AvidThink 2026 August Recap

(Featured Image - clockwise from left: image from Roy’s recent mystery travels; Meta’s Networking@Scale; IEEE Hot Chips at Stanford; the salad that may or may not have been the source of Roy’s recent gastrointestinal adventures; Jim Carroll and Roy at FMS 2026; Andrew Feldman, CEO Cerebras, with Jayshree Ullal, CEO Arista, at SuperNova)

We apologize for the delayed send this month — Roy traveled the first week and caught a bad stomach bug on his return. He’s finally well enough to submit his travelogue.

August saw AvidThink and our media partner, Converge Digest (Jim Carroll), at the Future of Memory and Storage (FMS) conference in Santa Clara, CA; the Cerebras SuperNova event; and the online IEEE Hot Interconnects and IEEE Hot Chips (Stanford). Upcoming in the next few weeks are the AI Infra Summit in Santa Clara, where we have dozens of interviews scheduled, and FutureNet Asia, where Roy will be hanging out with Asia-Pacific carriers. Unfortunately, we will miss Mplify Member Summit - Americas in Boston because that’s the week of FutureNet Asia. Still, we can’t wait to see who won the NaaS Excellence Awards 2026 (Roy was one of the judges).

And if you haven’t checked out the revamped NextGenInfra.io, try the improved navigation, including browse by topic and browse by company. We’re continuing to make upgrades. This coming week, we’ll launch the AI Infra Summit showcase as we grab video interviews from the show floor. Suggestions for how to further improve? Drop us a line at [email protected]. We’d really appreciate it.

Our upcoming reports include the AI for Networking (AIOps and automation) report slated for Oct and our Enterprise Connectivity (SASE, SD-WAN, Campus) slated for Nov. And our Data Center Networking Report 2026 Fall Edition will be in the Nov/Dec timeframe (after OCP Global Summit) (how’s that for creativity).

In parallel, we’re looking for sponsor interest/content for our (1) Satellite/Non-Terrestrial Networks (NTN), (2) Next-Gen Edge Computing with AI Inferencing, and (3) Physical AI and the Role of Telcos. If you’re interested in participating in any of these, 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 give us your feedback on our new site.

—The AvidThink Team

First, apologies for delaying our August recap by over a week. I was traveling for a client project in late August/early September, and when I returned, a stomach bug took me out of commission (with remarkably similar symptoms to the recent outbreak that the US FDA just declared over). I’m just glad that I’m back on my feet again. As for where I traveled, you can peek at the photo collage below for a hint. In one of the places, there’s clearly multiple generations of mobile technology - from 4G to (gasp) 7G! Forget 6G, that’s passe.

Anyhow, August saw the beginning of the Fall conference season, with a series of technology-centric events including FMS (formerly known as the Flash Memory Summit but now aptly renamed Future of Memory and Storage) in Santa Clara, Cerebras’ private SuperNova event in San Francisco, then the virtual IEEE Hot Interconnects, the in-person IEEE Hot Chips at Stanford, and in the same week, Meta’s Networking@Scale in Santa Clara. Plenty of content and learnings, and a chance to catch up with friends in the space and make new ones.

FMS 2026 (Santa Clara, ~3,500 attendees)

FMS is essentially an AI memory-and-data infrastructure event now. Attendance has spiked (unsurprisingly). What used to be a niche show under Lance Leventhal and Chip Stockton is now another AI show from London-based Terrapinn (which acquired it). When you thought you were AI-ready and understood what HBM (high bandwidth memory) was, along comes High Bandwidth Flash (HBF), a new memory layer that sits between ultra-fast HBM and high-capacity SSDs. We’ll hear more about HBF at OCP Global 2026 in October since that’s who will host the standard. As you can imagine, plenty of discussion around KV-cache placement, inference economics, and how to keep those GPUs fed. As well, plenty of optimization across every tier of the memory/storage hierarchy. BTW, another new taxonomy for you AI watchers to track: the context memory G1-G4 taxonomy NVIDIA formalized during its development of the Dynamo framework, and gaining in popularity.

Other takeaways for me from FMS include multiple shipping PCIe 6.0 products and PCIe 7.0 test gear on the horizon, CXL coming back into the spotlight for memory pooling and sharing within AI inference, and GPU-initiated SSD access (yes, NVIDIA’s involved again, and it’s called SCADA - no, not the one for industrial controls).

Cerebras SuperNova (San Francisco)

Super buzzy, long queue to get in for the in-person event, plus a bunch more people on the virtual webstream. Most of the event was a rehash of content and positioning for those not already familiar with Cerebras, and I almost gave up. A highlight for me was seeing Arista’s Jayshree Ullal joining Cerebras CEO Andrew Feldman on stage (Arista powers their rack-level networking). But our patience was rewarded when CTO Sean Lie wrapped up and finally unveiled their new CS-4 rack with its innovative Backpack design.

CS-4 is claimed to have 2x the throughput of CS-3, with the ability to push 1,000+ tok/s on frontier-class parameter models (like those from OpenAI). Note that the new CS-4 is built to house three of their current-generation WSE-3 Turbo processors (900K cores each, 44 GB on-wafer SRAM, 43.2 PB/s memory bandwidth). The rack delivers a claimed 750 PFLOPs, 7.2 Tb/s system I/O, and wafer-to-wafer latency as low as 2 µs.

The company’s positioning is that high-speed interactivity unlocks different modes of operations and encourages new use cases. For example, OpenAI’s Thibault Imbert was on stage to share how GPT-5.6 Sol's "Ultrafast mode" is enabled by Cerebras. And how this removes the speed-vs-intelligence tradeoff (substituting cost instead, I’m sure).

IEEE Hot Interconnects 2026 virtual (1,500 registrations with 1,000+ live attendees)

This virtual event was kicked off by Omar Baldonado of Meta. Omar shared how Meta's dedicated training environments have progressed from ~24K–32K GPUs to a 129K-GPU system and now toward multi-building, multi-region gigawatt-scale deployments. The co-design theme that pervades all AI conferences continued here, with Omar stressing that networking can no longer be designed independently from GPUs, CPUs, PCIe, CXL, memory, storage, cooling, power, or workload scheduling.

NVIDIA’s Gilad Shainer shared their full-stack vision, from NVLink for scale-up, Spectrum-X/InfiniBand for scale-out, Spectrum-XGS for scale-across, BlueField-4 for infrastructure processing and an emerging context-memory storage tier (CMX). He also shared that Spectrum-X Ethernet Photonics CPO is now in production.

Four lightning talks covered multiple optics-related standards (MSAs), including OCI (optical scale-up), XPO (extreme-short-reach pluggables toward 25.6T), Open CPX (socketed optical engines), and SDM4 MCF (four-core multicore fiber for 4x fiber density in scale-out).

Key takeaway from HOTI is that many of the engineering concepts across the scale-up, -out, and -across domains will be increasingly shared, with continued optimization spanning latency, energy per bit, cost, throughput, and copper-vs-optics (esp. on scale-up/out).

IEEE Hot Chips 2026 (Stanford)

I didn’t catch how many live or online attendees there were this year, but it felt crowded in the outdoor booths area (the main hall wasn’t always packed, but audience size was healthy). Perhaps the best indicator of how oversubscribed it was was the super long lunch line.

As an ongoing example of the extreme co-design theme, multiple presentations showed that the rack/system is the new unit of AI compute. The other theme was a focus on “agentic AI” where CPUs are once again ascendant because of tool use, tool creation/code compilation, and non-ML work. We saw various architectural blends of GPUs with CPUs in different presentations.

Also, given the focus on the rack as the computing unit and the spotlight on AI workloads, we saw more mentions and comparisons of tokens per MW as the new metric, with power efficiency now playing an increasing role (remember, power is still the binding constraint across the AI ecosystem).

Some cool presentations you can watch in the online archives are the Waymo talk on its custom 5-nanometer sensor fusion chip designed to process raw data for its robotaxis.

Meta’s Networking@Scale (Santa Clara)

Back at the Santa Clara Convention Center (where FMS was just held), Meta held its annual Networking@Scale conference. Omar Baldonado was one of the keynotes and went in greater depth than the topics for HOTI. A new element for the @Scale conference was the use of AI for networking as opposed to networking infrastructure for AI data centers. AIOps and AI for data center networking planning were additional topics covered at the conference.

The @Scale conference juxtaposed two overlapping approaches to succeed RoCEv2 and to bridge us to UEC/UET (UltraEthernet). First, Multipath Reliable Connection or MRC (now hosted at the OCP and proposed by OpenAI, AMD, Broadcom, NVIDIA, and Microsoft) is an RDMA-based transport that uses multi-pathing and offers SRv6 static source routing for path determinism (to reliably detect and bypass failures and congestion). The presentation shows the overall system's resilience in the face of network failures. Then, Meta showcased MetaRoCE—many similar concepts to MRC, like multi-pathing, out-of-order delivery, etc. However, MetaRoCE simplifies the network requirements to just ECMP and ECN support and allows multiple independent logical streams across a single connection, each with its own congestion window controls. We’ll see how the ecosystem embraces either or both.

Part of the conference took a holistic debugging perspective, starting at the communications layer and moving down to hardware, across both front-end and back-end networks. Finally, operations and observability are catching up to the build-out.

And that’s a wrap for this month — again, sorry this was so delayed. I’m off to AI Infra Summit this coming week, then FutureNet Asia. Next up are Corewave’s Fully Connected conference, Network Field Day 41, then OCP Global Summit in October, plus more events after.

In the meantime, 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:

Coming up this week are interviews captured on the show floor of the AI Infra Summit 2026 in Santa Clara. Stay tuned!

5G

Verizon 5G SA Test Shows 2.2x Uplink Efficiency Gain
Signals Research Group testing on Verizon’s commercial Ericsson network found 5G standalone (SA) delivered 2.2x higher upstream spectral efficiency than non-standalone (NSA), potentially giving operators more than twice the uplink capacity from the same spectrum. The result strengthens the case for SA as video analytics, industrial robotics, AI services, and other uplink-intensive applications grow.

Lockheed Martin and Verizon Turn 5G Network Into Drone Detection System
Lockheed Martin, Verizon, Nvidia and partners tested NetSense Airspace Awareness-as-a-Service, which uses Verizon’s commercial 5G network, ODC Networks AI-RAN software, Nvidia AI Aerial and Keysight RF simulation tools to detect and track drones without purpose-built radar. Following successful Miami testing, pilots are planned for later 2026 and early 2027, with commercial availability for select customers targeted by the end of 2027.

Ookla Finds Urban-Rural Performance Gap in U.S. 5G FWA
Ookla found rural median 5G FWA download speeds trailed urban performance by 26–49 Mbit/s across AT&T, T-Mobile and Verizon in Q2, with rural latency 7–13 ms higher. AT&T delivered 174.2 Mbit/s urban versus 132.6 Mbit/s rural, Verizon 148.3 versus 99.3 Mbit/s, while T-Mobile led overall with the smallest gap at 230.7 versus 204.7 Mbit/s.

EE Launches Consumer 5G SA Network Slice With Fast Lane
BT-owned EE launched Fast Lane, a roughly $7-per-month service that moves eligible customers onto a dedicated slice of its 5G SA network during congestion, following two years of trials at high-traffic venues. Ericsson’s June 2026 Mobility Report counted 84 commercially available SA slicing offers, up from 65 six months earlier, including roughly 34–35 from European operators.

Nokia Shifts Investment Toward AI and Optical as China Business Contracts
Nokia is redirecting investment toward AI infrastructure and optical networking as it reduces its China presence and moves to divest its enterprise campus edge private-networking business. Its AI and cloud segment generated $521.8 million in quarterly sales, up 105% year-over-year and representing 9% of revenue, while Nokia confirmed closure of its Hangzhou R&D facility, eliminating about 1,600 jobs amid a broader reported restructuring of its roughly 7,200-person Greater China workforce.

6G

Industry Aligns Around AI-Native, Standalone 6G Networks
Ahead of 3GPP’s September 2026 plenary meetings, the industry is converging around three 6G principles: standalone from day one, AI-native by design and increasingly service-aware networking. A key unresolved issue is whether 6G requires an entirely new core or an evolution of 5G, potentially using “service intent” and AI orchestration to dynamically configure network functions and AI agents around application requirements.

Japanese Operators and Vendors Secure ITU-T Framework for All-Photonics Networks
KDDI, NTT, Rakuten Mobile, NEC, 1Finity, Oki Electric and Sumitomo Electric secured ITU-T approval for an all-photonics network framework defining configurations and technical requirements for low-latency, energy-efficient optical connectivity across operators and countries. The work supports Japan’s Beyond 5G strategy, with a follow-on ITU-T recommendation defining detailed functional architecture targeted for completion by October 2028.

AI

Nvidia Reportedly Acquires Hugging Face for $12.9B
Nvidia reportedly acquired open-source AI platform Hugging Face for $12.9 billion, extending its position across models, datasets, libraries and AI developer distribution. Analysts see potential benefits for open-model funding but risks to hardware neutrality, while the roughly 86x-revenue valuation and Nvidia’s expanding control of the AI stack are likely to attract antitrust scrutiny in the U.S., EU and UK.

NGMN Calls for Guardrails Before Telcos Embrace Agentic AI Autonomy
The NGMN Alliance is warning operators against moving too quickly toward autonomous agentic AI for network operations, arguing that governance and safety controls must precede deployment in telco-grade infrastructure. Its position contrasts with vendors rapidly embedding agents into network management, orchestration and customer-service platforms, highlighting growing tension between autonomy and operational safety.

AI Network Buildouts Face Capacity, Permitting and Supply-Chain Constraints
Executives and policy experts from Lumen, AT&T and others warned that AI network expansion faces capacity, permitting and supply-chain bottlenecks, with Lumen identifying permits as its biggest challenge. Lumen plans to expand from 17 million fiber miles at the end of 2025 to 58 million by 2031, while the FCC’s ban on new foreign-made routers is adding approval and component complexity for equipment vendors.

SK Telecom Creates SK Horizon for AI Data Centers and Submarine Cables
SK Telecom is carving SK Broadband’s AI data center and submarine cable assets into SK Horizon, with KKR and an IMM Investment-Stonebridge consortium investing about $2.22 billion for 29% and 20% stakes, respectively; SK Telecom retains 51% and management control. SK Horizon will manage existing infrastructure and expansion, complementing SK Hyper’s plan to develop up to 15 GW of new AI data center capacity by 2035, including a gigawatt-scale Ulsan campus with AWS.

IDC: Only 13% of Enterprise Leaders Widely Understand Sovereign AI
An IDC survey commissioned by Cohere found only 13% of more than 500 senior AI decision-makers at $1 billion-plus enterprises said sovereign AI was widely understood internally, despite 99% using generative AI and 86% using embedded AI agents. Among respondents able to define it, 52% associated sovereignty with local or national control and 35% with digital independence, while 37% of telecom respondents viewed sovereignty as a competitive differentiator.

Huawei Boosts R&D Spending 25% as Profit Falls 36%
Huawei’s first-half revenue increased 9.6% year-over-year to $69.5 billion, but net profit fell 36% to $3.5 billion as R&D spending jumped 25% to $18.1 billion, nearly 26% of revenue. The spending supports Huawei’s push into AI computing, chip design and communications technology as it seeks to reduce dependence on foreign technology.

AWS Expands Nvidia Commitment by Another 2 Million GPUs
AWS agreed to deploy 2 million additional Nvidia Blackwell Ultra, Rubin and Rubin Ultra GPUs during 2027–2028, on top of more than 1 million GPUs pledged at GTC 2026. The expanded partnership adds Nvidia Vera CPUs to AWS and integrates NVLink Fusion and custom high-bandwidth memory with AWS Trainium racks, while also covering AI factories, robotics, simulation and training-data infrastructure.

Graphiant CEO: Telcos Should Move Aggressively on AI Infrastructure
Graphiant CEO Ali Shaikh argues AI gives operators an opportunity to move beyond commodity connectivity and become infrastructure providers for the AI economy. Others remain cautious: FiberLight CEO Bill Major warns against telcos stretching into too many businesses, while IDC argues operators should focus on AI opportunities where they have a genuine “right to win” and can execute effectively.

Relativity Networks Raises $22M and Lands $40M Hyperscaler Order
Hollow-core fiber startup Relativity Networks raised $22 million through a SAFE round, more than double its original target, and secured a $40 million follow-on order from an unnamed hyperscaler. The order followed testing of its ChronoCore technology connecting geographically separated data centers with low-latency optical links intended to make distributed GPUs behave more like a single system.

LG Uplus Tests AI-RAN Handoffs for Physical AI Devices
LG Uplus and Yonsei University demonstrated a physical AI device automatically handing off between Yonsei’s Nvidia GPU-based AI-RAN 5G network and LG Uplus' public 5G network while maintaining connectivity. The partners view seamless multi-network mobility for robots and AI-driven equipment as foundational for 6G and plan to expand testing with smaller companies.

Rebellions Targets Telcos With Lower-Cost, Lower-Power AI Inference
Korean AI chip startup Rebellions says it is talking with 10–20 operators globally after nearly three years deploying systems at SK Telecom, pitching hardware at one-half to one-third the price of Nvidia alternatives. Rebellions claims roughly 20 kW per rack versus more than 120 kW for Nvidia’s DGX GB200 NVL72 and about 6x lower operating costs, while using vLLM, PyTorch and OpenShift to maintain a “no forks” open-source software strategy.

T-Mobile Adds Open-Weight Models to Fit-for-Purpose AI Strategy
T-Mobile is combining frontier, open-weight and traditional machine-learning models based on workload complexity, latency, reliability, performance and required control rather than cost alone. The strategy reflects growing emphasis on AI model routing; AT&T, for example, has developed a proprietary Smart Router that evaluates cost, cloud environment and caching across thousands of benchmark decisions.

Cloud

Verizon Expands Google Cloud Partnership Around Gemini and Agentic Network Operations
Verizon is expanding its Google Cloud partnership to deploy Gemini more broadly across customer operations while developing an autonomous network intelligence framework that gives AI agents programmatic access to network systems for anomaly prediction, patching, rerouting and configuration. Google’s Agentic Data Cloud will also consolidate previously siloed customer, network and operational data to support additional AI applications across marketing, security and employee productivity.

RAN

Omdia: $17B First-Half RAN Market Grows Outside China
Omdia estimates the global RAN hardware and software market at roughly $17 billion in the first half of 2026, with growth outside China driven by new 5G deployments, densification, spectrum additions, modernization, vendor swaps and FWA. The top five vendors control about 95% of global revenue, with Huawei leading worldwide and Ericsson leading outside China; Omdia expects single-digit growth outside China for 2026 to be largely offset by Chinese weakness.

Dell’Oro: RAN Revenue Posts Third Consecutive Quarter of Growth
Worldwide RAN revenue grew modestly year-over-year in Q2 2026, marking a third consecutive growth quarter after more than two years of contraction and suggesting the coverage-to-capacity correction is largely over. Huawei, Ericsson, Nokia, ZTE and Samsung accounted for about 96% of the first-half market, while Dell’Oro continues to forecast roughly flat worldwide RAN revenue for full-year 2026.

Private Networks

T-Mobile Targets Private 5G With Hybrid Edge Control Service
T-Mobile’s Edge Control combines its 5G-Advanced infrastructure with network slicing and multi-access edge computing (MEC) to deliver “private network-like” performance without dedicated private infrastructure. The hybrid MNO-led model aims to provide enterprise customers with much of the performance and isolation of private 5G while reducing capital cost and operational complexity.

Celona Orion Converges Private 5G, Wi-Fi 7, Cellular and Satellite
Celona launched Orion, an AI-managed platform combining private 5G, Wi-Fi 7, cellular and satellite connectivity for enterprises deploying physical and agentic AI. Orion’s Orchestrator AI automates provisioning, configuration, diagnosis and remediation, while the open-source AerConnect initiative will provide SDKs for embedding multi-network connectivity into robotics and IoT devices; availability is planned for later in 2026.

Satellite (NTN)

FCC Authorizes AST SpaceMobile 800 MHz Direct-to-Device Tests
The FCC granted AST SpaceMobile a 30-day temporary license to test satellite-to-smartphone connectivity using the 800 MHz band with up to 100 commercially available, unmodified devices, with authorization running through September 12. Partner AT&T called the approval a milestone for direct-to-device (D2D) connectivity as satellite networks increasingly become a complement to terrestrial mobile coverage.

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?