OpenAI and Hugging Face incident, tens of thousands of unauthorized autonomous actions reconstructed only after the fact, is what that training gap looks like in production. Three concrete suggestions for the curriculum side:
First, a short course on runtime governance of autonomous systems: pre-execution authorization, scoped permissions, and fail-closed design, taught as engineering discipline rather than policy abstraction.
Second, incident-based workshops using real cases like the Hugging Face intrusion, the way safety engineering has always taught from failures.
Third, treating governance literacy as a core competency in every AI track rather than an elective, because the EU AI Act’s enforcement this August makes it a job requirement, not a specialization.
…………………………………………………………………………………………………………….
Udaya Bhaskar Vemuri, Corteva Agriscience:
IEEE can better support members by combining seminars with more practical, hands-on learning that people can immediately apply in their jobs.
Technology is evolving quickly, especially in AI, cloud computing, DevSecOps and software security. Professionals need short, focused learning paths that combine foundational concepts with real-world labs, case studies and demonstrations rather than relying mainly on theoretical courses.
I would also encourage IEEE to create learning tracks for different career stages. Early-career professionals have different learning needs than experienced engineers who want to expand into areas such as AI security or cloud-native architecture. Personalized learning paths and industry-recognized micro-credentials could help members build skills step by step.
IEEE could also strengthen its member community by encouraging peer learning through workshops, technical forums, and mentoring. Many of the best lessons come from engineers sharing practical experiences, challenges and solutions from real projects. That type of collaboration can help members keep pace with technology while also building a stronger professional network.
……………………………………………………………………………………………………………………
Iryna Kurkina, Academy Smart:
Your questions have several crucial points – which content to propose, how to build a relevant content – and I also think that another important here – which infrastructure can ensure engagement and what’s more important – ROI.
In terms of content and its types – it’s hard to overestimate the need for AI learning – as a tool that allows engineers to focus on architectural and business-related questions. In our own team we consider AI as a tool, not as engineer substitution. It helps to prototype dramatically faster – which allows to assess business impact of newly create solutions or features. Consequently, if we look at DevOps part – MLOps and AIOps is that part of cloud infra that every engineer has to be ready to deal with. In terms of content types – from our experience – engineers learn best with interactive tools – SCORM courses, simulations with real coding exercises, AI-powered recommendation engines that analyzes learning progress and behavior and recommends the next steps.
Video-only training is still good, but for engineering training – from our perspective – it is not sufficient. They have to have hand-on experience for better progress. And webinars are also very efficient – where engineers can not only share their experience, but also brainstorm, discuss.
In terms of infrastructure – besides classic LMS, the systems have to have those labs or spaces that can provide that coding / hand-on training experiences – and thus, the systems (LMS) must be able to track the progress to give the realistic analytics to the managers. Another challenge – consolidation of training information. Even if some team uses such labs – they quite often reside on a separate platform or environment which are not synced with the major learning progress. So either integrations must be done, or a new type of learning platform must be adopted.
I am happy to provide example of our projects and our internal approaches about how we manage continuing learning.
………………………………………………………………………………………………………………….
Arjun Sunke, Central New Mexico Community College (CNM):
On format, not just topics:
It does not appear to be the matter of which technologies are currently being addressed during training in the field of AI, cloud-native IT and multi-cloud learning. Most professional associations have kept up with recent technological developments in terms of technology selection. It seems to be the problem of the format of delivery of the courses in question. Many of them still rely on a relatively traditional approach to learning material development based on presentation of static slides with an occasional explanation. Such an approach might work reasonably well for stable and mature subjects but does not work for cloud computing and artificial intelligence learning. It is impossible to learn such a subject as multi-cloud architecture just by observing someone doing it.
The most successful approach in my own teaching is learning through scenarios and labs: placing learners in front of a live console in a real cloud environment where they have to work on solving the problem in the real world, not the one of PowerPoint slides about the best practices. It makes reskilling much easier to master for those members who are engineers or IT professionals and need to change their job profile to a more software-based one. It allows bridging the confidence gap quicker compared to lecture-style learning.
On integrating security into cloud/AI training, not treating it as a separate track:
Also worth noting is that cloud-native and multi-cloud training usually happens in isolation from security, as if “how do I architect this” and “how do I secure this” are separate curriculums for different audiences. They are not. All architectural decisions made in a multi-cloud scenario how the network segmentation is done, how identity and access management is handled across multiple clouds, how service to service authentication works is also an exercise in security. And teaching those separately results in people who are able to create something they cannot secure themselves, which turns into reality soon enough. Any change in ongoing education curriculum has to include security consideration as a part of cloud/AI training.
On prioritizing training for AI:
With respect to AI and skills adjacent to AI, I think training which recognizes that AI systems are infrastructure and requires governance is critical not training which teaches people how to use the tool. With organizations increasingly adopting AI and automation, issues such as “What does this system have access to, and how will we know if it behaves in an unexpected manner?” become equally important to “How can I use this technology?” Training which focuses exclusively on capability and not governance trains people to build systems faster than they can control them.
On format for delivery (seminars vs. workshops vs. short courses):
Considering the fast-paced nature of this community, it might be better to opt for short and regular workshops rather than seminar-style events. An event that lasts two or three hours on a specific topic (for example, “service-to-service authentication in a multi-cloud environment”) is much more likely to engage the members and provide them with practical skills than a long seminar on theoretical concepts. Short courses can serve as an intermediate step for those members who aim to develop their skills in order to gain a certain qualification in a few weeks’ time.
Why this matters for IEEE member retention:
Those who are transitioning to skills in software and clouds are probably making that transition because they have to, not because they want to, due to changes or disappearance of their existing careers. The implication here is that there needs to be immediacy in gaining confidence as part of the training process or these individuals may move elsewhere for training, such as boot camps, vendor certification training, or even YouTube. The important element of the IEEE training that sets it apart from all of these training options is credibility and sense of community, but this element is dependent on immediacy as well.
Would love to discuss this further or delve more into the lab-oriented teaching methods I’ve developed at CNM, should that be helpful in crafting your story.
………………………………………………………………………………………………………………………
Julie Scotland, Gravi AI:
A few things I’d tell IEEE:
1) Hands-on practice that applies their enterprise AI tools directly to their own day-to-day workflows.
2) Live cohorts with AI builds beat on-demand for faster skills and higher adoption. If you do go with on-demand, keep modules short and to the point.
3) Judgement and flexibility is as, if not more, important than teaching the tools themselves. That does not mean you don’t teach within the tools they use daily, but specific tool skills age quickly. Learners need foundational education that spans whatever tool they work with and learn how to adapt quickly as technology continues to evolve.
4) You will be constantly iterating, another reason why live cohorts work well because you enable near immediate industry and technical adaptation in real time.
Our association partner signups land well above usual course benchmarks. Happy to walk you through how we structure it. Free for 20 minutes?
……………………………………………………………………………………………………….
Rhys Higgs, The Discourse AI:
From my perspective in EdTech and AI-enabled workforce development, I think IEEE has an opportunity to rethink continuing education for engineers transitioning into software-defined, cloud-native, and AI-driven environments. The pace of change means professionals need learning experiences that are practical, flexible, and immediately applicable—not just technical presentations.
A few areas where I believe IEEE could differentiate itself:
– Shift from one-off seminars to structured learning journeys with hands-on labs, real-world projects, and peer collaboration.
– Focus on applied skills in AI, cloud-native architectures, Kubernetes, platform engineering, multi-cloud strategy, and AI governance rather than technology overviews alone.
– Build stronger connections between industry practitioners, academia, and employers so members understand not only emerging technologies but how they’re being adopted in production environments.
– Create communities of practice where learning continues beyond a workshop through mentorship, technical roundtables, and collaborative problem-solving.
One trend I’ve seen repeatedly is that experienced engineers aren’t struggling to learn new concepts—they’re struggling to connect those concepts to practical implementation and evolving job roles. The organizations that succeed in continuing education provide contextual, experience-based learning that helps professionals build confidence while staying current.
I’d be glad to share additional insights on designing AI and cloud training programs, learner engagement strategies, and how professional organizations can better support continuous upskilling in today’s technology landscape.
……………………………………………………………………………………………………………………
Navnit Kumar Shukla, Snowflake:
I can speak candidly to what IEEE’s continuing education is missing — the gap between IEEE’s traditionally hardware/standards-focused curriculum and what cloud architects and AI practitioners actually need today is significant and growing.
Key perspectives I can offer:
— Why most technical training fails practitioners mid-career (format problem, not content problem)
— What cloud-native and AI curricula need that IEEE doesn’t currently provide
— How the DeepLearning.AI model (hands-on labs + theory) compares to traditional certification approaches
— What engineers reskilling from hardware to cloud/AI need most
Happy to contribute for the IEEE Techblog article and Region 6 Newsletter. I’m also based in Southern California — available for the September 26th Town Hall at Santa Clara University if that’s useful.
………………………………………………………………………………………………………………………………
Ankit Pathak, ConsultAdd Inc:
Ankit can provide practical insights on how organizations and professional communities like IEEE can better prepare engineers and IT professionals for the next generation of technology careers, including:
* Why AI education should prioritize critical thinking, governance, evaluation, and responsible deployment—not just prompt engineering or tool demonstrations.
* The cloud-native and multi-cloud competencies today’s engineers need as AI workloads become increasingly distributed across enterprise environments.
* How training formats can evolve beyond traditional seminars into hands-on, scenario-based workshops that reflect real enterprise deployment challenges.
* The skills hardware-focused engineers should develop as their roles increasingly intersect with software, cloud infrastructure, and AI-powered systems.
* How professional organizations can create continuous learning pathways that keep members relevant as AI technologies evolve rapidly.
His perspective comes from advising enterprises on AI transformation and workforce readiness, where technical capability must be combined with governance, security, and practical implementation skills to deliver successful outcomes.
………………………………………………………………………………………………………………………………..
Kuber Sharma, Senior Director of Product Marketing at UiPath
What I keep watching in enterprise AI deployments is a specific failure mode that most continuing education programs don’t address: people learn how a technology works, and then get stuck the moment the deployment hits conditions the training never covered.
The gap is not in tool knowledge. It is in what I’d call operational judgment: knowing when to trust the system’s output, when to override it, when to escalate, and who is accountable when something goes wrong at the edge. You can teach cloud architecture in a short course. Teaching someone how to reason through a situation where the AI produced a plausible-looking answer with a bad input is a different problem, and it is much closer to what engineers and IT professionals actually face on day 90 of a deployment.
What IEEE could do that AWS and Azure training programs structurally cannot: teach the failure modes, not just the architecture. The vendor success stories are easy to find. A curriculum built around why enterprise AI and cloud projects fail, what the consistent patterns look like, and how to recognize them early is genuinely hard to find, and IEEE is neutral enough to teach it honestly.
The harder thing to design for is the context that training programs rarely simulate: what happens when the decision model meets a real organization where the data is incomplete, the ownership is disputed, the engineers aren’t the ones deciding what the system is allowed to do, and the legal team finds out about the deployment six months after it went live. That is not an unusual scenario. It is the median enterprise AI deployment. A curriculum that stops at the framework is preparing people for a world that does not exist. The session that would genuinely move people is the one that starts after the framework runs out.
…………………………………………………………………………………………………………………………………………………
References:
Sept 26, 2026 Town Hall Event Description to be forthcoming soon
IEEE SCV March 28th Event: A Conversation with IEEE President and IEEE Region 6 Director Elect
IEEE President’s Priorities and Strategic Direction for 2024
IEEE President Elect: IEEE Overview, 2024 Priorities and Strategic Plan
Palo Alto Networks: Frontier AI Critical Defense Program + Collaboration with NTT DATA for secure AI adoption
Palo Alto Networks Introduces Frontier AI Critical Defense Program:
Yesterday, cybersecurity leader Palo Alto Networks joined Nvidia and Anthropic in assembling a high-profile coalition focused on defending critical infrastructure against AI-enabled cyberattacks.
Gartner defines AI in cybersecurity as: “The application of AI technologies and techniques to enhance the security of computer systems, networks, and data to protect from potential threats and attacks. AI enables cybersecurity systems to analyze vast amounts of data, identify patterns, detect anomalies, and make intelligent decisions in real time to prevent, detect, and respond to cyberthreats.”
Using AI in cybersecurity solutions leads to faster and more accurate threat detection along with greater scalability and cost efficiencies. Palo Alto Network’s Frontier AI Critical Defense Program expands on its existing collaborations with IBM, Red Hat, Microsoft, Siemens, and Idaho National Laboratory. Anthropic, OpenAI, and Mitsubishi have now joined the initiative, which is focused on protecting operational technology (OT), health-care systems, commercial software, and open-source ecosystems from AI-driven exploits.
Participating organizations will work with Palo Alto Networks to identify and mitigate vulnerabilities at network scale. One element of the program is the deployment of “virtual patches”—network-level controls designed to neutralize known or newly discovered security weaknesses before software fixes can be developed, tested, and widely deployed.
Palo Alto Networks said its work with compute-intensive frontier AI models has already identified more than 14,000 previously unknown vulnerabilities in open-source software. By comparison, Anthropic reported that its Claude Mythos Preview Model had uncovered more than 23,000 flaws across more than 1,000 open-source projects.
IBM and Red Hat’s related Project Lightwell has not yet disclosed comparable findings. However, the initiative remains in its early stages, making direct comparisons premature.
These efforts reflect a broader shift in the cybersecurity threat landscape. AI systems can automate reconnaissance and exploit development while compressing attack timelines from weeks or days to minutes or seconds. Palo Alto Networks describes the objective of its Frontier AI Critical Defense Program as enabling critical infrastructure operators to “patch at ID speed”—that is, at the speed at which vulnerabilities can be identified—thereby narrowing the exposure window between discovery and remediation.
The emerging model represents a transition from predominantly human-paced cybersecurity operations toward a more compute-intensive and increasingly autonomous approach. AI agents can continuously search for vulnerabilities across complex software and network environments, potentially identifying weaknesses before they are discovered and exploited by adversaries using similar technologies.
“In the age of frontier AI, the traditional, reactive race to build and deploy software patches before adversaries exploit a flaw is a losing battle,” Palo Alto Networks Chief Product Officer Lee Klarich explained. “Protecting critical infrastructure requires a structural shift from isolated patching to collective, proactive intelligence. Through initiatives like our Frontier AI Critical Defense Program, we can neutralize threats at the network layer before they are weaponized.”
………………………………………………………………………………………………………………………………………………………………………………
NTT DATA and Palo Alto Networks Form Strategic Alliance to Accelerate Secure AI Adoption:
Today, NTT DATA, a global provider of AI, digital business, and technology services, and Palo Alto Networks have announced a multiyear strategic alliance aimed at helping organizations adopt AI securely, modernize cybersecurity operations, simplify complex technology environments, and strengthen cyber resilience for the AI era.
The agreement represents Palo Alto Networks’ first strategic alliance of this type with a global systems integrator. The companies expect the partnership to generate up to $1 billion in joint business by the end of the three-year period in 2029. The alliance combines Palo Alto Networks’ AI-powered cybersecurity platforms with NTT DATA’s consulting, systems engineering, and managed services capabilities.
Through joint engineering, co-innovation, and coordinated global delivery, the companies will help customers assess cyber risk, deploy AI securely, and continuously optimize their security environments. The resulting solutions are intended to provide an integrated path from cybersecurity strategy and implementation through ongoing managed operations.
Building on the companies’ existing collaboration through the Frontier AI initiative, the alliance will combine Palo Alto Networks’ Unit 42® threat intelligence with NTT DATA’s global cybersecurity expertise, AI-governance capabilities, and managed services. The effort will be supported by joint investments, more than 2,000 certified professionals, and dedicated Forward Deployed Engineers.
Direct engineering collaboration will also give NTT DATA early access to new Palo Alto Networks platform features, enabling the systems integrator to accelerate the development and delivery of AI-security services. The companies initially will focus on highly regulated and critical industries, including financial services, health care, manufacturing, and the public sector.
The initial portfolio will address six strategic transformation areas:
-
Autonomous security operations centers (SOCs): Modernize security operations through agentic AI and managed services that help organizations detect, investigate, and respond to increasingly sophisticated, machine-speed threats while reducing operational complexity.
-
AI governance: Integrate governance, security, and risk management across the AI lifecycle, enabling organizations to address emerging risks and scale AI initiatives with greater accountability, transparency, and control.
-
Identity security: Protect human, machine, and AI-agent identities—including workloads and devices—through an identity-security framework designed to discover, manage, secure, and govern identities across the enterprise.
-
Zero Trust and SASE: Secure users, applications, and data across an increasingly distributed attack surface through a unified Zero Trust and secure-access service edge architecture that incorporates AI-driven threat detection and prevention.
-
Resilient cloud: Improve visibility, compliance, and autonomous risk reduction across multicloud environments through AI-enabled security-posture management and stronger governance.
-
Firewall modernization: Modernize firewall infrastructures to reduce operational complexity, improve visibility, and strengthen enterprise-wide security.
“AI is reshaping both business and cybersecurity, making deep ecosystem collaboration more important than ever,” said Nikesh Arora, Chairman and Chief Executive Officer, Palo Alto Networks. “Expanding our alliance with NTT DATA allows us to operationalize platformization at true global scale, helping enterprises eliminate legacy complexity and move fast without sacrificing safety.” “AI is redefining every aspect of the enterprise, but it is also transforming the threat landscape at unprecedented speed. Organizations need a new approach to cyber resilience that combines AI-driven security, deep industry expertise and global scale,” said Abhijit Dubey, Chief Executive Officer and Chief AI Officer, NTT DATA, Inc.
“Together with Palo Alto Networks, we’re bringing AI-powered cybersecurity innovation together with NTT DATA’s consulting, engineering and managed services capabilities to help clients securely accelerate AI adoption and stay ahead of evolving threats.”
NTT DATA brings world-class cybersecurity expertise to the collaboration, backed by over 7,500 cybersecurity professionals, 70+ delivery centers and 20+ Autonomous Cyber Defense Centers. Paired with Palo Alto Networks AI-powered platforms and Unit 42 threat intelligence, the alliance delivers the technology, expertise and global reach enterprise organizations need to securely deploy AI across complex environments.
About NTT DATA:
Fortune Global 100. We are committed to accelerating client success and positively impacting society through responsible innovation. We are one of the world’s leading AI and digital infrastructure providers, with unmatched capabilities in enterprise-scale AI, cloud, security, connectivity, data centers and application services. Our consulting and industry solutions help organizations and society move confidently and sustainably into the digital future. As a Global Top Employer, we have experts in more than 70 countries. We also offer clients access to a robust ecosystem of innovation centers as well as established and start-up partners. NTT DATA is part of NTT Group, which invests over $3 billion each year in R&D. Visit us at nttdata.com
About Palo Alto Networks:
Palo Alto Networks (NASDAQ: PANW), the global AI cybersecurity leader, protects our digital way of life with a comprehensive portfolio of cybersecurity solutions and platforms across Network, Cloud, Security Operations, AI and Identity. Trusted by 70,000+ customers and powered by Unit 42 threat intelligence, our AI-driven platforms eliminate complexity, empowering enterprises to modernize with confidence and securing the speed of innovation. Explore the future of security at www.paloaltonetworks.com.
………………………………………………………………………………………………………………………………………………………………………………………………
References:
Anthropic’s Project Glasswing aims to reshape IT cybersecurity
Palo Alto Networks and Google Cloud expand partnership with advanced AI infrastructure and cloud security
Highlights and Analysis of July 30th U.S. Senate hearing on AI and telecommunications
Applying Zero Trust at the Wireless Edge: Securing Mixed WPA2 and WPA3 IoT Fleets
Fortinet and Palo Alto Networks are leaders in Gartner Magic Quadrant for Network Firewalls
Key Differences Between Network Cybersecurity and Control System Cybersecurity & Why It Matters
SHIELD-6G with AI-native cyber threat intelligence platform to enhance cybersecurity for Europe’s future 6G networks
Countdown to Q-day: How modern-day Quantum and AI collusion could lead to The Death of Encryption
Cybersecurity threats in telecoms require protection of network infrastructure and availability
Network X Americas: AT&T and Comcast reveal huge AI impact on network operations
Sovereign AI infrastructure for telecom companies: implementation and challenges
Dell’Oro: Data Center Physical Infrastructure revenue to grow at 22% CAGR from 2025-2025/forecast comparisons, analysis, risks
According to Dell’Oro Group, global Data Center Physical Infrastructure (DCPI) manufacturer revenue is projected to grow at a 22% compound annual growth rate (CAGR) from 2025 to 2030, reaching $120 billion by the end of the period. This growth is driven by net additions to installed IT capacity, which account for the large majority of the forecast. Additionally, the infrastructure content per megawatt will have a smaller effect as higher-density and liquid-cooled architectures redistribute spend across DCPI categories.
“The AI buildout has moved past the point where it can be treated as a surge. It is now the baseline against which the rest of the market is measured,” said Alex Cordovil, Research Director at Dell’Oro Group. “What has changed in this forecast is where the risk sits. Demand is no longer the open question—delivery is. Equipment lead times, construction labor, grid interconnection, and community consent all remain constrained, especially with the first statewide data center moratorium now in force.”
Additional highlights from the Data Center Physical Infrastructure 5-Year July 2026 forecast report:
- Capacity Additions Peak in 2026: Annual net capacity additions peak in year-over-year growth terms in 2026 and moderate steadily thereafter, remaining in double-digit growth territory through 2030. The market is still expanding quickly, but no longer accelerating. North America leads capacity additions over the period, followed by China.
- Thermal Management Leads Segment Growth: Thermal Management remains the fastest-growing DCPI segment, with liquid cooling the fastest-growing technology as rack densification moves the technology from an option to a precondition. Heat rejection coverage has been expanded in this edition, with water-cooled chillers expected to grow faster than air-cooled units on scalability rather than efficiency. Chillers remain a staple of data center specifications, even in warm-water designs, since free cooling loses effectiveness during the hottest days of the year.
- UPS Growth Concentrates in Larger Systems: Growth within the UPS segment concentrates in higher power rating three-phase systems, which are expected to expand faster than smaller units as the larger building blocks of AI clusters push deployments up the capacity curve. Medium-voltage designs are gaining ground, connecting UPS systems closer to the grid and attracting new entrants alongside established suppliers. Solid-state transformers are projected to weigh meaningfully on UPS demand beginning in 2029, initially focusing on large AI factories that have largely moved away from UPS-based architectures.
- Hyperscalers and Colocation Anchor Demand: Hyperscalers end the period as the largest single contributor to DCPI revenue, although their growth has slowed compared to the pace seen in 2025–26, as they lean more heavily on colocation partners to serve workloads, particularly outside the United States. Colocation remains central to the buildout, and the spread of powered shell development is shifting equipment procurement onto the tenant, moving revenue among customer segments without altering building occupancy. Newly separated in this forecast, AI-specialized Cloud—the neoclouds and AI model builders—becomes one of the fastest-growing lines in our coverage. Enterprise demand continues to grow, but more slowly than the rest of the market.
- Regional Diversification Builds: North America continues to lead regional growth, with China the next largest contributor. EMEA is the only region revised downward from the January forecast, reflecting slower power availability and a more difficult permitting environment. Community opposition has become a material constraint on siting, blocking or delaying a meaningful share of announced projects. Together with the expected repricing of U.S. natural gas, are expected to support faster growth in CALA and Asia Pacific excluding China.
Dell’Oro Group’s Data Center Physical Infrastructure 5-Year Forecast report provides a complete overview of the Data Center Physical Infrastructure market. This covers market sizes and forecasts for uninterruptible power supplies (UPS), thermal management, cabinet power distribution and busway, rack power distribution, IT racks and containment, and software and services. Allocation of manufacturer revenues by hyperscaler, other cloud, colocation, telco, and enterprise customer segments is also provided, alongside a forecast of data center capacity additions by region. For more information about the report, please contact us at [email protected].
………………………………………………………………………………………………………………………………………………………………………………………
Forecast Comparisons:
Dell’Oro’s $120 billion DCPI forecast through 2030 is at the high end of published physical-infrastructure manufacturer-revenue estimates, but it is directionally consistent with other firms’ forecasts for adjacent power, cooling, and mechanical/electrical (M&E) categories. The differences largely reflect market definition: DCPI is not interchangeable with total data-center capex, construction, IT equipment, or facilities real estate.
The 22% Dell’Oro CAGR should not be read as a consensus CAGR for every DCPI component. Power equipment forecasts around 7.5% and broader support-infrastructure forecasts around 8.2% suggest more moderate growth in legacy categories, while AI-linked liquid cooling is projected to grow in the mid- to high-teens.
The key forecasting judgment is therefore AI infrastructure content per MW: if GPU density keeps climbing and liquid cooling, high-voltage distribution, energy storage, and modular power systems become standard rather than niche, DCPI revenue can grow substantially faster than data-center floor space or even installed MW. Conversely, grid constraints, AI-demand normalization, and lower equipment dollars per watt from scale and engineering improvements could constrain manufacturer revenue growth even as deployed capacity continues to expand.
Comparable Forecasts:
A useful interpretation is that Dell’Oro’s $120 billion is plausible only if the market increasingly captures high-value AI-ready electrical and thermal systems—not merely traditional UPS, air-conditioning, and rack revenue. Adding standalone power and cooling forecasts cannot produce a clean “DCPI total,” because analysts differ in whether they include services, software/DCIM, integration, installation, generators, switchgear, rack infrastructure, and edge facilities.
………………………………………………………………………………………………………………………………………………………………………………………………………..
Image Generated by Perplexity.ai
…………………………………………………………………………………………………………………………………………………….
Analysis – main spending drivers:
-
AI accelerator density. GPU/accelerator clusters raise rack power from conventional enterprise levels to much higher levels, increasing demand for power distribution, UPS capacity, switchgear, busways, backup generation, and energy storage. ABI Research expects AI-dedicated active data-center capacity to rise from 11.5 GW in 2026 to 43.6 GW in 2031, and projects that AI will represent more than half of total data-center capacity in the early 2030s.
-
Shift from air cooling to liquid cooling. Higher-density AI systems cannot be served economically—or sometimes technically—by conventional room-level air cooling alone. Direct-to-chip cold plates, coolant-distribution units, rear-door heat exchangers, liquid loops, heat-rejection equipment, and controls raise cooling-system content per MW. Cooling equipment is therefore forecast to grow faster than the more mature broad power-equipment category.
-
Rapid capacity additions by hyperscalers and colocation operators. JLL expects roughly 97 GW of data-center capacity to be added globally from 2025 to 2030, approximately doubling the sector to about 200 GW. Every new MW requires a physical plant, even where the IT stack is sourced separately.
-
Power availability is becoming the binding constraint. Global data-center electricity consumption is expected to roughly double to 945 TWh by 2030 in the IEA base case. This puts a premium on grid interconnection equipment, substations, medium-voltage distribution, on-site generation, batteries, and energy-management systems—and can cause operators to overbuild or deploy infrastructure earlier than their server installations.
-
Resilience requirements and time-to-power. AI facilities require high availability alongside enormous load ramps. Operators are spending on redundant electrical paths, backup generation, battery systems, microgrids, and modular/skid-based electrical infrastructure to shorten construction schedules and reduce exposure to grid-connection delays.
-
Retrofitting the installed base. Demand is not solely greenfield. Existing hyperscale, colocation, and enterprise sites must upgrade electrical distribution and thermal plants to host AI pods, often retaining conventional infrastructure for legacy workloads while adding liquid-cooling islands.
-
Efficiency, water, and carbon constraints. Higher energy costs, grid constraints, water availability, and sustainability targets push investment toward more efficient thermal architectures, heat reuse where feasible, advanced controls, and power-management systems. These are often capital-intensive even when they lower lifetime PUE, water use, or operating cost.
……………………………………………………………………………………………………………………………………………………………….
Huge Risks to the forecast:
The central downside to Dell’Oro’s forecast is an AI-demand and funding reset: if OpenAI, Anthropic, or other frontier-model providers fail to turn extraordinary usage growth into durable, high-margin cash flows, capacity commitments could be deferred, resized, or cancelled. Because the DCPI forecast assumes nearly 200 GW of added data-center capacity through 2030, even a partial reduction in the AI build plan would materially affect power, cooling, and electrical-equipment orders.
The OpenAI/Anthropic risk:
The potential issue is not that either company vanishes overnight. It is that they—and the hyperscalers and GPU-cloud firms supporting them—may discover that the revenue and gross-margin trajectory does not justify the scale of previously contracted compute.
The risk chain is: AI monetization miss→lower compute utilization / pricing→capex deferrals→fewer energized MW→lower DCPI revenue.
The exposure is unusually concentrated. Advanced AI demand is dominated by a small number of hyperscalers and frontier-model providers; McKinsey estimates that 60–65% of AI workloads in the United States and Europe will be hosted on hyperscaler infrastructure by 2030. Thus, a retrenchment by a few large buyers can have an outsized effect on the physical-infrastructure supply chain.
Why OpenAI is a focal point:
OpenAI’s downside case would be a mismatch between compute obligations and customer monetization:
-
Consumer AI usage may remain high, but paid conversion, enterprise seat expansion, API volume, or willingness to pay for frontier-model performance may disappoint.
-
Inference costs may not decline fast enough relative to prices, leaving growth without attractive contribution margins.
-
New models may yield diminishing commercial differentiation, shortening product cycles and weakening customers’ willingness to pay premium prices.
-
Its financing requirements could become harder to meet if capital markets reassess terminal valuations, the cost of debt rises, or strategic partners limit exposure.
Some reporting and commentary point to very large continuing compute costs and funding needs relative to reported revenue, but the precise economics are opaque because OpenAI remains private and uses non-GAAP and run-rate measures inconsistently across reports. That opacity is itself a risk: DCPI vendors can see announced projects and committed capacity, but cannot fully observe the ultimate cash-flow support for the tenant’s demand.
Why Anthropic is not a complete hedge:
Anthropic’s enterprise orientation and reported revenue growth could diversify the sector’s demand base, but it does not eliminate systemic risk. It faces many of the same conditions:
-
Revenue is substantially concentrated in a relatively early enterprise-AI adoption cycle.
-
Enterprise customers can trial models broadly but consolidate suppliers quickly if performance differences narrow.
-
Model-price competition could reduce revenue per token or per API call faster than cost-per-token declines.
-
Large training runs are discretionary. A pause in the cadence of frontier-model releases would immediately reduce the urgency of new GPU clusters and associated electrical/cooling plant.
Recent reports describe unusually rapid revenue expansion and positive adjusted operating income for Anthropic, but the sustainability and definition of those measures are not independently transparent in the way public-company financial statements are. The relevant question for DCPI is not just whether Anthropic grows revenue, but whether its long-term contracted compute load and its own capital support remain sufficient to sustain multi-year facility commitments.
Other downside mechanisms:
An AI crash is the sharpest downside scenario, but Dell’Oro’s bullish outcome also depends on several more gradual assumptions.
Dell’Oro itself reportedly frames the immediate risk as delivery—equipment lead times, construction labor, grid interconnection, and community consent—rather than demand. Those bottlenecks can cut near-term revenue even if AI demand is real, because DCPI is recognized when facilities are physically delivered and commissioned, not when a GPU cluster or capacity plan is announced.
Efficiency is a double-edged sword:
Dell’Oro’s premise benefits from high rack density: AI systems require more substantial electrical architecture and move cooling from conventional air systems toward liquid cooling. McKinsey notes that direct-to-chip cooling can address roughly 60–120 kW racks, and that immersion can support still higher densities; those architectures increase DCPI content per rack and often per MW.
But efficiency can reverse the volume implication. Better accelerators, model distillation, mixture-of-experts approaches, lower-precision inference, and power-system improvements can reduce electricity and infrastructure required per unit of AI output. The IEA explicitly models a “High Efficiency” pathway in which technology and software efficiency gains materially restrain data-center electricity demand, while its “Headwinds” case assumes slower AI uptake and capacity growth that plateaus beyond 2030, with efficiency offsetting much of the effect of increased IT use.
The key analytical distinction is:
-
Revenue per MW can rise because AI racks need liquid cooling, high-capacity UPS, switchgear, busways, and sophisticated controls.
-
Total MW deployed can fall if model efficiency improves or commercial demand disappoints.
Dell’Oro’s $120 billion outcome requires both substantial net new MW and elevated DCPI content per MW. A positive outcome on only the second factor would not fully protect the forecast.
What would signal trouble:
For a forward-looking DCPI thesis, monitor leading indicators rather than announced headline capex:
-
OpenAI and Anthropic: paid enterprise adoption, API demand, realized—not merely annualized—revenue, gross margin, cash burn, and financing terms.
-
Hyperscalers: capex guidance, AI-service revenue, remaining performance obligations, capacity utilization, and disclosure of power or data-center commitments.
-
GPU-clouds and colocation firms: customer concentration, lease pre-commitments, cancellations, financing costs, and the ratio of contracted versus speculative capacity.
-
Physical deployment: utility interconnection queues, energized MW rather than planned MW, transformer/switchgear order cancellations, and data-center construction starts.
-
Economics: inference price declines versus cost declines, GPU utilization, and evidence that enterprise AI deployments generate measurable productivity or revenue returns.
A particularly bearish signal would be simultaneous model-price deflation, falling GPU utilization, and delayed data-center energization. That combination would mean the sector is not merely supply constrained; it would imply that the financial rationale for capacity has weakened.
Bottom line:
A failure by OpenAI or Anthropic to meet expectations could trigger a classic capital-cycle correction: capacity was ordered on expectations of demand, but the cash flows needed to validate the investment arrive later, at lower margins, or not at all. In that scenario, DCPI’s most vulnerable segments are discretionary greenfield power and cooling deployments attached to single large AI tenants or thinly capitalized GPU-cloud providers.
However, a single lab’s disappointment would not necessarily collapse the entire market. DCPI demand also comes from hyperscaler internal workloads, enterprise AI, cloud migration, conventional data growth, colocation expansion, and infrastructure upgrades. The most likely downside is therefore a lower and lumpier growth path, with project delays and inventory/order corrections, rather than zero growth. The more severe Dell’Oro downside requires a broad AI-ROI failure that causes multiple frontier labs and hyperscalers to retrench at the same time.
Dell’Oro: 2H2026 Data Center Capex to Accelerate due to massive AI Deployments
Nvidia CEO Huang: AI is the largest infrastructure buildout in human history; AI Data Center CAPEX will generate new revenue streams for operators
Huge Risks for the proposed $500B AI Investments from Giant Wall Street firms
Expose: AI is more than a bubble; it’s a data center debt bomb
Will Google Cloud’s AI and data analytics revenue +TPU IP licensing income offset huge AI CAPEX to produce a decent ROI?
Inside Amazon’s new data center network architecture: quasi random network topology and passive optical devices
Big Fiber’s $250M financing deal to buildout dark fiber routes for AI Data Center expansion
Analysis: Ethernet gains on InfiniBand in data center connectivity market; White Box/ODM vendors top choice for AI hyperscalers
Fiber Optic Boost: Corning and Meta in multiyear $6 billion deal to accelerate U.S data center buildout
How will fiber and equipment vendors meet the increased demand for fiber optics in 2026 due to AI data center buildouts?
Omdia’s 2025 Mobile Core Network Leaders: Huawei #1 in market share; Nokia #1 for portfolio competitiveness
Highlights of Omdia’s “Market Landscape: Core Vendors” Report:
1. Market Share:
Five vendors control the 4G and 5G core market with a combined market share of 86.1% in 2025, down from 87.4% in 2024. The top three players collectively captured 70.9% of total revenue. With such a concentration of market power, each vendor is increasingly focused on asserting leadership, particularly as competition intensifies around the next-generation 5G core. However, measuring market leadership in this space has its challenges
In 2025, Huawei was the market leader. It was followed by Ericsson, ZTE, Nokia, NEC, and Samsung Electronics.
• Ericsson, Huawei, and ZTE captured a combined 70.9% of global core revenue in 2025, up from 68.3% in 2024. When Nokia and NEC are added, these top five companies generated 86.1% of total core revenue in 2025, up from 87.4% in 2024. This is very high but still less than the 95.4% captured by these vendors in 2020, which indicates that upcoming vendors are collectively gaining market share.
• Huawei gained 5.0 percentage points in market share thanks to a more favorable geographical mix and market share gains in emerging markets. Ericsson, ZTE, and Nokia, however, lost share.
• Market share remains an important component of vendor strategy. Some vendors deliberately trade short-term margins for increased share, leveraging competitive pricing to secure new business, especially within the 5G core space, which presents fresh monetization opportunities. Conversely, others may prioritize margin protection over top-line growth, consciously sacrificing share to maintain profitability. A single vendor may use both approaches depending on the geography, project scope, or timing.
Samsung Electronics and four other vendors including the new participant, AxyomCore, are in the upcoming mobile core vendors group. However, caution is required when benchmarking these vendors, given the relatively fewer network functions (NFs) that they develop and their smaller market reach compared with the larger players.
Note: For 4G and 5G core revenue, Omdia did not include Communications Service Providers (CSPs)’ spending in NF virtualization infrastructure (NFVi) or server and management software.
…………………………………………………………………………………………………………………………………………………………………………………………………………………………………….
2. Portfolio Competitiveness:
Nokia was ranked No. 1 for mobile core portfolio competitiveness with the Finland headquartered company a leader across all seven competitiveness categories: core portfolio breadth, cloud-native maturity, signaling, automation, core as a service, AI/ML and analytics, and implementations of other network functions. The recognition reflects Nokia’s continued investment in mobile core technologies that help operators modernize their networks for the AI era. Recent deployments include the world’s first commercial mobile telco service based on 5G Core SaaS; Core SaaS Edge enabling local breakout for roaming subscribers; core network resilience solutions; telecom core modernization programs, and mission-critical network upgrades supporting the IoT, rail and utilities sectors.
“Core networks are becoming the intelligence layer of modern communications, connecting cloud-native operations, AI-driven automation and application innovation. This recognition from Omdia highlights the breadth and maturity of our portfolio across every major category operators are using to evaluate connectivity partners who can help them move toward more autonomous, resilient and programmable networks.” Kal De, SVP, Core Networks, Nokia.
“Nokia continues to distinguish itself as a technology leader in core networks, with advanced capabilities in categories spanning cloud native maturity, automation, AI/ML and analytics, and Core as a Service. These are no longer optional innovations but strategic requirements for telecommunication providers pursuing greater operational efficiency, service agility and monetization opportunities. Nokia’s comprehensive approach demonstrates a deep understanding of both current operator challenges and the future direction of the telecom market. Roberto Kompany, Principal Analyst, Mobile Infrastructure, Omdia.
3. Selected Mobile Core Vendor Strategies:
Huawei and ZTE dominate the global 5G mobile core landscape, combined representing over half of the global market revenue. Unlike Western vendors who heavily rely on US-based hyperscalers (AWS, Microsoft Azure, Google Cloud) to realize a cloud core, Huawei and ZTE use an entirely different playbook. Their strategy relies on vertically integrated, proprietary Telco Clouds combined with intensive AI-native automation, while focusing geographically on domestic, Middle Eastern, African, and Asia-Pacific markets due to geopolitical restrictions in Europe and North America.
Huawei and ZTE sell the entire mobile core vertical stack: the underlying hardware, the cloud virtualization layer, and the mobile core software.
- Huawei’s Strategy: Huawei actively positions its Huawei Cloud and proprietary platform (Telco Intelligent Converged Cloud – TICC) as the direct alternative to AWS or Azure for global operators. Instead of validating their 5G core on American hyperscalers, they build end-to-end “Cloud-Network Synergy” environments using their own Kunpeng and Ascend chipsets.
- ZTE’s Strategy: ZTE deploys its 5G Common Core Solution. It uses their in-house Distributed Cloud infrastructure, which fully abstracts 2G/3G/4G/5G pipelines into a singular, containerized platform. They target extreme reliability through proprietary data layers (like their stateless cloud database) rather than outsourcing data management to a public cloud.
In sharp contrast, non-Chinese mobile core vendors (Nokia, Ericsson, Samsung) treat the mobile core as pure application software meant to run on hyperscaler public clouds.
Nokia’s core portfolio supports deployment models spanning private, public and hybrid cloud environments and helps operators simplify core operations through automation, AI-driven analytics, resilient architectures and network exposure capabilities. The company’s approach enables telecom providers and mission-critical enterprises to accelerate service innovation while reducing operational complexity and improving network agility. Nokia is a primary driver of the “Core on Multi-Cloud” strategy, intentionally designing its cloud-native 5G Standalone (SA) core to be completely infrastructure-agnostic.
- AWS: Nokia collaborates with AWS to deploy automated, cloud-native packet core and IMS voice functions natively on AWS infrastructure.
- Google Cloud: Nokia utilizes Google Cloud infrastructure to run its 5G core while integrating Google’s advanced Generative AI and data analytics tools for autonomous network operations.
- Microsoft Azure: Nokia tightly integrates its core applications with Azure’s carrier-grade hybrid cloud platforms, optimizing hosting configurations for low-latency enterprise and edge applications.
-
- Google Cloud: Ericsson deeply collaborates with Google Cloud to evolve its cloud-native packet core stack, optimizing it to run on Google’s autonomous cloud infrastructure paired with Google’s Cloud TPUs and GPUs for network AI scaling.
- AWS: Ericsson partners with AWS to validate its 5G core functions on AWS Outposts, targeting hybrid cloud deployments for tier-1 operators.
- AWS: Samsung optimizes its cloud-native 5G Core on AWS to allow operators to quickly spin up network slices and private 5G instances.
- Microsoft Azure & Google Cloud: Samsung partners with both hyperscalers to deliver end-to-end virtualized network functions (VNFs) at the enterprise edge
…………………………………………………………………………………………………………………………………………………………………………………………………………………
References:
https://www.nokia.com/asset/215526/
Telco investments in mobile core networks surge 83% in 2025-Q4, but what about ROI?
5G in Europe: Broad coverage but 5G SA cloud native core network lags other major markets (Table)
Dell’Oro: Telecom carriers are on a 5G SA spending spree with more to come
GSA: 5G Non Terrestrial Networks, 5G SA and 5G Advanced gain momentum
Dell’Oro: Mobile Core Networks +15% in 2025; Ookla: Global Reality Check on 5G SA and 5G Advanced in 2026
Dell’Oro: RAN market stable, Mobile Core Network market +14% Y/Y with 72 5G SA core networks deployed
Applying Zero Trust at the Wireless Edge: Securing Mixed WPA2 and WPA3 IoT Fleets
By Iftikhar Javed khan with Ajay Lotan Thakur
Abstract
Zero-trust architecture is a security model that eliminates implicit trust based on network location and instead requires every access request to be continuously authenticated, authorized, and validated before it is granted (NIST SP 800-207), Yet while zero-trust architecture is intentionally network-agnostic, practitioners still need concrete design patterns to apply its principles at the wireless edge. This challenge is acute in IoT deployments that span multiple hardware generations, where newer devices support WPA3 while legacy endpoints remain limited to WPA2 and may be unable to support 802.1X supplicants, certificate-based authentication, or endpoint posture agents.
Based on an anonymized critical-infrastructure sensor deployment, this article presents a control pattern that segments devices according to their maximum supported security capability rather than forcing the entire fleet into a single compatibility-mode WLAN. WPA3-capable devices are placed in a WPA3-enforced domain, while WPA2-only devices are confined to a tightly restricted legacy domain.
The pattern combines five controls: capability-aware wireless segmentation, per-device Multi-Pre-Shared Key credentials, least-privilege policy enforcement, RF-exposure reduction, and access-point-integrated wireless intrusion detection. The central argument is that WPA3 is an important wireless security control, but it is not, by itself, a zero-trust architecture. Instead, the article shows how zero-trust principles can be translated into practical controls for networks that cannot immediately replace every legacy endpoint.
The Core Principle: Capability Dictates Posture, Not the Reverse
Most wireless security design begins with a chosen standard and then asks how to make the device fleet conform to it. In a homogeneous estate that works. In a mixed-generation IoT fleet it fails, because the fleet contains devices that physically cannot meet modern baseline sensors that support only WPA2-Personal, cannot run an 802.1X supplicant, and cannot host a posture agent. When a single standard is imposed on such a fleet, one of two things happens: either the network is dragged down to the capability of its weakest device, or the weakest devices are quietly excluded and left unmanaged. Neither is zero trust.
The inversion this article argues for is simple: let each device’s maximum supported security capability determine which policy domain it belongs to, and architect the network around that reality rather than against it. A WPA3-capable sensor and a WPA2-only sensor are not two configurations of the same policy; they are two different risk profiles that deserve two different domains. Once capability is treated as the independent variable, the rest of the design segmentation, credentialing, least-privilege enforcement, RF exposure, and monitoring follows from it.
Step One Is Always Visibility
Before any of this can be designed, the fleet has to be seen. In practice, the first problem in a mixed-capability wireless estate is not choosing controls, it is not knowing, with confidence, what is actually associated with the network and what each device can support. A design built on assumptions about the fleet is a design built on sand.
Visibility therefore comes first, and it has two parts as shown in figure 1. The first is a wireless inventory: enumerating the devices present on the medium, their association state, their supported security modes (WPA2-only versus WPA3-capable, SAE support, Protected Management Frames), and their physical distribution. The second is monitoring the medium itself for what should not be access points and clients that are not part of the sanctioned fleet. Only once the estate is known can devices be grouped by capability and confined to the right domain; and only once the medium is continuously observed can the segmentation be trusted to hold over time.
This reframes the usual order of operations. Segmentation and credentialing are what most WLAN-security discussions start with, but they are the second move. The first move is establishing and maintaining an accurate inventory because you cannot correctly assign a device to a capability domain that you have not yet discovered, and you cannot detect a rogue or misclassified device without ongoing observation.
Figure 1: Visibility
Why WPA3 Alone Is Not Zero Trust
WPA3 materially improves the wireless link. Under WPA2-Personal the pairwise master key is derived directly from the passphrase, so an attacker who captures the 4-way handshake can mount an offline dictionary attack against it. WPA3-Personal replaces that PSK authentication with Simultaneous Authentication of Equals (SAE), a password-authenticated key agreement that establishes the pairwise master key without exposing a crackable value the 4-way handshake still runs afterwards to derive and install session keys, but it no longer leaks an offline-attackable target. Protected Management Frames additionally harden management traffic against deauthentication and disassociation abuse. These are real gains, and WPA3-capable devices should use them.
But link protection is not access control. Whether a sensor is on WPA2 or WPA3 says nothing about which application it may reach, whether it may talk to its neighbours, whether it can reach a management interface, or how anomalous behaviour is detected after it has associated. A device can hold a perfectly valid WPA3 credential and still be a compromised endpoint. Zero trust asks a different set of questions than “is the link encrypted”: what is this device, what is it allowed to do, and is it still behaving as expected. Answering those requires credential context, an explicit policy decision, an enforcement point, and continuous monitoring none of which WPA3 provides on its own.
Figure 2: Controls
Control 1 : Capability-Aware Wireless Segmentation
Devices are grouped by their maximum supported wireless security capability and assigned to distinct SSIDs and security zones accordingly: WPA3-capable sensors to a WPA3-enforced domain, WPA2-only devices to a separate, narrowly scoped legacy domain, each with its own firewall and access-control policy.
It is worth being precise about the alternative, because the single-SSID case is often overstated. A WPA2/WPA3 transition-mode SSID is technically possible and would preserve legacy compatibility. The problem is not that one SSID mechanically forces one shared WPA2 key; it is that a transition-mode SSID cannot enforce WPA3-only operation for capable devices, and it places endpoints with materially different risk profiles inside the same wireless policy domain. Separating by capability is a deliberate choice for policy separation and to prevent silent downgrade of capable devices, not a workaround for a technical impossibility. The SSID is not the security boundary; it is the first sorting step, with real enforcement occurring downstream at the firewall or microsegmentation gateway.
Control 2: Per-Device Credential Granularity
A single fleet-wide pre-shared key is replaced with a per-device or narrowly-scoped-group credential scheme, so that a compromised credential exposes one device or a small group rather than the entire fleet, and any one device can be revoked without re-keying everything.
The security benefit is credential granularity and accountability, not cryptographic identity. A pre-shared key remains a possession-based secret: holding it proves possession, not verified device identity, and device identity should never rest on a spoofable MAC address alone. The honest framing is that per-device keying shrinks blast radius and enables granular revocation and gradual rotation, a meaningful improvement over a shared key, but a step below certificate-based authentication, which constrained legacy sensors often cannot support. That gap is precisely why the surrounding controls (segmentation, least-privilege enforcement, monitoring) carry the rest of the load.
Control 3: Least-Privilege Policy Enforcement
This is the control that turns WLAN hardening into a zero-trust pattern, and it is the one most often missing. Each sensor is permitted to reach only the services its function requires, enforced downstream of the wireless layer at a firewall or microsegmentation gateway:
- Permitted: its designated telemetry collector, an authorized DNS resolver, an approved NTP source, and a required update service where applicable.
- Denied: general internet access, sensor-to-sensor communication, any access to wireless or infrastructure management interfaces, and any reach into user, server, or administrative networks.
- Logged: denied flows and policy violations, forwarded to central monitoring for correlation.
Stated plainly: a sensor should be able to send its readings to exactly one collector, resolve names, keep time, and nothing else. Segmentation decides which domain a device lives in; this layer decides what it is allowed to do once there. Without it, capability-aware SSIDs are just better-organized flat networks.
Control 4 : RF-Exposure Reduction
Because the exposure boundary of a wireless network is defined by radio propagation rather than by cabling, access-point placement, transmit power, and minimum data rates are optimized to reduce unnecessary signal propagation beyond the intended service area, validated through an RF survey rather than controller settings alone.
This is defense-in-depth, not a perimeter. RF exposure reduction lowers casual reachability, but a sensitive receiver or a directional antenna can still detect a usable signal outside the intended area, so it must never be presented as containment. Any power or rate change must also preserve required coverage, client uplink performance, roaming behaviour, redundancy, and application reliability; an over-aggressive reduction that breaks associations trades a small exposure gain for an availability loss.
Control 5 : Wireless Intrusion Detection and Continuous Monitoring
Continuous observation of the medium closes the loop opened by the visibility-first principle. Access-point-integrated WIDS/WIPS, or dedicated monitor-mode sensors, observe the over-the-air environment for conditions that IP-layer inspection cannot see, because rogue-AP and evil-twin detection depends on access to raw 802.11 management and control frames rather than post-association traffic.
A practical detection methodology for this environment prioritizes, in order:
- Rogue-AP identification: Unsanctioned access points advertising reachable SSIDs, the highest-priority wireless threat because it can bypass the entire wired policy stack.
- Evil-twin / SSID impersonation: Access points spoofing a legitimate SSID to lure client associations.
- Unauthorized or misclassified clients: Devices associating outside their expected capability domain, or appearing where the inventory says they should not be.
- Association and authentication anomalies: Repeated authentication failures, unexpected device appearance or movement, and deauthentication patterns consistent with over-the-air attack.
Wireless events are forwarded to a central monitoring or SIEM platform and correlated with firewall, authentication, and telemetry logs, so an over-the-air anomaly and a policy violation on the same device can be seen together. Where a network IDS is also used, its role is post-association IP-traffic inspection a separate function from over-the-air 802.11 monitoring, and the two should not be conflated
Reference Architecture
The controls compose into a single flow: capability determines domain, credential and policy determine access, and the medium is continuously observed.
Figure 3: Reference Architecture
Mapping Wireless Properties to Zero-Trust Principles
| Wireless Property | Control Applied | Zero-Trust Principle |
| Heterogeneous device capability | Capability-aware segmentation | Group by risk; do not downgrade capable devices. |
| Shared credential / broad blast radius | Per-device credential granularity | Bounded compromise; granular revocation. |
| Implicit post-association trust | Least-privilege policy enforcement | Explicit per-device authorization. |
| Propagation-defined exposure | RF-exposure reduction | Reduce reachable attack surface (defense-in-depth). |
| Unobserved medium | Visibility + WIDS/WIPS | Assume breach; continuous over-the-air monitoring. |
Applicability to Other IoT Environments
Although derived from a remote environmental-monitoring fleet, the pattern generalizes to any heterogeneous wireless estate that cannot hold every device to one standard healthcare and medical IoT combining modern and legacy or safety devices, building-management and facilities systems, industrial and operational-technology sensors, and smart-city or asset-tracking deployments acquired across long procurement cycles. In each, the same five questions apply, and in the same order: is the estate actually visible; can endpoints be separated by capability and risk; can credentials be scoped per device or small group; can access be restricted to explicitly authorized services; and is the wireless medium continuously monitored.
Limitations and Residual Risks
- WPA2-only legacy devices remain a residual risk until lifecycle replacement; the pattern bounds that risk, it does not eliminate it.
- Per-device pre-shared keys improve granularity but do not provide the assurance of certificate-based authentication.
- RF-exposure reduction lowers casual reachability but cannot prevent reception by sensitive or directional equipment.
- WIDS/WIPS provides detection, not guaranteed prevention; false positives require tuning and operational handling.
- Posture assessment is often limited on constrained endpoints, which is why segmentation and least-privilege enforcement carry more of the load.
- Long-term remediation should include lifecycle planning to retire devices that cannot meet the required security baseline.
Implications for Standards and Practice
As IoT-specific zero-trust guidance matures, the wireless access layer where many IoT deployments are, in practice, most exposed deserves treatment as a primary zero-trust surface rather than a downstream detail. The pattern here is deliberately buildable with widely available capabilities (WPA3/SAE, per-device keying, firewall or microsegmentation policy, AP-integrated WIDS/WIPS), which matters for constrained, hard-to-patch fleets that cannot absorb heavyweight agents. Recent research on trust-boundary management in heterogeneous, multi-radio IoT environments makes a compatible argument that current zero-trust frameworks assume relatively stable networks and under-address dynamic wireless conditions and points to the same conclusion: shared-medium risk, capability-aware segmentation, and wireless monitoring belong in IoT zero-trust frameworks as first-class concerns.
Conclusions
Mixed-capability wireless IoT fleets should not be collapsed into one lowest-common-denominator policy. The more robust approach inverts the usual order: let each device’s capability dictate its security domain, establish and maintain visibility of the medium first, and then apply capability-aware segmentation, per-device credentialing, least-privilege enforcement, RF-exposure reduction, and continuous wireless monitoring on top of that foundation. WPA3 is an important control within this pattern, but it is not the pattern itself. The result is a reusable, achievable design for organizations that must secure legacy and modern wireless IoT devices at the same time without waiting for a fleet-wide hardware refresh that may never come.
…………………………………………………………………………………………………………………………………………………………………..
References
- Cloud Security Alliance, Zero Trust Guidance for IoT https://cloudsecurityalliance.org/artifacts/zero-trust-guidance-for-iot
- IEEE Std 802.11-2024 (WLAN MAC and PHY base standard) https://standards.ieee.org/ieee/802.11/10548/
- Wi-Fi Alliance WPA3 certification documentation (SAE, Protected Management Frames)
- J. Shelby, “Zero Trust for Multi-RAT IoT: Trust Boundary Management in Heterogeneous Wireless Network Environments,” arXiv:2602.08989
- Rose, S., Borchert, O., Mitchell, S., and Connelly, S., Zero Trust Architecture, NIST Special Publication 800-207, National Institute of Standards and Technology, August 2020.
- https://www.cwnp.com/uploads/1660_cwnp-wp_revisiting-wpa-2-understanding-its-security-gaps-and-evaluating-the-move-to-wpa3_ver120224.pdf
About the author:
Iftikhar Javed khan is an enterprise security architect specializing in wireless security and zero-trust architecture for IoT and critical-infrastructure environments. He holds the Cisco CCIE and CWNP CWNE credentials and is a Senior Member of the IEEE. https://www.linkedin.com/in/iftikhar-j-03aa5533/
Ericsson and MediaTek Demonstrate 3GPP-Based GNSS RTK Positioning with Sub-30cm Accuracy Over a Commercial 5G Network
Ericsson and MediaTek have completed a world-first end-to-end demonstration of high-precision outdoor positioning over a commercial 5G network, using standardized 3GPP Global Navigation Satellite Systems (GNSS) Real-Time Kinematic (RTK) assistance. The trial achieved stable outdoor positioning accuracy below 30 cm using a handset’s integrated antenna, demonstrating a scalable path to decimeter-level positioning without proprietary positioning infrastructure or dedicated external GNSS hardware.
The demonstration validates how 5G networks can distribute GNSS RTK correction information efficiently to devices for industrial automation, autonomous mobility, drones, robotics, and other applications requiring reliable, mission-critical location awareness.
Image courtesy of Ericsson
………………………………………………………………………………………………………………………………………………………………………………………………………………………
Technical highlights:
-
Sub-30 cm positioning performance: The device achieved stable outdoor accuracy below 30 cm with its built-in antenna. With a geodetic-grade GNSS antenna, the same RTK approach can support centimeter-class positioning performance.
-
3GPP-standardized architecture: The solution uses capabilities specified in 3GPP Release 16 for GNSS positioning assistance, enabling an interoperable architecture across compatible networks, positioning servers, chipsets, and devices.
-
End-to-end unicast and broadcast validation: Both delivery modes were demonstrated end to end:
-
Unicast GNSS RTK: LPP assistance messages are delivered through Secure User Plane Location (SUPL) over standard IP connectivity, enabling targeted delivery to individual UEs.
-
Broadcast GNSS RTK: RTK assistance data is transmitted through Positioning System Information Blocks (PosSIBs), allowing common assistance information to be delivered simultaneously to all capable devices within a cell.
-
-
Efficient scaling model: Broadcast delivery can reduce repeated transmission of identical RTK correction data in high-density areas, while unicast can be used for lower-traffic scenarios, individualized service handling, or LTE-connected devices. This supports flexible selection of the most efficient assistance-delivery method based on coverage, device population, and radio-resource conditions.
-
Secured broadcast operation: While broadcast information is radio-accessible to devices in the coverage area, 3GPP mechanisms support encrypted positioning assistance. Authorized subscribers obtain the required decryption material through the 5G Core, including the Access and Mobility Management Function (AMF), protecting access to premium high-precision positioning services.
GNSS correction support:
The trial supports two standardized GNSS correction-data models:
-
Observation State Representation (OSR): Provides corrections derived for a specific receiver or location context, including GNSS observation-related corrections.
-
State Space Representation (SSR): Delivers correction parameters associated with GNSS satellite orbit, clock, bias, and atmospheric error states. SSR is particularly suitable for scalable service delivery because the common state information can be applied by many devices across a service area.
In both cases, the UE applies the received RTK assistance in real time to mitigate satellite, orbital, clock, ionospheric, tropospheric, and other GNSS error sources that limit conventional standalone GNSS accuracy.
Network and device implementation:
The test used Ericsson’s 5G Core, Ericsson Network Location (ENL), Ericsson 5G Advanced Location Services, and Ericsson RAN to generate, manage, and distribute OSR and SSR correction information through the mobile network. MediaTek validated the device side using its latest 5G modem technology with integrated GNSS capability, including real-time processing of advanced GNSS correction data with power-optimized implementation.
“This is a milestone for the 5G ecosystem,” said Johan Hultell, Head of Product RAN Software at Ericsson. “Together with MediaTek, we make high-precision location more accessible to device makers and enterprises, accelerating use in manufacturing, transport and beyond.”
Dr. HC Hwang, General Manager of Wireless Communication Systems and Partnerships at MediaTek, added: “The collaboration with Ericsson demonstrates how 5G Advanced technology can unlock new value for consumers and enterprises alike. Our modems with integrated GNSS are ready to support these advanced positioning services, paving the way for the next generation of intelligent devices.”
Technology backgrounder:
Conventional standalone GNSS typically provides positioning accuracy in the range of several meters. RTK improves this performance by using measurements from accurately surveyed reference stations to generate correction assistance. In the 3GPP architecture, this assistance is delivered through the cellular network using LPP-based positioning procedures and, where applicable, broadcast system information.
By standardizing the delivery of GNSS RTK assistance over 4G and 5G networks, operators can offer high-precision location capabilities at network scale. This creates a standards-based foundation for digitalized industrial operations, autonomous systems, intelligent transportation, asset tracking, and spatially aware consumer devices.
Accuracy to within tens of centimeters is a significant improvement compared to traditional GNSS, which is accurate to within several meters, and therefore paves the way for more advanced use cases. These include the safe operation of autonomous vehicles, drones and industrial robots in complex outdoor environments, Ericsson said. Complex environments encompasses places like factories, logistics hubs, ports, and mines etc.
…………………………………………………………………………………………………………………………………………………………………………………………………………………………………………
References:
https://www.ericsson.com/en/news/2026/8/ericsson-and-mediatek-raise-the-bar-for-location-accuracy
https://www.telecoms.com/5g-6g/ericsson-and-mediatek-put-precise-5g-positioning-through-its-paces
MediaTek overtakes Qualcomm in 5G smartphone chip market
MediaTek will use TSMC to make its Dimensity SoC’s in 2024
AT&T’s 600 MHz Deployment with Ericsson: Turning Low-Band Spectrum Into Coverage and Uplink Capacity
AT&T/Ericsson Demonstrate 5G-Based ISAC for Drone Detection at World Cup Stadium
Ericsson leads SK Telecom AI RAN vs. NVIDIA’s GPU centric AI RAN Alliance
AI-RAN and Agentic AI get real: Ericsson, Nokia, Verizon & other operators enter into a new network automation era
Analysis: Ericsson’s leading role in French INTENTION 6G project
5G in Europe: Broad coverage but 5G SA cloud native core network lags other major markets (Table)
Europe’s 5G market status appears increasingly uneven. The region has achieved broad 5G population coverage, but its transition to the 5G Standalone (SA) core network—and the cloud-native 5G Core required to realize 5G features and capabilities—remains materially behind several major markets.
A recent workshop report from the EU-backed European 5G Observatory highlights stakeholder concern that European investment remains weighted toward radio-access coverage rather than toward core-network modernization. The underlying Observatory assessment also characterizes EU deployment as still predominantly NSA, meaning much of the installed 5G radio layer continues to rely on an LTE/EPC anchor rather than a service-based 5GC architecture. Participants estimated that only 18% of European 5G investment is directed to the core network, compared with 40% in China and 34% in the United States, South Korea, and Japan. Although European 5G core investment reportedly rose 31% from 2024, the prevailing view was that the next phase of 5G deployment must prioritize SA-capable core infrastructure rather than further expansion of basic 5G coverage.
This distinction is technically significant. Much of Europe’s existing 5G footprint remains based on Non-Standalone (NSA) deployment, in which 5G NR radio access is anchored to the LTE/Evolved Packet Core domain. The 3GPP-defined SA architecture instead pairs 5G NR with the 5G Core (5GC), including its service-based architecture specified principally in 3GPP TS 23.501 and TS 23.502. A full SA implementation enables functions that NSA cannot support as comprehensively, including end-to-end network slicing, native 5G QoS flows, ultra-low-latency service support, exposure of network capabilities through APIs, and more flexible edge and enterprise-service integration.
Europe’s aggregate 5G investment level is reported at 50.6%, substantially above the proportion allocated to the core but below China’s 72.0%, the United States’ 62.0%, and Japan’s 58.0%. The imbalance helps explain why headline coverage statistics do not yet translate into comparably mature 5G SA availability or widespread advanced-service monetization.
The Observatory report indicates that basic 5G coverage now reaches nearly all EU citizens, corresponding to 96.8% of households overall and 88.9% of rural households. However, deployment of SA-capable sites remains limited: on average, only 21.6% of European base stations are reported to operate in 5G SA mode, compared with 36% in the United States and 35% in China.
The gap becomes more pronounced when considering commercial service availability. While 21.6% of European base stations may be SA-capable, commercial SA availability to end users is reported at only 2.8%. India provides a notable contrast: the report places its SA-capable base-station share at 10%, below Europe’s, but commercial SA availability at 50%. The comparison suggests that Europe’s challenge is not only deployment of SA-capable RAN and 5GC infrastructure, but also the operationalization, device enablement, service launch, and commercial scaling of SA offerings.
Stakeholders attributed the low European availability figures partly to the current concentration of SA deployment in enterprise and business-to-business applications. Germany reportedly has the EU’s highest proportion of SA-capable base stations, at 63.2%, yet commercial availability stands at only 2.5%. Austria has the highest reported EU SA availability, although that figure remains modest at 8.7%.
This enterprise emphasis is understandable. A 5G SA network provides the architectural basis for differentiated connectivity services specified across the 3GPP 5G system framework, including network slicing and dedicated QoS treatment. In particular, 3GPP TS 23.501 defines network-slice concepts and service requirements, while 3GPP TS 23.503 specifies policy and charging-control mechanisms that can support service differentiation. Private 5G networks, campus deployments, industrial automation, logistics, utilities, and critical-infrastructure applications are therefore among the most immediate candidates for SA-led value creation.
Workshop participants nevertheless agreed that broader 5G SA deployment, together with private-network expansion, will be necessary if Europe is to capture the full economic value of 5G. That requires investment not only in 5GC functions, but also in cloud infrastructure, transport capacity and synchronization, orchestration, security, operational automation, and interoperable exposure frameworks. In practical terms, the opportunity is to shift from a coverage-centric 5G model to one capable of delivering programmable, assured, and differentiated connectivity services.
The European Commission established the 5G Observatory in 2018 as an evidence-gathering and policy-support mechanism. The Commission reports that the EU had reached 75.3% harmonized spectrum assignment by 2025, with several Member States approaching or achieving full assignment. Spectrum progress is important, but it does not by itself ensure SA maturity: operators must still convert spectrum assets and broad NR coverage into commercially available 5GC-based services.
The investment challenge is substantial. A GSMA assessment published earlier this year estimates that Europe will require approximately $550 billion in mobile-network investment over the next decade, while operators may have access to only about $312 billion. That implies an investment gap of roughly $238 billion—one that could constrain Europe’s ability to close the SA, cloudification, and advanced-network-services gap with leading global markets.
Mr. Johannes Theiss, DG CNECT Head of Sector for Advanced Networking Technologies and Applications at the European Commission noted that 5G standalone (SA) will be crucial on the road towards 6G, though further work is needed to develop indicators that will be both meaningful and manageable. As preparations begin for how 6G progress will eventually be measured, the experience of tracking 5G provides a useful basis for identifying what worked well and what should be approached differently.
…………………………………………………………………………………………………………………………………………………………………………
5G SA -Europe vs RoW Comparison:
A 5G SA system requires both 5G NR radio access and a 3GPP 5G Core (5GC), rather than NR anchored to an LTE EPC as in NSA. The relevant architecture is defined principally in 3GPP TS 23.501 and associated procedures in TS 23.502. Accordingly, the table is best interpreted as an indication of how extensively each market has extended SA capability into the RAN—not as a direct count of deployed cores.
The EU’s 5G SA deployment footprint represented 20.9% of all mobile base stations in 2025, compared with 36.2% in the United States, 34.8% in China, 26.3% in Japan, and 26.2% in South Korea
5G SA deployment proxy – 5G SA base stations as a share of all mobile base stations (Source: Perplexity.ai):
…………………………………………………………………………………………………………………………………………………………………………
References:
Ookla: Uneven 5G deployment in Europe, 5G SA remains sluggish; Ofcom: 28% of UK connections on 5G with only 2% 5G SA
Ookla: Europe severely lagging in 5G SA deployments and performance
Dell’Oro: Telecom carriers are on a 5G SA spending spree with more to come
GSA: 5G Non Terrestrial Networks, 5G SA and 5G Advanced gain momentum
Dell’Oro: Mobile Core Networks +15% in 2025; Ookla: Global Reality Check on 5G SA and 5G Advanced in 2026
Dell’Oro: RAN market stable, Mobile Core Network market +14% Y/Y with 72 5G SA core networks deployed
AT&T deploys nationwide 5G SA while Verizon lags and T-Mobile leads
Impact of optical component shortages & bottlenecks explained + Hyperscaler’s CAPEX
LightCounting’s August 2026 market and component report highlights how severe physical supply shortages for high-speed optical components are reshaping data-center architectures and forcing cloud builders to expand their vendor pools from the traditional 2-3 suppliers up to 5-7 specialized hardware manufacturers. Qualifying and managing all of them is a new challenge.
- Applied Optoelectronics, Macom, and MaxLinear have already reported accelerating growth in Q2 2026 and we expect to see more examples in the upcoming earnings reports.
- Coherent and Lumentum have also reported improvements in growth rates for the last quarter, catching up with Eoptolink and Innolight (reporting at the end of August). Eoptolink has already disclosed a sharp increase in the profits expected for Q2. Accelink and CIG also report sharply higher profits.
- Tower Semiconductor reported triple-digit growth (y-o-y) in silicon photonics revenue in Q2-2026.
- GlobalFoundries plans to double its silicon photonics business in 2026.
- Cisco reported “a remarkable 28% y-o-y increase” in the networking segment revenues, up from 25% in Q1. Cisco reported three new hyperscale AI design wins and 40% growth in orders during the quarter. The company also reported record orders for campus networking – up 20% y-o-y.
- Arista Networks reported Q2 2026 revenue of $3.04 billion, marking its first-ever $3 billion quarter. This represented a roughly 38% year-over-year increase. Management also raised its full-year 2026 revenue growth outlook to 40%.
- Calix reported 21% y-o-y growth in Q2 and guided for another 15% in the current quarter. Growth in company’s revenue is attributed to sales of broadband equipment and AI-enhanced software to telecom operators.
- Extreme Networks reported 14% y-o-y growth in product sales, in part driven by success of its agentic AI networking platform.
Surging data-center traffic driven by AI queries has created acute optical component shortages, fundamentally altering supply chain qualifications for hyper-dense network switches and interconnects. The primary optical component bottleneck is a severe production shortage of Indium Phosphide (InP) laser chips and EML (Electro-Absorption Modulated Laser) components, which are failing to keep pace with a projected 53% surge in total optical transceiver and hardware demand, reaching $39 billion. Current demand for high-speed datacom optics is outstripping available supply by roughly 30%, forcing hyperscalers and system vendors to fundamentally restructure their supply chains. More details in the Addendum at the end of this article.
………………………………………………………………………………………………………………………………………………………………………………………………
Dell’Oro Group says that the rise of agentic AI and inference workloads is driving new demand and introducing network requirements that differ from those associated with training workloads. This shift is leading to significant expansion of front-end networks to support the broader buildout of AI infrastructure. We estimate that more than half of the growth in Front-end Data Center Switch sales over the next few years will be driven by AI-related opportunities. These deployments represent net-new expansion opportunities for both established vendors and new market entrants.
Accton, Arista, Celestica, Cisco, HPE/Juniper, H3C, Huawei, NVIDIA and others—emerge as the primary beneficiaries of this growth, according to the market research company.
“As AI infrastructure shifts from large-scale training to inference and agentic workflows, there is an increasing demand for general-purpose infrastructure, and expanded front-end network requirements,” said Sameh Boujelbene, Vice President at Dell’Oro Group. “The traditional assumption of a 10-to-1 ratio of XPU to CPU no longer applies across all deployments, with some environments moving closer to a 1-to-1 ratio. CPUs are becoming increasingly important for workloads orchestration and data movement. Additionally, networking for KV caching storage rack is also needed for inferencing applications,” added Boujelbene.
……………………………………………………………………………………………………………………………………………………………………………………….
As for the network providers/ISPs:
- Verizon announced a dark-fiber deal with Google to support the hyperscaler’s data center traffic growth. The deal is valued at over $1 billion and Verizon said there will be other deals announced before year-end.
- Deutsche Telekom said its AI factory for Germany’s public institutions and businesses that opened in February, has already sold the site’s 10,000 Blackwell GPUs capacity. DT is now looking at increasing capacity by another 20,000 GPUs.
- Comcast cites that its upstream broadband traffic is growing at 2.5x the rate of downstream data, which it links to AI-driven queries.
- AT&T is seeing rapid increases in large-scale data traffic requiring high-capacity metro and intercity fiber infrastructure.
Capex of Telecom Service Providers is still expected to be flat (or down 1%) in 2026, but AT&T and Comcast reported 16% and 20% y-o-y growth in capex for Q2, respectively. Both companies are investing more in broadband access. Verizon’s capex was also up, but only by 5% in Q2.
……………………………………………………………………………………………………………………………………………………………………………………….
Hyperscaler CAPEX:
The chart below illustrates the capital intensity of the TOP 4 Cloud companies. It is up sharply: Meta is already at 51% and Microsoft is at 46% in Q2. Oracle (not included in the figure) would have been off the chart at above 80% in Q1. The company reports earnings in early September and the key question is how it intends to finance future investments. Some moderation in spending growth is well needed. Amazon is the largest spender with $54.2 billion for the quarter, up 73% y-o-y. Amazon also increased guidance for 2026 capex from $200 billion to $220 billion.
Recent $ Trillion fundraising activities of Anthropic and OpenAI, supported by Nvidia, were widely covered yet both companies continue to lose money. Some financial experts refer to these activities as the “future for financial engineering.” Any innovation comes with some risk, but we all hope for the best.
………………………………………………………………………………………………………………………………………………………………………………………
Addendum: Optical Network Bottlenecks Explained:
-
- InP Epitaxial Production Limits: The primary physical constraint lies in the fabrication and epitaxy capacity for Indium Phosphide wafers used to build high-frequency lasers, where specialized manufacturing equipment (such as MOCVD systems) remains heavily backordered.
- Electrical Trace and Power Walls: Inside massive AI clusters, traditional pluggable optical modules face severe latency, thermal, and signal-integrity walls due to long copper electrical traces running from switch ASICs to the optical engine.
- Packaging and Testing Complexities: Shifting from legacy designs to dense 800G, 1.6T, and co-packaged optics (CPO) introduces intricate multi-die integration and hybrid bonding hurdles that slow down final module validation and packaging yields.
- Expanding Vendor Pools: Cloud builders are aggressively expanding their sourcing lists, moving from a concentrated pool of 2–3 legacy optical vendors out to 5–7 specialized component manufacturers to secure allocation.
- Ecosystem Pivot to Photonics: Major semiconductor players (such as Marvell’s acquisition of Celestial AI) and optical fabricators are pouring capital into silicon photonics and optical interconnect architectures to bypass standard copper and discrete laser constraints.
-
- Mechanism: LPO removes the power-hungry Digital Signal Processor (DSP) and clock/data recovery (CDR) chips directly from inside the optical transceiver module.
- The Fix: Raw, linear analog signals are driven straight from the switch or network interface card (NIC) ASIC.
- Benefits: Slashes module power consumption by 50% or more and cuts latency while preserving the hot-swappable, multi-vendor front-panel pluggable form factor.
- Trade-off: Requires host ASICs with advanced analog front-end and signal-equalization capabilities to process the raw electrical signals.
- Mechanism: CPO moves the optical engine (which converts electrical signals into light) off the front-panel cage and places it directly onto the same substrate or interposer as the switch or accelerator ASIC.
- The Fix: Electrical traces shrink from centimeters down to mere millimeters, completely bypassing lossy copper-clad circuit boards and high-power SerDes requirements.
- Benefits: Reduces optical-interface power consumption by up to 75% and maximizes bandwidth density for ultra-dense GPU scale-up fabrics.
- Trade-off: Reworks serviceability—if an optical engine fails, the repair domain shifts from a simple two-minute transceiver swap to board- or switch-level replacement
………………………………………………………………………………………………………………………………………………………………………
References:
Agentic AI and Inference to Supercharge Front-End Networks Growth, According to Dell’Oro Group
Networking chips and modules for AI data centers: Infiniband, Ultra Ethernet, Optical Connections
Goldman Sachs report: Optical Networking is the next mega trend in AI infrastructure
Cisco Execs: New “Network Supercycle” as Agentic AI Workloads Reshape Telecom Infrastructure
Cisco report: Agentic AI to reshape WAN traffic, AI inference will be ~25% of total traffic by 2035
Meta’s “Iris” AI Chip for MTIA: Implications for Telecom-Grade Optical Networking, DCI and High Capacity Ethernet Fabrics
Oriole Networks photonic networking platform to be integrated with AMD GPUs/CPUs for next-gen AI data center fabrics
Cheap Chinese AI Models: Unappreciated Threat to U.S. Hyperscaler AI Dominance
Huge Risks for the proposed $500B AI Investments from Giant Wall Street firms
Disclaimer: Perplexity.ai was used for research and analysis in this article.
……………………………………………………………………………………………………………………..
Executive Summary:
This past Monday, six giant Wall Street asset managers, private-equity firms and banks announced an effort to raise $500 billion to keep fueling the A.I. boom by financing more data centers, power plants and chips. The proposed platform would direct capital to Nvidia customers—including AI startups and data-center operators—at precisely the point when many have struggled to obtain funding through ordinary credit channels.
We take that as a huge warning sign for the proposed AI investments. Here’s why: If the underlying projects offered clearly proven cash flows, predictable utilization and collateral with durable value, lenders would not need a specially assembled consortium, headline-scale commitments and Nvidia’s direct involvement to make the loans happen. The initiative appears designed to overcome a financing bottleneck created by the extraordinary gap between AI infrastructure spending and demonstrated AI revenue.
This proposed $500 billion AI-financing initiative is less a validation of durable AI economics than an admission that the sector’s spending plans have outgrown its customers’ ability—or willingness—to finance them conventionally. Rather than demonstrating independently sustainable demand, the arrangement risks extending an investment cycle increasingly dependent on vendor-enabled credit, opaque commitments and financial engineering.
The firms—Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR—said they were working together to come up with that huge sum to lend to Nvidia’s customers, including the start-ups that use the company’s chips in data centers to develop and operate A.I. software. These customers, Nvidia said, have been struggling to secure financing for chips and data centers.
Nvidia will connect its customers with one of the six lenders, which will provide financing that could range from loans to credit. The financing will be “at attractive rates,” Nvidia said in a blog post.
In practical terms, this is a vendor-financing mechanism, even if the capital technically comes from third parties. Nvidia is helping its customers obtain the money required to buy Nvidia-dependent infrastructure. The more readily startups and data-center operators can borrow, the more equipment they can order; the more equipment they order, the more Nvidia can sell. That does not mean the demand is fictitious, but it does make it harder to distinguish independent end-user demand from demand supported by an ecosystem that is financing itself.
Circularity is the principal concern. When suppliers, investors, cloud providers, AI labs and lenders all have financial incentives to keep capital circulating within the same small group of counterparties, reported growth can look more robust than the ultimate economics justify. The risk is not simply that projects fail individually. It is that a shortfall in AI-service revenues, utilization or pricing could spread simultaneously through hardware vendors, AI developers, cloud operators, private-credit vehicles and the securities backed by their cash flows. Circular financing can blur the line between real external demand and investment-funded purchases.
Executives from the lenders joined Jensen Huang, Nvidia’s CEO & founder, for an unusual, extended interview on CNBC, where they talked up their new, seemingly insatiable desire to finance infrastructure for A.I. Huang said on the air. He added that “A.I. labs” and “A.I. start-ups” would have access to the financing. He did not name those companies or whether Nvidia would receive any money as part of the effort. David M. Solomon, Goldman’s chief executive, said the consortium was Mr. Huang’s idea.
“We need to raise this money as fast as possible,” said Larry Fink, BlackRock’s chief executive. He also stated: “There’s quite a bit of negativity around A.I. and data centers right now, but let’s be clear: This is going to be creating a huge amount of jobs.”
Skeptical Analysis:
The urgency in those statements deserves more scrutiny. Speed is not a substitute for underwriting, and job-creation claims do not establish that investments will meet their cost of capital. The industry is attempting to finance assets whose useful economic life may be much shorter and more uncertain than that of traditional infrastructure. A GPU fleet can lose competitiveness rapidly when a new architecture, memory standard or systems design emerges. Its resale value can fall sharply if capacity demand weakens. Treating such hardware as collateral comparable to a toll road, utility asset or long-lived building is a major assumption—not an established fact.
The announcement punctuates a head rush on Wall Street and in Silicon Valley into anything that even vaguely resembles A.I. The stocks of tech giants and chipmakers have soared for most of this year, and a pair of the biggest names in the space, Anthropic and OpenAI, are expected to file for initial public offerings that could value them at $1 trillion apiece.
But soaring equity valuations and enormous projected IPO valuations do not answer the basic return-on-investment question: who will pay enough, for long enough, to justify the total cost of the data centers, power plants, networking, GPUs and debt now being assembled? AI vendors may generate impressive revenue growth while still failing to earn enough to cover the depreciation, energy, financing and replacement costs of the infrastructure required to produce it.
Nvidia’s financing narrative appears to conflate broad interest in AI-enabled services with broad, independent demand for capital-intensive AI infrastructure. Governments, enterprises and startups may all seek access to AI capabilities, but the current demand for hyperscale GPU clusters, dedicated power capacity and purpose-built “AI factories” remains concentrated among a relatively small group of frontier-model developers and cloud platforms.
That distinction matters because broad adoption of AI applications does not automatically translate into economically sustainable demand for vast new data-center capacity. Enterprises can consume AI through APIs, hosted platforms, smaller models and open-source software without owning—or financing—dedicated high-density compute infrastructure. Nvidia itself describes demand as spanning enterprises, startups, governments, nations and AI labs, but the critical question for investors is how much of that demand converts into durable, contracted infrastructure revenue rather than experimentation, pilots or subsidized consumption.
As Bloomberg reported last week, OpenAI accounted for somewhere between 50% and 70% of Microsoft’s AI revenue in the 12 months to 30 June.
That degree of concentration is significant. Microsoft disclosed $24.1 billion in sales from OpenAI during the year ended in June; outside estimates place that at more than half, and perhaps about 70%, of Microsoft’s AI-related revenue. Much of the revenue reflects OpenAI’s spending on Microsoft cloud and model-development services, meaning a substantial portion of the apparent AI revenue base may be generated within a tightly linked commercial relationship rather than by a diversified population of external enterprise customers.
That could be problematic if the picture is similar at other hyperscalers and at other frontier AI companies like Anthropic, for example. A market in which a handful of loss-making model developers drive an outsized share of cloud and infrastructure revenue is inherently more fragile than one supported by a broad base of profitable end users. It exposes infrastructure providers and lenders to customer concentration, correlated capital spending and the possibility that a reduction in financing at one frontier lab quickly reduces demand across the supply chain.
Indeed, as has been reported by multiple correspondents – most notably staunch AI critic Ed Zitron – LLM makers like OpenAI are losing money hand over fist and can only charge so much for tokens before customers either curb their usage or switch to open source models. Zitron wrote:
“It’s estimated that 70% or more of the AI revenues of Microsoft, Google, and Amazon were from either OpenAI or Anthropic. UBS estimated that next year, Anthropic and OpenAI’s compute spend would be 48% of all Google Cloud revenues — which means that they likely account for even more than 70% of its AI revenues.”
The commercial challenge is not whether frontier models have value- they do. It is whether they can deliver that value at prices that exceed the combined costs of training, inference, electricity, networking, cloud capacity, depreciation and ongoing model development. If token pricing rises too far, customers may reduce usage, shift workloads to lower-cost models, or use open-source alternatives. If pricing remains low, frontier-model providers may struggle to cover their infrastructure bills. Open-source fine-tuning can materially reduce costs for specialized enterprise workloads, reinforcing the competitive pressure on premium proprietary-model pricing.
The most plausible downside is that the technology becomes broadly useful but insufficiently profitable to support today’s extreme capital intensity. That outcome would leave the sector with too much high-cost capacity, thin margins, declining GPU collateral values and lenders dependent on assumptions about utilization and cash flows that have not yet been tested through a downturn. The foundational financial risk is straightforward: AI-service revenues may prove insufficient to service the fixed obligations incurred to construct the infrastructure. Notably, Nvidia stock dipped modestly on Monday after The Financial Times reported that the company was nearing the mammoth financing deal.
While the contours of the arrangements were announced Monday afternoon, details remained scarce. A joint news release referred only to “memorandums of understanding” to “create dedicated pools of capital at significant scale.”
That language is important. Memorandums of Understanding (MoU’s) are not equivalent to binding, fully funded commitments. Until investors know the actual terms—capital committed, leverage permitted, collateral requirements, Nvidia’s role in losses, loan maturities, customer concentration limits and underwriting standards—the $500 billion figure is better understood as an ambitious financing target than as validated capital deployment. Reporting has described the initiative as a multiyear target rather than cash already committed, while questions remain about the scale of any Nvidia backstop.
During their television interview, the lenders’ executives alluded vaguely to “yield-based products” and even securitization, or the creation of bonds that would divvy up the revenue from A.I. labs into risky and less risky categories. There were several references to A.I. as a new asset class and to allowing smaller investors an opportunity to invest in debt backed by the data centers.
This is where the proposal moves from aggressive investment to potential systemic-risk creation. Securitization can distribute risk, but it does not eliminate it; it can instead diffuse difficult-to-value exposure across a wider investor base. Packaging AI-data-center debt into yield products may make financing more available, but it also risks obscuring the quality of the underlying cash flows, the degree of correlated exposure among borrowers and the vulnerability of rapidly depreciating hardware collateral. The more complex the capital stack becomes, the greater the danger that investors mistake engineered liquidity for genuine economic value.
Executive Quotes:
“NVIDIA has reached an important milestone. We began by building chips; today, we are helping create a new class of productive, investable infrastructure: AI factories,” said Jensen Huang, founder and CEO of NVIDIA. “In AI, compute is revenue. NVIDIA compute is uniquely suited for this role. It is broadly adopted, flexible across models and workloads, fungible and transferable across customers and operators, and continuously improved through CUDA software — extending its useful life and improving its economics over time. It is supported by a deep global ecosystem of developers, customers and offtakers. That is why we are bringing the world’s leading long-term capital providers together to independently underwrite AI infrastructure. These financing platforms will help customers access scarce compute at scale and build the DSX AI factories that will power every industry and country in the age of AI.”
“Modern compute has emerged as a scarce, mission-critical asset class with compelling investment characteristics that is positioned to drive significant long-term economic growth and productivity gains,” said Apollo President Jim Zelter. “The combination of NVIDIA’s proprietary technology ecosystem and Apollo’s flexible, long-term capital base provides a strong foundation to support the next stage of the AI buildout as part of the broader Global Industrial Renaissance.”
“The AI buildout will require unprecedented investment and a skilled workforce to turn that investment into the infrastructure that will help power future growth,” said Larry Fink, Chairman and CEO of BlackRock. “This partnership deepens our relationship with NVIDIA, including through the AI Infrastructure Partnership, and brings together NVIDIA’s leadership in accelerated computing with BlackRock’s ability to connect long-term capital to essential infrastructure. Together, we can help deliver the compute capacity that companies need to grow and create more jobs, supporting the continued growth of the U.S. and global economies, while creating attractive, long-term investment opportunities for our clients.”
“NVIDIA has created extraordinary demand for its compute through an intense focus on customer value and versatile technology,” said Jon Gray, President and COO of Blackstone. “We continue to be enormous investors globally across the NVIDIA ecosystem, and this announcement further underscores our confidence in their platform and the future of AI infrastructure.”
“As our strategic partner, NVIDIA is enabling us to scale AI factories. We are excited about further collaboration to build and fund the backbone of AI globally,” said Bruce Flatt, CEO of Brookfield. “With demand for large-scale AI compute growing significantly as adoption scales across industries, compute is fast becoming the essential layer of infrastructure and a core pillar of the Brookfield AI infrastructure strategy.”
“We’re in a pivotal moment of a historic AI investment cycle. NVIDIA’s full-stack platform is in high demand and uniquely positioned at the center of that global buildout,” said David Solomon, Chairman and CEO of Goldman Sachs. “Our investment and distribution roles reflect our confidence in NVIDIA’s leadership, and we’re excited for the new opportunity to create a market for credit backed by NVIDIA compute.”
“Compute has become a critical infrastructure asset. As we’ve scaled our approach to digital infrastructure, we’ve learned that delivery, not ambition, is the hard part. That’s why we’re excited to build on our strategic partnership with NVIDIA, a founding investor in Helix Digital Infrastructure, to bring together NVIDIA’s accelerated computing platform with KKR’s long-duration capital, infrastructure expertise and capital markets capabilities to turn growing demand into real capacity at extraordinary scale,” said Joe Bae and Scott Nuttall, Co-Chief Executive Officers of KKR.
“This is the very beginning — like what it was when I started in the mortgage-backed securities market in the 1970s,” Mr. Fink said. “I look upon this as a next future for financial engineering.”
That analogy should be treated as cautionary, not reassuring. Financial engineering can expand access to capital and spread risk efficiently when assets have transparent valuations, stable cash flows and conservative underwriting. It becomes dangerous when it is used to finance unproven revenue models, rapidly obsolescing assets and demand forecasts that must remain exceptionally optimistic simply to justify the initial investment.
Huge Risks Explained:
If AI demand stalls, Nvidia faces a sharp reversal in hardware demand and a potentially damaging credit overhang, while lenders could be left financing underutilized data centers secured by equipment whose value can decline far faster than conventional infrastructure. The common vulnerability is that the same uncertain AI revenue streams would be expected to support chip purchases, data-center leases, project debt and securitized investment products. Nvidia is especially exposed to risks:
-
Order cancellations and lower pricing power. Cloud providers, AI labs and startups would likely slow GPU orders, defer deployments or renegotiate capacity commitments. Nvidia could face weaker revenue growth, inventory risk and pressure on the high margins that have supported its valuation.
-
Vendor-financing and counterparty risk. If Nvidia is arranging or backstopping financing for customers buying its systems, a demand slowdown could turn what appeared to be hardware sales into indirect credit exposure. Customers unable to earn adequate returns from AI services may struggle to repay loans used to buy Nvidia equipment. The risk is magnified where the company’s commercial success depends on borrowers gaining access to financing in the first place.247wallst+1
-
Collateral impairment. GPUs are not durable, slow-depreciating infrastructure assets. A new chip generation, a shift toward more efficient models, or weak utilization can materially reduce the resale value of installed systems. If lenders rely on those systems as collateral, a default could leave them holding equipment worth substantially less than the loan balance—and Nvidia could face lower demand for both new and prior-generation products. Moody’s identifies rapid capacity expansion and fast-changing chips, cooling and computing architectures as sources of overbuilding and technology-obsolescence risk.moodys
-
Feedback-loop risk. A slowdown could create a negative cycle: AI customers reduce spending; Nvidia’s sales weaken; lenders become more cautious; financing availability tightens; customers cut orders further. Where vendors, customers and capital providers are financially intertwined, the decline can be more abrupt than a normal inventory correction.
-
Equity-valuation risk. Nvidia’s market value reflects unusually strong assumptions about the longevity of AI spending, margins and growth. A reassessment of those assumptions could compress the valuation sharply even if Nvidia remains profitable. BIS warns that disappointment in AI returns could trigger a sudden financing pullback and turn the capex boom into a prolonged investment bust.bis
Lenders’ exposure:
The most immediate lender risk is a mismatch between long-lived debt obligations and unstable, technology-dependent revenue. A data center might be financed over many years, but its GPU fleet may need continual upgrades to remain competitive—requiring additional capital expenditure before the original debt is repaid. Moody’s notes that this combination of increasing capital intensity, uncertain compute requirements and structured finance can pressure developers, landlords and investors through execution, renewal and refinancing risk.moodys
Construction and power risks:
A demand slowdown could arrive before projects enter service. That creates a particularly difficult situation for lenders because construction interest, cost overruns and power-reservation charges may accumulate before revenue begins.
Permitting delays, local opposition, water constraints and electricity-grid limitations further compound this exposure. Reuters reported that lenders increasingly treat project readiness—including approvals, permits and local community support—as a credit factor because delays can increase costs, jeopardize covenants and prevent projects from reaching revenue-generating operation.
System-wide scenario:
The more serious scenario is a correlated unwind:
-
AI applications fail to deliver enough monetizable demand.
-
AI labs and cloud providers reduce compute commitments.
-
Data-center utilization and expected rental income fall.
-
Borrowers cannot service or refinance project debt.
-
GPUs and related infrastructure lose collateral value.
-
Losses reach private-credit funds, banks, securitized vehicles, insurers and institutional investors.
-
New financing becomes unavailable, causing further cuts in infrastructure orders and Nvidia sales.
This would not necessarily resemble the 2008 banking crisis: some first-loss exposure sits outside regulated banks. But Chicago Booth research estimates that a severe re-rating of AI-related debt could still produce roughly $60 billion to $140 billion in realized credit losses, alongside large equity-market effects.
What matters most:
The decisive question is not whether AI is useful or whether data centers remain necessary. It is whether cash-paying end users will generate sufficient, durable revenue to justify the full cost of the hardware, power, real estate, construction and financing now being committed. If that answer is no, the sector may discover that it financed capacity—not returns.
……………………………………………………………………………………………………………………………………………………………………………………
From Google’s Gemini:
-
- The Revenue Void: If OpenAI or Anthropic defers or cancels capacity commitments, hyperscalers are left with energized Megawatts (MW) that have no immediate, high-paying tenant.
- Stranded Capital: The specialized power, liquid cooling systems, and electrical equipment tailored for dense AI clusters cannot easily be repurposed for traditional cloud workloads without lowering returns on invested capital (ROIC).
| Financial Metric | Baseline Forecast (Bull Case) | Partial Monetization Reset | Severe Crash / Structural Downside |
|---|---|---|---|
| Cloud Revenue Growth | Accelerated (driven by AI APIs & enterprise seats) | Flattens out as enterprise customers consolidate suppliers | Decelerates sharply; circular cloud revenue loops collapse |
| CapEx Infrastructure | Full execution of Dell’Oro’s ~200 GW build out by 2030 | 20–30% of capacity commitments deferred or resized | Mass cancellations of orders; multi-year build freezes |
| Operating Margins | Expands as inference costs decline relative to software prices | Compresses due to underutilized GPU clusters and high DCPI fixed costs | Severe contraction; massive asset write-downs and depreciation drag |
| ROIC | Historic highs driven by high-density compute demand | Diluted; longer payback periods for specialized data centers | Tanks; billions in stranded physical capital and obsolete hardware |
References:
https://www.nytimes.com/2026/08/10/business/ai-nvidia-lenders-500-billion.html
https://www.wheresyoured.at/dont-look-up/
Curmudgeon: Caveat Emptor: Huge Debt and Circular Financing Deals Dominate AI Build-Outs (07/23/26)
Merry-go-round of dog chasing its tail: Relationship between U.S. hyperscalers and private Gen AI companies
AI infrastructure spending boom: a path towards AGI or speculative bubble?
Expose: AI is more than a bubble; it’s a data center debt bomb
Will Google Cloud’s AI and data analytics revenue +TPU IP licensing income offset huge AI CAPEX to produce a decent ROI?
Amazon’s Jeff Bezos at Italian Tech Week: “AI is a kind of industrial bubble”
Big Tech AI spending binge results in massive job cuts!
AI spending boom accelerates: Big tech to invest an aggregate of $400 billion in 2025; much more in 2026!
FT: Scale of AI private company valuations dwarfs dot-com boom
Big tech spending on AI data centers and infrastructure vs the fiber optic buildout during the dot-com boom (& bust)
AI Data Center Boom Carries Huge Default and Demand Risks
Can the debt fueling the new wave of AI infrastructure buildouts ever be repaid?
Gartner: AI spending >$2 trillion in 2026 driven by hyperscalers data center investments
Will billions of dollars big tech is spending on Gen AI data centers produce a decent ROI?
