AvidThink 2026 September Recap

September was another month of back-to-back conferences. First, AI Infra Summit in Santa Clara, then FutureNet Asia in Singapore, followed by CoreWeave’s Fully Connected event in San Francisco.

If you missed the AI Infra Summit, don’t fret because we have a full showcase of ~30 videos for you to browse (with a few stragglers that we’re still working through). Check out the interviews that Jim grabbed on the show floor — Astera Labs, Ayar Labs, Broadcom, Cornelis Networks, Credo, d-Matrix, DriveNets, Lightmatter, Marvell, Nutanix, Qualcomm, Rebellions, NVIDIA, Synopsys, UALink Consortium, and many more. Suggestions for how to further improve? Drop us a line at [email protected]. We’d really appreciate it.

Upcoming events for us in October include Roy participating in Network Field Day #41 in Santa Clara Oct 6-9, OCP Global Summit (Oct 12 to 15) in San Jose, then PyTorchCon (co-located AGNTCon + MCPCon) (Oct 20-21) in San Jose, followed by Megaport Connect (Oct 22) in Chicago. After that, we have FutureNet Americas in New York, KubeCon North America in Salt Lake City, Network Automation Forum AutoCon 6 in Tucson, before we break for the US Thanksgiving holiday.

Our upcoming reports include the AI for Networking (AIOps and automation) report slated for later this month in Oct, our Data Center Networking Report 2026 Fall Edition in the Nov/Dec timeframe, and the Enterprise Connectivity Report we’re actively working on, which will ship sometime in Q4 (most likely).

If you’re interested in participating in any of these (or the others we have in our pipeline covering Telco AI factories, NTN, and Physical AI), drop us a note at [email protected]. Or fill in the form at: https://nextgeninfra.io/contact/.

Here's what's in store for this month's newsletter:

Don’t forget to visit NextGenInfra.io and check out the latest AI Infra Summit videos!

—The AvidThink Team

Before I dig into my travel from last month, I want to do a last-minute callout about a webinar on 6 October (Tuesday) at 1500 CEST (0900 US Eastern), where I join some well-known telco pundits for a candid discussion on the real state of autonomous networks. Why have many operators backed away from their autonomy levels, moving from L4 to L2+? And can we collectively get past L2? Should we? Join Mark Newman (TM Forum), Dean Ramsay (Omdia), Tony Poulos (multi-hyphenated telco expert), host Christian Maitre (Caspera-Lab), and me for 30-40 minutes of debate. Sign up here.

AI Infra Summit (Santa Clara, 8,000+ attendees)

The AI Infra Summit has gone by several names over the years (AI Hardware & Edge AI Summit, EDGE AI Summit, AI Hardware Summit) since Kisaco Research first kicked off the event at the Computer History Museum in Mountain View in 2018 with 200+ attendees. Until the last two years, it had been a relatively niche show. Not anymore. Kisaco says the 2026 show drew 8,000+ registrants onsite. With headliners David Patterson (Google, ex-Berkeley professor), Lip-Bu Tan (Intel), Ian Buck (NVIDIA), Peter DeSantis (AWS), and others, it’s definitely become a must-attend show in the AI line-up.

Its edge and IoT roots still show through, given the ample presence of IoT and physical AI solutions on the show floor: Ambarella, Synopsys physical AI offerings, Cognichip, and others. But the organizers now tout it as a full-stack AI show. Analysts and media alike bore the full brunt of a news and briefing deluge from silicon, optical, networking, storage, compute, ML models, harnesses, and AI applications vendors. It was fun, nevertheless, to see the trends across the board. There’s clearly interest in inferencing at the edge (Qualcomm was all over that, along with other startups pushing custom silicon, inferencing silicon, re-programmable silicon) and physical AI (the next big thing, supposedly). Many are hoping that the diverse, distributed, and fragmented inferencing market, which has been estimated at 4-5X the pre-training opportunity, will be open to NVIDIA competitors looking for a foothold. The need for on-device inferencing with highly responsive, energy- and compute-efficient models/platforms is clear. The business case for edge-based inferencing (near, far, local) is still a little murky to me — at least until some AR/VR-type or drone-video augmentation workloads become more prevalent.

Networking vendors were out in full force, pushing scale-up, scale-out, and scale-across networking, along with optics improvements. You can watch the 30+ recorded interviews and briefings on the NextGenInfra.io showcase. And special thanks to Sid and the d-Matrix for hosting us at their offices next door for lunch and a lab tour! Congrats again on securing the unique partnership with NVIDIA.

FutureNet Asia 2026 (Marina Bay Sands, Singapore)

Kudos to the FutureNet World team for a successful event! The event’s grown again since last year (official numbers to come later),, and the APAC telco 'executives' willingness to be candid and share real-world experiences was much appreciated. Never one to pass up an excuse to visit home in Singapore and hang out at the Marina Bay Sands, it was well worth making that 20-hour trek both ways (yes, I know there’s a non-stop from SFO, but my butt and back hurts when I do that — aging isn’t kind).

Congrats to Kim Krogh Anderson of Telstra on winning the Technology Leader of the Year Award, and thanks to him for helping set the tone (along with the always stellar Thomas Hundt of Axiata and insightful Prakash Parekh of Amdocs). By honestly (and bravely) sharing that after a TM Forum re-benchmark, Telstra is "embarrassingly below 2" where the team had previously self-assessed at 3–3.5, he allowed everyone else to be open about their autonomous network efforts. The rest of the content was predictably evolutionary, with speakers sharing their progress on AI automation, though costs and sovereignty made multiple appearances. For a quick recap of key conference takeaways, I’d encourage you to scroll down to the Webinars section below and watch the closing analyst session that I had with S&P Global Market Intelligence's Divya Keshwani and Appledore's Shanthi Ravindran. I really enjoyed that chat with Divya and Shanthi!

For me, the key takeaway was Kim (Telstra)’s clarion call for telcos to rethink what we’re selling to our customers: if the hyperscalers can repackage compute and storage (and other infrastructure services) into a rich portfolio of on-demand offerings, why can’t we?

CoreWeave Fully Connected 2026 (Moscone Center, San Francisco)

wrapped up September with a trek up to Moscone Center in San Francisco for the neocloud show put on by CoreWeave (the largest neocloud and a public company) and its partners. It was two days of CoreWeave and friends showcasing their customer traction and touting their uniqueness beyond just being a GPU farm. Two major announcements were that CoreWeave customer Cognition was the first in the world to run production workloads on NVIDIA Vera Rubin NVL72, and CoreWeave Forge. This unified environment accelerates agent development by bringing together training, inference, evaluation, and agent development.

Some fun takeaways were CoreWeave demonstrating innovations in the move to liquid-cooled racks (we expect to see more of this at OCP Global Summit shortly) with Racky and Valvey, their row-level liquid cooling monitoring and management solution — complex challenges dynamically managing fluid (PG25 - 25% propylene glycol + 75% pure water + corrosion inhibitors) that could short and destroy the expensive electrical systems of dense GPU racks.

And that’s a wrap for this month. Up next for me are Network Field Day 41, then OCP Global Summit, PyTorchCon, Megaport Connect (please register) in Chicago, FutureNet Americas in New York, KubeCon North America in Salt Lake City, AutoCon 6 in Tucson, and more! Hope to see you at one of these events.

If you are interested in any of the reports we’re working on (check out our report roadmap), or if you would like to talk and catch up, drop us a line at [email protected].

– Roy for the AvidThink Team

Check out our most recent articles on the AvidThink website:

Check out the ~30 videos we have from interviews at the AI Infra Summit 2026 in Santa Clara. Our media partner, Jim Carroll, braved the 8,000 (or so)-strong crowds to secure these interviews with key executives at hot startups, established infrastructure leaders, and all involved in shaping the next generation of AI infrastructure — from the data center to the edge.

Featuring Astera Labs, Ayar Labs, Broadcom, Cornelis Networks, Credo, d-Matrix, DriveNets, Lightmatter, Marvell, Nutanix, Qualcomm, Rebellions, NVIDIA, Synopsys, UALink Consortium, and many more companies.

3GPP Defers Decision on Migration from 5G to 6G
At its September meetings in Madrid, 3GPP left unresolved whether 6G should use a lightly modified 5G core, a substantially evolved core with new functions, or entirely new infrastructure. The choice will affect operators still deploying 5G standalone (SA) and the introduction of AI-native orchestration through a dedicated AI domain in the core. 3GPP has until March 2027, when 6G work moves from studies to standards development, to resolve the migration approach.

T-Mobile Seeks 6G Tests in 2.7GHz and 4GHz Bands
T-Mobile has requested FCC permission to test 6G use cases in the 2.7GHz and 4GHz bands, joining Verizon and other operators pursuing early mid-band trials. Propagation and interference data from these tests could influence 3GPP standards and future FCC spectrum allocations, although commercial 6G standards and radios remain years away.

Samsung Opposes Major 6G Redesign for Network Sensing
Samsung argues that redesigning 6G for integrated sensing and communication (ISAC), which uses radio networks to detect and track objects, would add excessive cost and complexity for limited use cases. Its position contrasts with vendors promoting sensing as a major 6G capability and highlights 3GPP’s debate over transformative features versus an evolutionary approach based on 5G-Advanced.

Verizon Seeks FCC Approval for 4GHz 6G Trials
Verizon has requested FCC approval to test 6G use cases in the 4GHz band, gathering propagation and application data ahead of commercial deployment expected around 2030. Its overlap with T-Mobile’s requested spectrum signals potential competition for mid-band frequencies adjacent to existing commercial and federal uses.

GSMA Projects 2.4 Billion 6G Connections by 2035
GSMA forecasts approximately 2.4 billion global 6G connections by 2035, with standards expected to be finalized around 2030. The projection implies faster early adoption than 5G and a more mature infrastructure and device ecosystem at launch, while raising investment questions for operators still deploying 5G SA and seeking returns on existing networks.

T-Mobile and Jio Claim First Commercial 5G SA International Roaming Connection
T-Mobile and Reliance Jio say they completed the world’s first commercial 5G standalone international roaming connection between the U.S. and India, preserving voice over New Radio (VoNR) and faster session setup. Coordinating their 5G cores and signaling enabled subscriber recognition and SA capabilities abroad, addressing interoperability gaps that typically force roaming devices onto 4G or non-standalone 5G (NSA).

GSA Reports More Than 2,000 Private Mobile Network Customers
The Global mobile Suppliers Association reports that private mobile networks now serve more than 2,000 customers worldwide across manufacturing, logistics, mining, ports and other sectors. These 4G and 5G deployments often incorporate edge computing and AI automation, while adoption is broadening beyond heavy industry. Celona, Nokia and T-Mobile are introducing private wireless and AI-management platforms as the industry weighs dedicated networks against enterprise slices on public 5G-Advanced networks.

NTT Docomo Deploys Nokia Open RAN Base Station
NTT Docomo deployed a Nokia open RAN base station as part of its effort to demonstrate that interoperable, multi-vendor networks can match traditional single-vendor performance. Its disaggregated architecture aims to reduce costs and supplier dependence while encouraging common interfaces. SNS Insider estimates the AI-native open RAN market will grow from $3.2 billion in 2025 to $24.5 billion in 2035.

Rakuten Mobile Positions Open RAN as a Foundation for AI
Rakuten Mobile says its software-defined, cloud-native open RAN architecture simplifies AI integration by running virtualized network functions on commodity infrastructure. At the Intelligent RAN Forum, RAN head Sudhakar Pandey said the network has expanded from its 2019 O-RAN launch to more than 152,000 sites and over 330,000 cells. Originally intended to reduce vendor lock-in, the architecture now supports broader AI automation, with human oversight retained for higher-risk decisions.

U.S. Carriers Diverge on GPU Investment for AI-RAN
T-Mobile expects Nokia AI-RAN prototype radios later this year through its work with Nokia and Nvidia; CTO Ankur Kapoor reports early spectral-efficiency gains of at least 30%, with Nokia targeting 50% by the end of 2027 and 100% by the end of 2028 to improve fixed wireless capacity and speeds. Verizon CTO Yago Tenorio favors GPUs for edge applications such as smart glasses but expects CPUs to remain sufficient for radio processing and embedded AI inference for several years. AT&T, whose main RAN vendor is Ericsson, is evaluating Nvidia and other GPU platforms but requires a business case for GPU adoption or hosting third-party workloads at cell sites.

T-Mobile Uses AI to Strengthen 5G Network Resilience
T-Mobile is extending its self-optimizing network (SON) capabilities with AI that continuously detects anomalies and automatically remediates problems during outages, severe weather and other disruptions. Tests of AutoPilot, its DynamicCX platform, showed shorter resolution times for outages and power failures. The operator says its early 5G SA deployment provides a cleaner data foundation for automation than networks retaining substantial NSA infrastructure.

Optus Reports AI-RAN Field Gains with Ericsson
Optus reports that its Ericsson AI-RAN deployment has improved link adaptation, coverage prediction and interference compensation, demonstrating machine learning applied directly to radio-layer decisions in commercial operation. At RCRTech’s Intelligent RAN Forum, panelists cautioned that nationwide deployment will require investment in data pipelines, model management, staff training and operational engineering.

Orange Moves Nokia AI-RAN into Live Network Testing
Orange is preparing to test Nokia’s AI-RAN platform on its live commercial network to assess radio-efficiency gains under real traffic conditions. The trials will also evaluate sharing GPU infrastructure between radio optimization and other AI applications, a key economic consideration because dedicated GPUs for radio optimization alone can be difficult to justify.

KT and SK Telecom Select Samsung for AI-RAN Projects
KT and SK Telecom, South Korea’s two largest mobile operators, have selected Samsung for AI-RAN initiatives, strengthening its position against Nokia, Ericsson, Huawei and ZTE as traditional hardware margins decline. The projects support Samsung’s strategy to embed AI in the radio layer and complement the carriers' broader AI infrastructure ambitions, including sovereign AI computing.

SK Telecom Urges Service-Led Investment in AI-RAN
At RCRTech’s Intelligent RAN Forum, SK Telecom 6G tech team director Dr Dongwook Kim warned that AI-RAN’s biggest risk is repeating early 5G investment mistakes. He cited limited revenue growth from edge computing and network slicing because few services required them, and urged operators to define real-time connectivity-plus-inference services before evolving their networks.

Agentic AI Adoption Outpaces Telecom Governance
An Omdia/Cisco study finds that telecom operators are adopting agentic AI, which takes autonomous actions, faster than many executives expected. Agents already handle tasks such as anomaly detection and configuration changes, but trust, governance and oversight processes lag deployment. The findings reinforce NGMN Alliance warnings about advancing autonomous network operations without adequate controls.

China Telecom Expands AI Token and Inference Business
China Telecom is expanding sales of large language model (LLM) inference capacity and plans to make AI tokens central to its business after becoming the first operator to sell them in May, followed by its two domestic rivals. Its strategy uses data-center and network capacity to serve enterprises and developers, reflecting Chinese carriers' shift toward AI infrastructure. Analysts caution that volatile token pricing and thin margins could undermine returns, while China Telecom is betting on scale and integration with its own infrastructure.

Nokia Embeds Autonomous AI Agents in Mobile Core Functions
Nokia is placing smaller AI models alongside mobile core network functions, enabling edge inference and autonomous decisions without human intervention. It says operational and security safeguards are essential, echoing NGMN Alliance calls for guardrails as operators seek lower costs and faster responses. Nokia is also using generative and agentic AI internally for root-cause analysis, code generation and productivity.

Huawei Targets Nvidia-Like Role in ICT Infrastructure
Huawei Chairman Guo Ping says the company aims to become the Nvidia of information and communications technology and describes AI as the final technology revolution. Its expansion into AI chips and computing infrastructure follows U.S. export restrictions limiting access to Nvidia hardware, with Huawei targeting domestic and international customers. Guo ruled out manufacturing satellites or rockets, pursuing space exploration, or building Huawei’s own cars.

AT&T, Boost Mobile and GSMA Warn of AI Train-Serve Skew
Experts from AT&T, Boost Mobile and GSMA warn that AI models can degrade in production when live data differs subtly from training data, a problem known as train-serve skew. Boost Mobile data scientist Priyank Jain and GSMA AI technologies director Louis Powell caution that the resulting accuracy loss can remain unnoticed until a visible failure occurs. Operators deploying AI for network operations and customer service should build monitoring and retraining pipelines into production systems from the outset.

Ofcom Finds 8% of Online Adults Use AI to Manage Telecom Services
Ofcom reports that 8% of UK online adults use AI to help manage phone, broadband or pay-TV services, indicating growing uptake of chatbots, virtual assistants and account-management tools despite mixed customer opinions. Providers remain cautious, applying guardrails and limiting AI functions. Ofcom expects more capable agentic systems could compare broadband deals, prepare transactions for approval, monitor accounts and eventually switch customers to better offers.

Rebellions NPUs Power Four SK Telecom Consumer AI Services
South Korean AI-chip startup Rebellions says its neural processing unit (NPU) infrastructure has moved from pilots into production across four SK Telecom consumer AI services after nearly three years of collaboration. The company claims its systems cost roughly one-half to one-third as much as comparable Nvidia hardware and offer power-efficiency advantages. Rebellions is using the SK Telecom relationship to pursue other operators globally.

Cloud

Dell’Oro: Data Center Capex Rose 92% in Q2 2026
Dell’Oro Group reports that global data center capital expenditures increased 92% year over year in Q2 2026, driven by AI infrastructure demand and rising memory-chip costs. Hyperscalers and neoclouds are competing for GPUs, high-bandwidth memory and power as component prices climb. Dell led OEM revenue for the quarter, followed by Supermicro and Lenovo, while white-box server revenue reached a record high.

AI Cloud Provider Nscale Raises $3.36 Billion Ahead of IPO
Nscale raised $3.36 billion as it prepares for an IPO, marking one of the largest financing rounds among neoclouds built around GPU-based AI infrastructure. The funding signals continued investor appetite despite uncertainty over whether enterprise demand for training and inference can sustain current investment levels. An IPO would make Nscale a prominent test of public-market demand for neocloud businesses.

Microsoft Commits $10 Billion to Middle East Cloud and AI Infrastructure
Microsoft has committed $10 billion to expand cloud and AI infrastructure across the Middle East, joining regional investments by AWS, Google and other providers. Affordable energy, sovereign wealth capital, enterprise and government AI demand, and data-sovereignty requirements are encouraging local infrastructure development. The commitment also supports Microsoft’s wider effort to secure GPU capacity and data-center space.

BNP Paribas Signs Five-Year Google Cloud and AI Deal with Hybrid Approach
BNP Paribas has agreed to a five-year Google cloud and AI partnership while reportedly retaining its own infrastructure and keeping sensitive customer data and critical operations outside the public cloud. Reuters reports that confidential information, including medical data held by its insurance business, will remain outside the public cloud. Its corporate and institutional banking group will use Google Cloud, Gemini models and Gemini Enterprise, integrating Gemini into its in-house generative AI assistant.

If you would like to meet with AvidThink at one of the following upcoming events, please click on the event below to request a meeting.

if this newsletter was forwarded to you, why not sign up for your own subscription?