The networking stories enterprise IT leaders actually read.
Volume 343, Issue 3 · 11 articles from 4 issues
As AI agents start making network changes, a mathematical digital twin is needed to predict the impact of a change before it reaches production.
Network engineers still often test changes in production, which becomes a serious liability once AI agents propose and execute changes, and the gap between design intent and reality leads teams to defer patches and firewall updates.
Drawing on a Forward Networks e-book, the analysis separates emulation-based twins from deterministic mathematical models that compute every possible forwarding path, and separates both from observability.
Cited data shows edge-device exploits rising from 3% to 22% of breaches, only 54% of network vulnerabilities remediated each year, and outages costing more than $500,000 per hour.
Generative AI demand pushed Q2 2026 datacenter Ethernet switch revenue up 64.5% as NVIDIA surged and 800Gb/sec sales tripled.
Overall Ethernet switch revenue rose 43.4% year-over-year to $18.91 billion in Q2 2026, with datacenter switches alone reaching $12.3 billion, up 64.5%.
NVIDIA nearly tripled its Ethernet switch sales to about $3.86 billion, outpacing Cisco ($2.24 billion) and Arista ($2.3 billion).
Sales of 800 Gb/sec switches grew 6.1x to $5.07 billion, representing an estimated 41.2% of all datacenter switch ports shipped.
The analysis also flags a growing imbalance: companies are deploying 6.2 times more server cores than high-speed switch ports, raising concerns about networking bottlenecks limiting server utilization.
Satellite direct-to-device services are spreading quickly, but the economics behind non-terrestrial networks lag behind.
Non-terrestrial networks have moved from proprietary islands to early commercial direct-to-device services in dozens of countries, but the economics have not kept up.
CCS Insight's Luke Pearce told RCR's NTN Integration Forum that consumer willingness to pay is real but modest, the add-on model does not work in low-ARPU markets, and constellations for continuous broadband are very capital intensive.
Demand drivers include operators seeking growth beyond 5G and governments mandating coverage, while cheaper launches and 3GPP standardization help supply.
Panelists discussed how open standards, partnerships and regulatory alignment could turn the technology into a commercial-scale business.
UEC 1.1 adds a lightweight transport for GPU scale-up networks and richer, programmable congestion control.
The Ultra Ethernet Consortium released UEC 1.1 on October 12, extending its AI and HPC Ethernet specification to scale-up networks that link accelerators inside a node.
Because scale-up traffic involves small messages at high packet rates, UEC 1.1 adds a simplified transport and options to reduce per-packet overhead, and the consortium is working with OCP on its ESUN specification.
The release also adds programmable congestion control, addressing limits in 1.0 such as single-bit ECN signals that indicate a queue is filling but not where or how badly.
The article explains how these mechanisms work.
Enterprise AI projects are stalling on network bandwidth and latency, as agents generate far more bursty traffic than humans.
This contributed column argues that after heavy spending on models, compute and GPUs, many organizations cannot reliably reach what they built.
It cites research showing that 59% of organizations report bandwidth issues, up from 43% a year earlier, while latency problems rose from 32% to 53%.
Traditional networks were designed for predictable human transactions, but an AI agent can generate up to 450% more network traffic per task than a human.
That traffic is distributed, bursty, concurrent and machine-speed, which is why AI pilots that succeed often struggle when scaled into production.
Analysts see SONiC, the open-source network OS, moving from hyperscalers into enterprise and AI scale-out networks.
Network World looks at why SONiC, the open-source network operating system Microsoft created for white-box switches, is gaining enterprise adoption.
650 Group forecasts SONiC-based data center switching revenue above $5 billion in 2026, Dell'Oro expects SONiC switch sales in AI scale-out networks to grow more than 50% a year through 2030, and Gartner estimates 2,500 organizations run it in production.
A key driver is that Arista, Cisco, Dell and Nvidia now support SONiC on their own hardware.
Gartner calls it a viable option for enterprises seeking flexibility, cost control and less vendor lock-in in leaf-spine and AI fabrics.
Dell'Oro: optics sessions nearly doubled at OCP 2026 as co-packaged optics entered production and 102.4 Tbps switching took shape.
Dell'Oro's analyst says optics dominated OCP 2026, with sessions nearly doubling as AI networks demand more bandwidth and density.
Co-packaged optics hit a milestone as NVIDIA's Spectrum-X Ethernet Photonics switches entered production, and Celestica showed a liquid-cooled 102.4 Tbps CPO switch on Broadcom's Tomahawk 6, part of a 102.4 Tbps generation forming across vendors.
Arista cited near-packaged optics as a path to 6.5 Pb/s per rack, while NVIDIA and Cisco proposed a 400 Gb/s-per-lane USFP pluggable.
The analyst stresses that early commercialization is not broad adoption, and that yields, serviceability and shipment volumes are the metrics to watch.
Lumentum, Astera Labs and Corning demonstrated optical PCIe at OCP, extending reach beyond copper for disaggregated AI racks.
At the 2026 OCP Global Summit, Lumentum, Astera Labs and Corning showed a rack-level proof of concept that carries PCIe over multimode fiber using VCSEL-based optics.
As PCIe data rates rise, copper links face tighter limits on reach, signal integrity, power and cable bulk, with passive copper limited to a few meters while optical links can span tens to hundreds of meters inside a data center.
The companies say optical PCIe could enable more flexible, disaggregated AI rack designs with fewer retimers.
Lumentum stresses that the demo is a proof of concept, not a product.
The Advanced Photonics Coalition approved standards for high-volume optical manufacturing as optical component shortages loom.
The Advanced Photonics Coalition has approved a set of manufacturing, interface and test standards meant to enable high-volume production of optical networking gear, as AI data centers hit copper's reach and power limits.
Lumentum says its indium phosphide optical components are sold out through 2029, and China's role in indium phosphide supply adds geopolitical risk.
The article warns that optical parts could face the kind of shortages and price spikes already hitting DRAM and flash.
APC says common practices will support multiple sourcing and broader supplier participation.
Rule-based automation wins on predictability, while AI-native networking adapts to nonstandard environments; most teams will use both.
TechTarget compares traditional network automation, which relies on admin-written scripts, templates and playbooks and behaves deterministically, with AI-native networking, which infers decisions from real-time and historical data.
Traditional automation is best when changes are known, repeatable and need tight control, but it needs heavy upkeep and struggles with the nonstandard configurations most networks have.
AI-native systems focus on desired state and adapt better, but they shift the admin role to setting guardrails, approving high-risk changes and tuning models.
The article concludes that most teams will combine the two approaches.
Surveys show AI agents in NetOps are mostly advisory; early adopters like Danfoss see gains from limited autonomous actions.
A Cisco and Omdia survey of 1,000 IT and network operations leaders found 51% use agentic AI in production, but mostly for diagnostics and investigations, with only 18% using agents for remediation and 17% for system changes.
An EMA survey found just 13.2% of IT organizations let agents act on anomalies flagged by observability tools.
Danfoss has deployed HPE's self-driving network agents at 20 to 30 sites for tasks such as isolating nodes, provisioning and handling overloads, and a capacity optimization agent has also improved wireless coverage and roaming.
Trust and governance remain the main hurdles to fully autonomous networks.