{"id":5775,"date":"2026-08-30T15:14:58","date_gmt":"2026-08-30T20:14:58","guid":{"rendered":"https:\/\/www.cybersecurityinstitute.com\/blog\/?p=5775"},"modified":"2026-08-30T15:14:58","modified_gmt":"2026-08-30T20:14:58","slug":"agentic-netops-august-30-2026","status":"publish","type":"post","link":"https:\/\/www.cybersecurityinstitute.com\/blog\/?p=5775","title":{"rendered":"Agentic NetOps \u2014 August 30, 2026"},"content":{"rendered":"<style>\n.single .entry-title,\n.single .entry-header .entry-title,\n.single .post-title,\n.single header.entry-header h1,\n.single h1.entry-title,\n.single .page-title,\n.post-template-default h1.entry-title,\n.post-template-default .entry-header,\narticle .entry-header,\narticle .entry-title { display: none !important; }\n.single .entry-header { margin: 0 !important; padding: 0 !important; }\n.single .entry-content { margin-top: 0 !important; padding-top: 0 !important; }\n<\/style>\n<table role=\"presentation\" class=\"wrapper\" cellpadding=\"0\" cellspacing=\"0\" border=\"0\" width=\"100%\">\n<tr>\n<td align=\"center\">\n<table role=\"presentation\" class=\"container\" cellpadding=\"0\" cellspacing=\"0\" border=\"0\" width=\"680\">\n<p>        <!-- Banner --><\/p>\n<tr>\n<td class=\"banner\" style=\"background-color:#312e81;background:linear-gradient(135deg,#312e81 0%,#4338ca 100%);padding:36px 32px;color:#ffffff;\">\n<p class=\"date\" style=\"color:#ffffff !important;\">August 30, 2026 &middot; Weekly Edition<\/p>\n<h1 style=\"color:#ffffff !important;\">Agentic NetOps<\/h1>\n<p class=\"tagline\" style=\"color:#ffffff !important;\">Cisco Cloud Control reaches general availability and the agentic operations pitch stops being a roadmap slide. TM Forum answers with the uncomfortable half of the story &mdash; governance first, or the autonomy programme stalls at the pilot. Meanwhile the fabric people had their own week: Meta published MetaRoCE and moved networking onto its custom silicon, Nvidia took Spectrum-X and NVLink Fusion to Hot Chips 2026, and the OIF set the optical interop agenda for ECOC. Two halves of the same job &mdash; the operations layer that decides, and the fabric layer that has to carry it.<\/p>\n<\/td>\n<\/tr>\n<p>        <!-- At a glance --><\/p>\n<tr>\n<td class=\"content\">\n<h2>This week at a glance<\/h2>\n<p>The week splits cleanly in two, and both halves matter to the same team. On one side, <strong>agentic network operations<\/strong> crossed from announcement into product. <strong>Cisco Cloud Control<\/strong> reached general availability in the US on 25 August with more than 300 customers already onboarded and over 5,000 sign-ups waiting on the global pre-GA list. It puts Meraki, Catalyst Center, Catalyst SD-WAN, Nexus Dashboard, Nexus Hyperfabric, Intersight, Security Cloud Control, ThousandEyes, Splunk and Cisco IQ behind one control plane, adds ServiceNow, Infoblox, BlueCat and Atlassian integrations with more than 60 further ecosystem commitments behind them, and layers four surfaces on top &mdash; AI Canvas, Cloud Control Studio, Cisco Agentic Workflows and the AI Assistant. A cross-product control plane is the precondition for any agent that has to reason about more than one domain at a time, and it is exactly the piece most operators do not have. Cisco paired it with a Secure AI Factory expansion aimed squarely at neoclouds and sovereign clouds: Silicon One switching on the front end, Nvidia Spectrum-X on the back, liquid-cooled N9000 systems and Supermicro racks arriving from October 2026 into cabinets drawing north of 200 kW. On the other side, <strong>AI-fabric engineering<\/strong> produced the densest technical publishing of the week: Meta&#8217;s MetaRoCE transport, Nvidia&#8217;s Spectrum-X and NVLink Fusion material around Hot Chips 2026, and the OIF&#8217;s interoperability agenda for ECOC. The connective tissue is that the operations layer is now writing cheques the fabric layer has to cash.<\/p>\n<p><strong>TM Forum<\/strong> supplied the week&#8217;s most useful corrective, and it came in two parts. Omantel&#8217;s Dr Sukrit Kalia argues for a <strong>governance-first approach to AgenticOps<\/strong> and, unusually, specifies the machinery: an eight-stage control plane running observe, correlate, reason, recommend, approve, execute, validate and learn, with an explicit human policy gate at approval, risk-tiered autonomy that keeps high-impact functions advisory-only while low-risk reversible actions run unattended, and the whole thing mapped onto NIST AI RMF and ISO 42001 rather than invented from scratch. P3 Communications CEO Hakan Ekmen supplies the other half: autonomy confined to network operations is already worth 30&ndash;60% less alarm noise and 20&ndash;40% faster mean time to resolution, but the compounding returns arrive only when the same autonomy crosses into customer care, supply chain and the commercial layer. Most operators are working through TM Forum Autonomy Levels 2 to 3; the measurable value sits between Levels 2 and 4. Read together they describe the same trap from two directions: govern too little and the programme is unsafe, scope too narrowly and it is not worth governing. <strong>Nokia<\/strong> made the domain-specific version of the point for optical networks, framing the move as one from cruise control to agentic operations &mdash; a good description of the difference between an optimiser that holds a set point and an agent that decides the set point should change.<\/p>\n<p>The telco AI commercial layer moved too. <strong>Verizon and Google Cloud<\/strong> expanded their partnership across customer service, network and data, with Gemini already handling the majority of inbound consumer calls and chats in some channels and the network half granting agents programmatic API access for automated patching and configuration changes &mdash; without, as the coverage notes, any published framework for how those agents get constrained, audited or overridden. <strong>AT&amp;T<\/strong> put <strong>OTel 2.0<\/strong> into production: not a telemetry release but the GSMA-backed open telecom model, 31 billion parameters on a Gemma 4 base, post-trained on roughly 400 billion telecom-specific tokens distilled from 3GPP, ETSI, CAMARA, ITU, O-RAN and TM Forum material. AT&amp;T now runs 45 billion tokens a day and more than a trillion a month across 100-plus generative AI models, and reports gateway routing and cached computation reuse cutting inference costs by up to 90% against frontier models. <strong>Lumen<\/strong> extended its Multi-Cloud Gateway past its own fibre to more than 10 million US business locations. And two CEOs made the demand-side case in plain terms: <strong>Calix<\/strong>&#8216;s Michael Weening says AI changes everything for broadband service providers, and <strong>Graphiant<\/strong>&#8216;s Ali Shaikh reports telcos moving aggressively to seize the AI opportunity rather than waiting to be disintermediated.<\/p>\n<p>Underneath all of it, the fabric. <strong>Meta<\/strong> published <strong>MetaRoCE<\/strong>, an RDMA transport for AI-scale Ethernet that abandons in-order delivery outright: packets spray across per-path UDP flows, carry their own destination address, and land straight in memory with no reorder buffer and no head-of-line blocking, while SACK bitvectors and per-path RTT state sort out loss without PFC. It held roughly 86% of throughput at 1% packet loss on a 64-node AMD cluster and scaled linearly across four- and eight-plane topologies with 4,000 concurrent connections. Meta also moved networking on-chip in <strong>MTIA 300<\/strong>, whose two network chiplets carry twelve custom 800 Gb\/s RDMA NICs for 1.2 TB\/s of I\/O that never touches PCIe. <strong>Nvidia<\/strong> had the loudest week: Spectrum-X Ethernet grew 2.6x year on year inside a $96.2 billion quarter, its developer blog laid out three hardware control loops sustaining 98% of line rate at 1.6 Tb\/s per GPU, <strong>Hot Chips 2026<\/strong> got the multiplane architecture that takes a scale-out domain from roughly 8,000 to 512,000 GPUs, and <strong>NVLink Fusion<\/strong> gained NVHBM &mdash; memory controllers moved into the HBM base die, claiming 30% more bandwidth than HBM4e, with AWS, Qualcomm, Marvell, Arm and Fujitsu named alongside. SDxCentral supplied the scepticism, questioning whether the newly promoted &ldquo;scale-in&rdquo; pillar is a real architectural category or a marketing one. On the physical layer, the <strong>OIF<\/strong> takes 448G, 800ZR, CMIS and AI interoperability to <strong>ECOC 2026<\/strong> in M&aacute;laga on 21&ndash;23 September, which is where the roadmap claims get tested against other people&#8217;s hardware.<\/p>\n<p>            <!-- Topic map --><\/p>\n<div class=\"topic-map\">\n              <img decoding=\"async\" src=\"https:\/\/www.cybersecurityinstitute.com\/blog\/wp-content\/uploads\/2026\/08\/topic-map-agentic-netops-2026-08-30.png\" alt=\"Topic map of this week's Agentic NetOps themes: agentic AI and autonomous networks anchored by Cisco Cloud Control and TM Forum's governance-first AgenticOps line, the telco AI commercial layer with Verizon and Google Cloud, AT&amp;T putting the GSMA-backed OTel 2.0 open telecom model into production, Lumen, Calix and Graphiant, optical autonomous operations at Nokia and VIAVI with the OIF and ECOC 2026 interop agenda, and the AI-fabric engineering thread running from Meta's MetaRoCE RDMA transport through Nvidia Spectrum-X, NVLink Fusion, BlueField and Hot Chips 2026\" loading=\"eager\"><\/p>\n<p class=\"caption\">This week&#8217;s topic map &mdash; the agentic operations half (Cisco Cloud Control, TM Forum AgenticOps governance, autonomous networks, Canonical, VIAVI, Versa and MCP, network APIs); the telco AI commercial layer (Verizon and Google Cloud, AT&amp;T with the OTel 2.0 open telecom model, Lumen&#8217;s Multi-Cloud Gateway, Calix, Graphiant); the optical autonomy and interop layer (Nokia, OIF, ECOC 2026); and the AI-fabric engineering thread (Meta and MetaRoCE, RDMA over Ethernet, Nvidia Spectrum-X, NVLink Fusion, BlueField, Hot Chips 2026). Node size reflects how often an entity is mentioned; edge weight reflects how often two entities are discussed together.<\/p>\n<p>              <!-- INTERACTIVE_MAP_LINK_START --><\/p>\n<p style=\"margin:10px 0 0;text-align:center;\"><a href=\"https:\/\/www.cybersecurityinstitute.com\/blog\/?p=5774\" target=\"_blank\" rel=\"noopener\" style=\"display:inline-block;padding:8px 18px;background-color:#1e1b4b;color:#ffffff !important;text-decoration:none;border-radius:6px;font-size:13px;font-weight:600;\">View interactive topic map &rarr;<\/a><\/p>\n<p><!-- INTERACTIVE_MAP_LINK_END -->\n            <\/div>\n<p>            <!-- Article index --><\/p>\n<h2>Article index<\/h2>\n<h3>Weekly News<\/h3>\n<h4>Cisco Cloud Control goes GA: agentic operations gets a product<\/h4>\n<div class=\"cluster-intro\">Cisco&#8217;s cross-domain control plane reaches general availability in the US, and the accompanying AI-factory infrastructure expansion shows where the company thinks the money is &mdash; neoclouds and sovereign clouds rather than the traditional enterprise core.<\/div>\n<table class=\"index-table\">\n<tr>\n<th>Article<\/th>\n<th>Source<\/th>\n<th>Published<\/th>\n<\/tr>\n<tr>\n<td>1. <a href=\"https:\/\/blogs.cisco.com\/ai\/cisco-cloud-control-us-general-availability\">A new chapter for IT operations: Cisco Cloud Control is now generally available<\/a><\/td>\n<td class=\"src\">Cisco Blogs<\/td>\n<td class=\"dt\">Aug 25, 2026<\/td>\n<\/tr>\n<tr>\n<td>2. <a href=\"https:\/\/www.sdxcentral.com\/news\/cisco-agentifies-suite-with-single-cloud-pane-of-glass\/\">Cisco agentifies suite with single cloud pane of glass<\/a><\/td>\n<td class=\"src\">SDxCentral<\/td>\n<td class=\"dt\">Aug 26, 2026<\/td>\n<\/tr>\n<tr>\n<td>3. <a href=\"https:\/\/siliconangle.com\/2026\/08\/25\/cisco-expands-rack-scale-secure-ai-factory-infrastructure-for-neocloud-and-sovereign-clouds\/\">Cisco expands rack-scale secure AI factory infrastructure for neocloud and sovereign clouds<\/a><\/td>\n<td class=\"src\">SiliconANGLE<\/td>\n<td class=\"dt\">Aug 25, 2026<\/td>\n<\/tr>\n<\/table>\n<h4>Governance first: TM Forum on AgenticOps and enterprise-wide autonomy<\/h4>\n<div class=\"cluster-intro\">Two TM Forum pieces that belong together &mdash; oversight has to precede deployment, and autonomy confined to the network domain leaves most of its value on the table &mdash; plus Nokia&#8217;s domain-specific version of the same argument for optical networks.<\/div>\n<table class=\"index-table\">\n<tr>\n<th>Article<\/th>\n<th>Source<\/th>\n<th>Published<\/th>\n<\/tr>\n<tr>\n<td>4. <a href=\"https:\/\/inform.tmforum.org\/features-and-opinion\/telecoms-operators-need-a-governance-first-approach-to-agenticops\">Telecoms operators need a governance-first approach to AgenticOps<\/a><\/td>\n<td class=\"src\">TM Forum Inform<\/td>\n<td class=\"dt\">Aug 27, 2026<\/td>\n<\/tr>\n<tr>\n<td>5. <a href=\"https:\/\/inform.tmforum.org\/features-and-opinion\/autonomy-across-the-enterprise-is-where-real-value-lies\">Autonomy across the enterprise is where real value lies<\/a><\/td>\n<td class=\"src\">TM Forum Inform<\/td>\n<td class=\"dt\">Aug 27, 2026<\/td>\n<\/tr>\n<tr>\n<td>6. <a href=\"https:\/\/www.nokia.com\/blog\/how-do-optical-networks-move-beyond-cruise-control-to-agentic-operations\/\">How do optical networks move beyond cruise control to agentic operations?<\/a><\/td>\n<td class=\"src\">Nokia Blog<\/td>\n<td class=\"dt\">Aug 26, 2026<\/td>\n<\/tr>\n<\/table>\n<h4>Operator AI programmes: the models, partnerships and reach underneath<\/h4>\n<div class=\"cluster-intro\">Verizon deepens its Google Cloud partnership across service, network and data; AT&amp;T puts OTel 2.0 &mdash; the GSMA-backed open telecom model, not a telemetry release &mdash; into production on its own hardware; Lumen pushes its Multi-Cloud Gateway past the core network.<\/div>\n<table class=\"index-table\">\n<tr>\n<th>Article<\/th>\n<th>Source<\/th>\n<th>Published<\/th>\n<\/tr>\n<tr>\n<td>7. <a href=\"https:\/\/rcrwireless.com\/20260826\/ai\/verizon-google-cloud-ai-partnership\">Verizon and Google Cloud expand AI partnership across service, network, and data<\/a><\/td>\n<td class=\"src\">RCR Wireless News<\/td>\n<td class=\"dt\">Aug 26, 2026<\/td>\n<\/tr>\n<tr>\n<td>8. <a href=\"https:\/\/rcrwireless.com\/20260828\/ai\/att-otel-2-0-production\">AT&amp;T takes OTel 2.0 into production<\/a><\/td>\n<td class=\"src\">RCR Wireless News<\/td>\n<td class=\"dt\">Aug 28, 2026<\/td>\n<\/tr>\n<tr>\n<td>9. <a href=\"https:\/\/www.sdxcentral.com\/news\/lumen-extends-multi-cloud-gateway-outside-of-core-network\/\">Lumen extends Multi-Cloud Gateway outside of core network<\/a><\/td>\n<td class=\"src\">SDxCentral<\/td>\n<td class=\"dt\">Aug 26, 2026<\/td>\n<\/tr>\n<\/table>\n<h4>MetaRoCE: Meta rewrites the transport and moves networking on-chip<\/h4>\n<div class=\"cluster-intro\">Meta&#8217;s engineering blog publishes a new RDMA transport designed for AI-scale Ethernet, the trade press unpacks why it deliberately tolerates out-of-order delivery, and a third piece covers Meta baking networking into custom silicon to cut CPU hops.<\/div>\n<table class=\"index-table\">\n<tr>\n<th>Article<\/th>\n<th>Source<\/th>\n<th>Published<\/th>\n<\/tr>\n<tr>\n<td>10. <a href=\"https:\/\/engineering.fb.com\/2026\/08\/24\/networking-traffic\/metaroce-rdma-transport-ai-ethernet\/\">MetaRoCE: A New RDMA Transport Built for AI-Scale Ethernet<\/a><\/td>\n<td class=\"src\">Meta Engineering<\/td>\n<td class=\"dt\">Aug 24, 2026<\/td>\n<\/tr>\n<tr>\n<td>11. <a href=\"https:\/\/www.sdxcentral.com\/news\/metas-custom-transport-protocol-embraces-packet-chaos-to-boost-ai-throughput\/\">Meta&#8217;s custom transport protocol embraces packet chaos to boost AI throughput<\/a><\/td>\n<td class=\"src\">SDxCentral<\/td>\n<td class=\"dt\">Aug 25, 2026<\/td>\n<\/tr>\n<tr>\n<td>12. <a href=\"https:\/\/www.sdxcentral.com\/news\/meta-bakes-networking-on-chip-in-custom-silicon-designed-to-cut-cpu-hops\/\">Meta bakes networking on-chip in custom silicon designed to cut CPU hops<\/a><\/td>\n<td class=\"src\">SDxCentral<\/td>\n<td class=\"dt\">Aug 28, 2026<\/td>\n<\/tr>\n<\/table>\n<h4>Nvidia&#8217;s fabric week: Spectrum-X, NVLink Fusion and Hot Chips 2026<\/h4>\n<div class=\"cluster-intro\">Sales claims, an architecture deep-dive, a Hot Chips multiplane session and a memory-reach expansion for custom silicon &mdash; with one piece asking whether the newly promoted &ldquo;scale-in&rdquo; pillar is an architecture or a slide.<\/div>\n<table class=\"index-table\">\n<tr>\n<th>Article<\/th>\n<th>Source<\/th>\n<th>Published<\/th>\n<\/tr>\n<tr>\n<td>13. <a href=\"https:\/\/developer.nvidia.com\/blog\/giga-scale-ai-ethernet-evolution-spectrum-x-ethernet-rewrites-rules\/\">Giga-Scale AI and the Ethernet Evolution: How Spectrum-X Ethernet Rewrites the Rules<\/a><\/td>\n<td class=\"src\">NVIDIA Developer Blog<\/td>\n<td class=\"dt\">Aug 24, 2026<\/td>\n<\/tr>\n<tr>\n<td>14. <a href=\"https:\/\/www.servethehome.com\/nvidia-spectrum-x-ethernet-multiplane-network-architecture-at-hot-chips-2026\/\">NVIDIA Spectrum-X Ethernet Multiplane Network Architecture at Hot Chips 2026<\/a><\/td>\n<td class=\"src\">ServeTheHome<\/td>\n<td class=\"dt\">Aug 25, 2026<\/td>\n<\/tr>\n<tr>\n<td>15. <a href=\"https:\/\/www.sdxcentral.com\/news\/scale-in-nvidias-fifth-pillar-of-ai-networking-rests-on-shaky-ground\/\">Scale-in? Nvidia&#8217;s &#8216;fifth&#8217; pillar of AI networking rests on shaky ground<\/a><\/td>\n<td class=\"src\">SDxCentral<\/td>\n<td class=\"dt\">Aug 25, 2026<\/td>\n<\/tr>\n<tr>\n<td>16. <a href=\"https:\/\/www.sdxcentral.com\/news\/nvidia-claims-fastest-growing-networking-title-as-spectrum-x-sales-soar\/\">Nvidia claims &#8216;fastest-growing&#8217; networking title as Spectrum-X sales soar<\/a><\/td>\n<td class=\"src\">SDxCentral<\/td>\n<td class=\"dt\">Aug 27, 2026<\/td>\n<\/tr>\n<tr>\n<td>17. <a href=\"https:\/\/www.sdxcentral.com\/news\/nvidia-expands-nvlink-fusion-strategy-to-supercharge-custom-silicon-memory\/\">Nvidia expands NVLink Fusion strategy to supercharge custom silicon memory<\/a><\/td>\n<td class=\"src\">SDxCentral<\/td>\n<td class=\"dt\">Aug 27, 2026<\/td>\n<\/tr>\n<\/table>\n<h4>Access, edge and the optics interop calendar<\/h4>\n<div class=\"cluster-intro\">Two CEOs on the demand side of telco AI &mdash; Calix on what AI does to broadband service providers, Graphiant on operators moving aggressively rather than defensively &mdash; and the OIF setting the interoperability agenda for ECOC 2026.<\/div>\n<table class=\"index-table\">\n<tr>\n<th>Article<\/th>\n<th>Source<\/th>\n<th>Published<\/th>\n<\/tr>\n<tr>\n<td>18. <a href=\"https:\/\/www.lightreading.com\/ai-machine-learning\/calix-ceo-is-a-firm-believer-that-ai-changes-everything-\">Calix CEO is a &#8216;firm believer that AI changes everything&#8217;<\/a><\/td>\n<td class=\"src\">Light Reading<\/td>\n<td class=\"dt\">Aug 27, 2026<\/td>\n<\/tr>\n<tr>\n<td>19. <a href=\"https:\/\/www.fierce-network.com\/cloud\/telcos-move-aggressively-seize-ai-opportunity-graphiant-ceo\">Telcos move &#8216;aggressively&#8217; to seize AI opportunity &mdash; Graphiant CEO<\/a><\/td>\n<td class=\"src\">Fierce Network<\/td>\n<td class=\"dt\">Aug 27, 2026<\/td>\n<\/tr>\n<tr>\n<td>20. <a href=\"https:\/\/convergedigest.com\/oif-448g-800zr-cmis-ai-interoperability-ecoc-2026\/\">OIF Brings 448G, 800ZR, CMIS and AI Interoperability to ECOC 2026<\/a><\/td>\n<td class=\"src\">Converge Digest<\/td>\n<td class=\"dt\">Aug 26, 2026<\/td>\n<\/tr>\n<\/table>\n<h3>Foundational Reading<\/h3>\n<h4>Harnesses, APIs and the agent interface layer<\/h4>\n<div class=\"cluster-intro\">Three pieces on the plumbing between an agent and a network: Canonical on harnessing AI for telco autonomous networks, TechTarget on why network APIs are the precondition for agentic AI, and Versa on MCP as the emerging operating model for enterprise infrastructure.<\/div>\n<table class=\"index-table\">\n<tr>\n<th>Article<\/th>\n<th>Source<\/th>\n<th>Published<\/th>\n<\/tr>\n<tr>\n<td>21. <a href=\"https:\/\/ubuntu.com\/blog\/ai-harnesses-for-telco-autonomous-networks\">AI harnesses for telco autonomous networks<\/a><\/td>\n<td class=\"src\">Ubuntu (Canonical)<\/td>\n<td class=\"dt\">Aug 28, 2026<\/td>\n<\/tr>\n<tr>\n<td>22. <a href=\"https:\/\/www.techtarget.com\/it-infrastructure\/tip\/Why-agentic-AI-needs-network-APIs-to-succeed\">Why agentic AI needs network APIs to succeed<\/a><\/td>\n<td class=\"src\">TechTarget<\/td>\n<td class=\"dt\">Aug 25, 2026<\/td>\n<\/tr>\n<tr>\n<td>23. <a href=\"https:\/\/versa-networks.com\/blog\/mcp-enterprise-infrastructure-zero-trust-governance\/\">Beyond the Console: MCP and the New Operating Model for Enterprise Infrastructure<\/a><\/td>\n<td class=\"src\">Versa Networks<\/td>\n<td class=\"dt\">Aug 27, 2026<\/td>\n<\/tr>\n<\/table>\n<h4>Self-aware RAN, and the fabric plumbing that portable workloads need<\/h4>\n<div class=\"cluster-intro\">VIAVI on what it takes to make a RAN self-aware rather than merely instrumented, and a practical piece on why VM mobility across Kubernetes clusters is a layer-2 problem before it is a platform one.<\/div>\n<table class=\"index-table\">\n<tr>\n<th>Article<\/th>\n<th>Source<\/th>\n<th>Published<\/th>\n<\/tr>\n<tr>\n<td>24. <a href=\"https:\/\/blog.viavisolutions.com\/2026\/08\/24\/the-journey-to-a-self-aware-autonomous-ran\/\">The Journey to a Self-Aware Autonomous RAN<\/a><\/td>\n<td class=\"src\">VIAVI Perspectives<\/td>\n<td class=\"dt\">Aug 24, 2026<\/td>\n<\/tr>\n<tr>\n<td>25. <a href=\"https:\/\/thenewstack.io\/kubevirt-evpn-vm-migration\/\">Why your KubeVirt VMs can&#8217;t move between clusters &mdash; and how EVPN fixes it<\/a><\/td>\n<td class=\"src\">The New Stack<\/td>\n<td class=\"dt\">Aug 8, 2026<\/td>\n<\/tr>\n<\/table>\n<p>            <!-- Detailed write-ups --><\/p>\n<h2>Detailed write-ups<\/h2>\n<div class=\"article\">\n<h4>1. Cisco Cloud Control reaches GA &mdash; the cross-domain control plane agentic operations has been waiting for<\/h4>\n<p class=\"meta\">Cisco Blogs &middot; SDxCentral &middot; SiliconANGLE &middot; August 25&ndash;26, 2026<\/p>\n<p>Cisco Cloud Control went generally available in the US on 25 August 2026, and it is the most consequential product event in this issue because of what it makes possible rather than what it does on day one. The adoption numbers Cisco published are modest but real: more than 300 customers onboarded, and over 5,000 sign-ups on a global pre-GA waitlist that the US-only phased rollout has not yet reached. What matters more is the surface area. Cloud Control spans Meraki, Catalyst Center, Catalyst SD-WAN, Nexus Dashboard, Nexus Hyperfabric, Intersight, Identity, Security Cloud Control, ThousandEyes, Splunk, Collaboration Control Hub and Cisco IQ, with ServiceNow, Infoblox, BlueCat and Atlassian integrations already in and more than 60 further ecosystem commitments behind them. Four surfaces sit on top: AI Canvas, a multiplayer workspace where people and agents work the same incident; Cloud Control Studio for building custom apps and agents; Cisco Agentic Workflows; and the AI Assistant throughout. Agentic operations is not blocked on model quality; it is blocked on the fact that the answer to almost any real production question lives in three or four separate systems that do not share a data model. An agent asked why a site is degraded needs routing state, security policy, wireless telemetry and application performance in the same reasoning context &mdash; and it is the ThousandEyes integration that puts latency, packet loss, jitter and availability into that context rather than in a different console. Cisco&#8217;s own case study cites an issue that took more than ten hours resolved in under ten minutes; treat the figure as a best case, but the mechanism behind it is the right one.<\/p>\n<p>Balaji Venkatraman, Cisco&#8217;s VP of product management for AI software and platform, framed the release more carefully than the category usually manages: &ldquo;We believe agenticOps will change how teams operate IT infrastructure, but that change will be earned one result at a time.&rdquo; It is still worth reading as a competitive move as much as a product one. Making the control plane the product and the agents a feature of it is a defensible position for an incumbent with a broad portfolio and an awkward one for point-tool vendors. It also raises the question every buyer should ask before signing: how much of the value depends on the estate being Cisco end to end. The twelve connected domains Cisco lists are all its own; the named third-party integrations number four. A control plane that reasons beautifully across Cisco networking, Cisco security and Cisco observability, and thinly across everything else, is a consolidation argument dressed as an operations argument. That may still be the right trade for many organisations, but it is a procurement decision with a long tail, and it should be made deliberately rather than discovered eighteen months in. Cisco used the same week to launch Sovereign Critical Infrastructure in Canada, aligned to the ITSG-33 control framework &mdash; the same platform, repackaged for buyers whose first question is jurisdiction rather than capability.<\/p>\n<p>Alongside the operations release, Cisco expanded its rack-scale Secure AI Factory with an explicit target: neoclouds and sovereign clouds. The build is specific. Cisco Silicon One switches handle the front end, Nvidia Spectrum-X switches the back-end fabric, and the liquid-cooled N9000 Series carries rack-to-fabric traffic, all managed through Nexus One. Compute is Nvidia Vera Rubin NVL72 and HGX Rubin NVL8, with Supermicro rack-scale liquid-cooled systems rolling out from October 2026 into cabinets drawing more than 200 kW apiece. Jeetu Patel, Cisco&#8217;s president and chief product officer, put the pitch as &ldquo;every organization is racing to scale AI &mdash; but speed only counts if it comes with control of data, managed token costs and real return on investment,&rdquo; with neocloud operator Sharon AI supplying the reference customer. Two details deserve an architect&#8217;s attention. Cisco is reselling Nvidia&#8217;s back-end fabric rather than competing with it, and reserving its own silicon for the front end &mdash; a division of labour that says a good deal about where each vendor&#8217;s leverage currently sits. And a 200 kW rack is a facilities decision before it is a networking one: if your data centre cannot deliver liquid cooling at that density, the fabric conversation is premature. Taken together, the two announcements describe a coherent bet: sell the fabric to the people building AI capacity, and sell the agentic control plane to the people who then have to operate it.<\/p>\n<p style=\"font-size:13px;color:#6b7280;margin:0;\">Sources: <a href=\"https:\/\/blogs.cisco.com\/ai\/cisco-cloud-control-us-general-availability\">Cisco Blogs (Cloud Control general availability)<\/a> &middot; <a href=\"https:\/\/www.sdxcentral.com\/news\/cisco-agentifies-suite-with-single-cloud-pane-of-glass\/\">SDxCentral (Cisco agentifies its suite)<\/a> &middot; <a href=\"https:\/\/siliconangle.com\/2026\/08\/25\/cisco-expands-rack-scale-secure-ai-factory-infrastructure-for-neocloud-and-sovereign-clouds\/\">SiliconANGLE (rack-scale secure AI factory infrastructure)<\/a><\/p>\n<\/p><\/div>\n<div class=\"article\">\n<h4>2. TM Forum&#8217;s two-part argument: govern before you deploy, and stop confining autonomy to the network<\/h4>\n<p class=\"meta\">TM Forum Inform &middot; August 27, 2026<\/p>\n<p>TM Forum published two pieces on the same day that read as one argument with two failure modes. The first, by Dr Sukrit Kalia, an AI and machine-learning subject-matter expert at Omantel, says operators need a <strong>governance-first approach to AgenticOps<\/strong>, and its premise is blunt: autonomous agents are already capable of operating telecom infrastructure faster than human teams can supervise them in real time. The usual sequencing &mdash; stand up agents, prove value, retrofit oversight &mdash; is exactly backwards, because the oversight machinery is what determines the blast radius of the first mistake, and the first mistake always arrives before the governance project is funded. What lifts the piece above the usual advice is that it specifies the machinery. It proposes an AgenticOps control plane of eight stages &mdash; observe, correlate, reason, recommend, approve, execute, validate, learn &mdash; in which approval is an explicit human policy gate rather than an implied one, sitting beneath an agent orchestration layer. Autonomy is tiered by risk: advisory-only for high-impact functions, fully autonomous only for the low-risk and reversible. Audit logging captures the decision rationale, not merely the resulting change. Data residency is treated as an architectural input rather than a compliance afterthought. And the whole structure is mapped onto NIST AI RMF and ISO 42001 instead of being invented locally. None of the individual controls are novel. What is novel is applying them to a non-deterministic actor working at machine speed against production infrastructure, when the existing RBAC and audit stack was built on the assumption that whoever made the change had a login and a manager.<\/p>\n<p>The second piece, from P3 Communications CEO Hakan Ekmen, pushes in the opposite direction and is the more commercially interesting of the two. Autonomy confined to network operations is genuinely worth something &mdash; he cites 30&ndash;60% reductions in alarm noise, 20&ndash;40% faster mean time to resolution and materially fewer duplicate tickets &mdash; but that is the floor, not the ceiling. The returns compound when the same autonomy runs <strong>across the enterprise<\/strong>: customer care, supply chain, order flow and the commercial layer, coordinated end to end rather than optimised in isolation, which is how automotive, manufacturing and logistics already run. TM Forum&#8217;s Autonomy Levels give the argument a yardstick, and the yardstick is unflattering: most operators are working through Levels 2 to 3, while the measurable value sits between Levels 2 and 4. The structural implication is about org design as much as technology. A network-scoped agent optimises inside constraints set elsewhere. An enterprise-scoped one can trade across them &mdash; deferring a truck roll because the order it would serve is being restructured, or reprioritising remediation because the affected service carries a different commercial commitment than the ticket queue suggests. That is where the money is, and it is also where the governance requirements get materially harder, because the actions now cross organisational boundaries that have their own approval cultures.<\/p>\n<p>Read together the two pieces describe a narrow path. Under-govern and the programme is unsafe and will be halted by the first incident that reaches a regulator or a major customer. Over-narrow the scope and the programme is safe, dull and impossible to justify at the next budget round &mdash; a 30% cut in alarm noise is a good quarter, not a business case. The practical read for a NetOps leader is that the governance work is not a tax on the autonomy programme; it is the thing that lets the autonomy programme expand past the network domain without the rest of the business blocking it. Build the identity, permissioning and audit substrate now, at network scope where the blast radius is understood and while you are still climbing from Level 2 to Level 3, and it becomes the asset that makes the enterprise-wide expansion approvable later.<\/p>\n<p style=\"font-size:13px;color:#6b7280;margin:0;\">Sources: <a href=\"https:\/\/inform.tmforum.org\/features-and-opinion\/telecoms-operators-need-a-governance-first-approach-to-agenticops\">TM Forum Inform (governance-first AgenticOps)<\/a> &middot; <a href=\"https:\/\/inform.tmforum.org\/features-and-opinion\/autonomy-across-the-enterprise-is-where-real-value-lies\">TM Forum Inform (autonomy across the enterprise)<\/a><\/p>\n<\/p><\/div>\n<div class=\"article\">\n<h4>3. MetaRoCE: a transport that stops fighting packet reordering &mdash; and networking moves onto Meta&#8217;s silicon<\/h4>\n<p class=\"meta\">Meta Engineering &middot; SDxCentral &middot; August 24&ndash;28, 2026<\/p>\n<p>Meta&#8217;s engineering blog published <strong>MetaRoCE<\/strong>, a new RDMA transport built specifically for AI-scale Ethernet, and it is the most substantive piece of network engineering in this issue. The problem it addresses is well known to anyone who has run RoCE at scale: classical RDMA assumes in-order delivery, which forces the fabric to preserve ordering with PFC and pause frames, which in practice means constraining how traffic is spread across available paths. MetaRoCE discards the assumption. It treats Ethernet as inherently lossy and needs no PFC at all, decomposing the fabric into fine-grained logical paths that the NIC tracks individually &mdash; per-path RTT, ECN state and utilisation, each path pinned to its own UDP source port, so path selection becomes something the endpoint controls rather than something a switch hash inflicts on it. Meta&#8217;s own summary of the design is the line worth keeping: the fabric sees packets, but the NIC sees intent. Because every packet carries its destination address, out-of-order arrivals write straight into memory with no reorder buffer and no head-of-line blocking, and SACK bitvectors identify the gaps so a retransmission goes back down the specific path that dropped it. Per-path windows and RTT estimates let the transport tell congestion apart from failure. Congestion control pairs sender-driven ECN-based AIMD with receiver-driven fair-share rate hints returned in acknowledgements, so a sender converges on its allocation rather than probing towards it.<\/p>\n<p>The measurements are what make it more than an architecture argument. On a 64-node AMD GPU cluster using Pensando programmable NICs, MetaRoCE held roughly 86% of its throughput at 1% packet loss &mdash; a rate at which RoCEv2 falls apart &mdash; and continued to deliver useful bandwidth at loss rates as high as 10%. It scaled linearly across four- and eight-plane topologies with 4,000 concurrent connections, and recovered autonomously through simulated plane failures. AMD&#8217;s Omar Baldonado, senior director for data centre and AI networking, put the principle plainly: design for loss from day one, push the intelligence to the edge, and you get a transport that is faster in good conditions and degrades gracefully in bad ones. Meta is not keeping it in-house &mdash; the specification, a compliance suite and a DPDK-optimised reference implementation are due at the OCP Global Summit in October 2026, and libsoftmetaroce already provides a complete functional stack over plain UDP on commodity Linux, which means the behaviour can be evaluated before anyone sells you silicon. For engineers running smaller estates the operational lesson generalises: if your fabric&#8217;s performance depends on flows landing evenly through a hash function you do not control, you have a fragility that will surface as unexplained tail latency long before it surfaces as a link utilisation alarm. The direction of travel is toward transports that assume the network will be uneven and compensate at the edge.<\/p>\n<p>The third Meta item completes the picture from the other end. <strong>MTIA 300<\/strong>, the latest Meta Training and Inference Accelerator, brings the NICs inside the package: two network chiplets, each carrying six custom 800 Gb\/s RDMA NICs, giving twelve Ethernet interfaces and 1.2 TB\/s of I\/O bandwidth that never crosses a PCIe bus, alongside 216 GB of HBM3E. Collectives execute autonomously through Meta&#8217;s HCCL library with the host uninvolved, and on a 150-billion-parameter recommendation model spread across 40 accelerators Meta reports communication running 3.9 times faster than on an equivalent GPU cluster. This is the same offload argument that produced SmartNICs and DPUs, applied by an operator with enough scale to justify its own ASIC. Every hop through a general-purpose CPU on the data path costs latency, costs cycles that could be running the workload, and adds a jitter source that is hard to characterise. Doing the transport work in silicon that sits next to the accelerator is how you keep the transport innovation above from being eaten by host overhead. Between the two pieces, Meta has published both halves of a coherent design: a transport that tolerates an uneven network, and hardware that executes it without taxing the compute it exists to serve.<\/p>\n<p style=\"font-size:13px;color:#6b7280;margin:0;\">Sources: <a href=\"https:\/\/engineering.fb.com\/2026\/08\/24\/networking-traffic\/metaroce-rdma-transport-ai-ethernet\/\">Meta Engineering (MetaRoCE RDMA transport)<\/a> &middot; <a href=\"https:\/\/www.sdxcentral.com\/news\/metas-custom-transport-protocol-embraces-packet-chaos-to-boost-ai-throughput\/\">SDxCentral (embracing packet chaos)<\/a> &middot; <a href=\"https:\/\/www.sdxcentral.com\/news\/meta-bakes-networking-on-chip-in-custom-silicon-designed-to-cut-cpu-hops\/\">SDxCentral (networking on-chip in custom silicon)<\/a><\/p>\n<\/p><\/div>\n<div class=\"article\">\n<h4>4. Nvidia&#8217;s fabric week: Spectrum-X momentum, a Hot Chips architecture reveal, NVLink Fusion for custom memory &mdash; and one sceptic<\/h4>\n<p class=\"meta\">NVIDIA Developer Blog &middot; ServeTheHome &middot; SDxCentral &middot; August 24&ndash;27, 2026<\/p>\n<p>Nvidia spent the week arguing that it is a networking company, and this time the financials carry the claim. In the quarter ended 26 July 2026 it reported $96.2 billion of revenue, up 18% sequentially and 106% year on year at 75% gross margin, with Spectrum-X Ethernet growing 2.6 times year on year &mdash; enough for CFO Colette Kress to describe Nvidia as the largest and fastest-growing networking company in the world. The underlying shift is real: when the fabric determines cluster efficiency rather than merely connecting it, the accelerator vendor has both the incentive to sell the network with the compute and the advantage of co-designing the two. Its developer blog put engineering behind the claim, describing three hardware-accelerated control loops that Nvidia argues become unavoidable above 800 Gb\/s &mdash; in-switch adaptive routing steering per packet with a quantised join-shortest-queue algorithm, targeted congestion control combining ECN marking with RTT-based rate adjustment, and a NIC-side plane load balancer distributing traffic across independent planes. The claimed outcome is 1.6 Tb\/s of scale-out bandwidth per GPU sustained at 98% of theoretical line rate, in a two-tier topology reaching beyond 128,000 endpoints and a three-tier one reaching 16 million. Note what those loops have in common with MetaRoCE: both conclude that software-timescale reaction is too slow, and both move the decision into hardware close to the traffic.<\/p>\n<p>ServeTheHome&#8217;s write-up of the <strong>Hot Chips 2026<\/strong> session filled in the topology, and it is the structural answer to the problem MetaRoCE attacks from the transport side. Rather than giving each GPU a single 1.6 Tb\/s uplink, the multiplane design splits it into eight 200 Gb\/s links, each landing on a different switch &mdash; eight planes across four rails, built from 512-port 200G switches with 100 Tb\/s of capacity. Nvidia&#8217;s claim is that this takes a scale-out domain from around 8,000 GPUs on conventional multi-rail to 512,000 on multiplane, a 64-fold increase, while using 1.7 times fewer scale-out switches and delivering 1.9 times higher training performance under multi-tenant conditions. The optics half is equally specific: a co-packaged optics engine using micro-ring modulators, 3D-stacked on TSMC&#8217;s COUPE process, claiming four times fewer lasers, lower power and a ten times longer mean time between interruptions. That last figure is the one to press vendors on, because at half a million endpoints optical link flap stops being a maintenance nuisance and becomes a job-failure rate. The same session put NVLink Fusion at 3.6 TB\/s of all-to-all bandwidth across 72 XPUs, which is the scale-up number against which any competing rack-level interconnect will be measured.<\/p>\n<p><strong>NVLink Fusion<\/strong> is the other half of the strategy and the more strategically aggressive one. This week&#8217;s expansion centres on NVHBM, which moves the memory controller off the accelerator board and into the HBM base die: Nvidia claims up to 30% more memory bandwidth than HBM4e, 15% lower power, and up to 30% more silicon freed on the compute die for whoever is designing the XPU. AWS&#8217;s Annapurna Labs, Qualcomm, Marvell, Arm, Fujitsu and Ayar Labs are named around it, with Annapurna VP Nafea Bshara calling NVHBM a new architectural approach to advancing high-bandwidth memory performance. Read the platform logic rather than the spec sheet. It attacks the memory wall directly &mdash; the practical constraint on large-context inference is not usually raw compute but how much fast memory a job can reach and how quickly &mdash; and it makes Nvidia&#8217;s interconnect the substrate for other people&#8217;s chips. If your custom accelerator reaches pooled memory through NVLink Fusion, Nvidia keeps a structural position in your rack even where it did not sell you the compute. Against all of this, SDxCentral supplied the corrective by questioning the newly promoted &ldquo;scale-in&rdquo; pillar. Using BlueField-4 DPUs, the DOCA stack and Spectrum-X to accelerate north-south traffic and offload security and storage from the main processors is a sensible product; calling it a fifth architectural category alongside scale-up, scale-out and scale-across looks more like taxonomy in service of a catalogue, particularly while the fourth pillar itself remains sparsely defined. Taxonomies matter here because they shape RFPs, and a vendor-coined category that maps onto one vendor&#8217;s product line quietly writes a sole-source requirement into a specification. When a new pillar appears, the test is whether it names a problem you had before the slide existed.<\/p>\n<p style=\"font-size:13px;color:#6b7280;margin:0;\">Sources: <a href=\"https:\/\/developer.nvidia.com\/blog\/giga-scale-ai-ethernet-evolution-spectrum-x-ethernet-rewrites-rules\/\">NVIDIA Developer Blog (giga-scale AI and Ethernet evolution)<\/a> &middot; <a href=\"https:\/\/www.servethehome.com\/nvidia-spectrum-x-ethernet-multiplane-network-architecture-at-hot-chips-2026\/\">ServeTheHome (Spectrum-X multiplane architecture at Hot Chips 2026)<\/a> &middot; <a href=\"https:\/\/www.sdxcentral.com\/news\/nvidia-claims-fastest-growing-networking-title-as-spectrum-x-sales-soar\/\">SDxCentral (fastest-growing networking claim)<\/a> &middot; <a href=\"https:\/\/www.sdxcentral.com\/news\/nvidia-expands-nvlink-fusion-strategy-to-supercharge-custom-silicon-memory\/\">SDxCentral (NVLink Fusion and custom silicon memory)<\/a> &middot; <a href=\"https:\/\/www.sdxcentral.com\/news\/scale-in-nvidias-fifth-pillar-of-ai-networking-rests-on-shaky-ground\/\">SDxCentral (the &#8216;scale-in&#8217; pillar)<\/a><\/p>\n<\/p><\/div>\n<div class=\"article\">\n<h4>5. Verizon, AT&amp;T and Lumen: the model, partnership and reach work that agentic operations actually rests on<\/h4>\n<p class=\"meta\">RCR Wireless News &middot; SDxCentral &middot; August 26&ndash;28, 2026<\/p>\n<p><strong>Verizon and Google Cloud<\/strong> expanded their AI partnership across three fronts: customer service, autonomous network operations and unified enterprise data. The customer-service half is already live rather than aspirational &mdash; Gemini now handles the majority of inbound consumer calls and chats in some channels, following earlier Vertex AI trials that Verizon reported as producing higher response accuracy and better agent productivity than its legacy service processes. The data half routes Verizon&#8217;s enterprise data through Google Cloud&#8217;s Agentic Data Cloud, with Google security and threat-detection tooling alongside. The network half is the one to watch, because it involves giving AI agents programmatic API access to network systems for automated patching and configuration changes. Two things are conspicuously absent: any disclosed financial terms, duration or firm timeline, and any published framework for how those network agents will be constrained, audited or overridden &mdash; precisely the gap TM Forum spent the same week arguing about, turning up in the largest operator deal of the week. Network data is the asset an operator has that a frontier model does not, and it is also the asset that, once it is flowing into a hyperscaler platform, becomes considerably harder to move. The strategic question for any operator signing a similar deal is which side of the boundary the domain knowledge ends up on: models tuned on your network&#8217;s behaviour are either a capability you own or one you rent, and the contract, not the architecture diagram, decides which.<\/p>\n<p><strong>AT&amp;T<\/strong> taking <strong>OTel 2.0<\/strong> into production is the item most likely to be misread from its name. This is not a telemetry release. OTel 2.0 is the GSMA-backed open telecom model &mdash; 31 billion parameters built on Google&#8217;s Gemma 4 31B-IT base and post-trained specifically for telecoms, with over a trillion tokens processed and distilled down to roughly 400 billion telecom-specific ones, drawing on a 15-billion-token curated dataset plus GSMA&#8217;s 10-billion-token Telco Corpus of 3GPP, ETSI, CAMARA, ITU, O-RAN and TM Forum material. Training ran on about 430 AMD Instinct MI300X GPUs through Microsoft Foundry on Azure; AT&amp;T runs it on-premises on Dell servers with MI355X parts, with roughly 40% of its AI operations now on AMD hardware. The scale is the point: AT&amp;T processes 45 billion tokens a day and more than a trillion a month across over 100 generative AI models, and reports that gateway routing and cached computation reuse cut inference costs by up to 90% compared with sending the same work to frontier models. Weights are published on Hugging Face with weekly updates. For a NetOps team the read is direct &mdash; a domain-tuned open model that already speaks 3GPP and TM Forum vocabulary removes much of the retrieval and prompt scaffolding a general-purpose model needs before it can be trusted with a standards question, and it does so on hardware you control rather than by shipping operational context to a third party. The 90% cost delta is the number that decides whether agentic operations is affordable at production volume rather than at demo volume.<\/p>\n<p><strong>Lumen<\/strong> extended its Multi-Cloud Gateway well past its own fibre footprint. The private layer-3 service, launched earlier this year and managed through the Lumen Connect portal and its APIs, is now available at more than 10 million US business locations, including qualified third-party connected sites outside Lumen&#8217;s owned network, with feature parity maintained across the wider footprint &mdash; routing policy, segmentation control and digital service control via Ethernet Fabric Connect all follow the customer off-net. AI workloads generate exactly the traffic pattern these products were built for &mdash; substantial, bursty, and moving between more than one provider &mdash; and pushing the gateway function out of the core toward customer sites shortens the path. This is also, quietly, the operator answer to the disintermediation worry that has run through this bulletin all summer: if the hyperscalers increasingly own their own long-haul between data centres, the durable carrier position is in the messy last stretch between enterprise premises and multiple clouds, a business hyperscalers have shown little appetite for. The interesting admission is structural. Building reach on other operators&#8217; access says the owned fibre is no longer the differentiator; the gateway, the portal and the API surface around them are.<\/p>\n<p style=\"font-size:13px;color:#6b7280;margin:0;\">Sources: <a href=\"https:\/\/rcrwireless.com\/20260826\/ai\/verizon-google-cloud-ai-partnership\">RCR Wireless News (Verizon and Google Cloud)<\/a> &middot; <a href=\"https:\/\/rcrwireless.com\/20260828\/ai\/att-otel-2-0-production\">RCR Wireless News (AT&amp;T takes OTel 2.0 into production)<\/a> &middot; <a href=\"https:\/\/www.sdxcentral.com\/news\/lumen-extends-multi-cloud-gateway-outside-of-core-network\/\">SDxCentral (Lumen Multi-Cloud Gateway)<\/a><\/p>\n<\/p><\/div>\n<div class=\"article\">\n<h4>6. Optical networks move past cruise control &mdash; and the OIF sets the interop agenda for ECOC<\/h4>\n<p class=\"meta\">Nokia Blog &middot; Converge Digest &middot; Light Reading &middot; Fierce Network &middot; August 26&ndash;27, 2026<\/p>\n<p><strong>Nokia<\/strong>&#8216;s framing &mdash; moving optical networks beyond cruise control to agentic operations &mdash; is the best metaphor of the week because it names a distinction vendors usually blur. Cruise control holds a set point: it is closed-loop, it is automated, and it has been shipping in optical control planes for years under names like automatic power balancing and restoration. What it does not do is decide the set point is wrong. Nokia&#8217;s answer is optical-domain agents inside its WaveSuite and Transcend automation portfolio, reasoning across alarms, KPIs, topology and logs to surface anomalies, capacity drift and silent degradation, then recommending corrective actions for human review rather than executing them &mdash; with real-time explainability and auditable logs attached, which is the governance-first pattern from earlier in this issue showing up as shipped product. Two implementation details deserve an architect&#8217;s attention. Nokia supports bring-your-own-cloud deployment across on-premises, private, hybrid and operator-hosted environments, which matters when the telemetry involved is commercially sensitive or subject to residency rules. And it ships token-consumption dashboards for tracking GPU resource use &mdash; an unglamorous acknowledgement that inference cost is now an operational line item, and that an agent left reasoning continuously over an optical estate produces a bill somebody has to defend. Optical is a good proving ground precisely because the state space is comparatively well-instrumented and the physics constrain the plausible actions; it is a bad one to be casual in, for the same reasons, because optical changes are slow to reverse and their blast radius is measured in services rather than sessions.<\/p>\n<p>The <strong>OIF<\/strong> takes its interoperability programme to <strong>ECOC 2026<\/strong> in M&aacute;laga on 21&ndash;23 September, with 39 member companies participating and 22 vendors in the optical showcase at booth 2126. The demonstration list is the roadmap made concrete: a multi-vendor 448G electrical interface demo pushing past 224G per lane; 400ZR and 800ZR coherent implementations shown across both data-centre interconnect and scale-across architectures; CEI-224G interoperability spanning very-short-reach, long-reach and linear links; CMIS common module management including firmware update and link training; energy-efficient interface work covering retimed, half-retimed RTLR and unretimed linear architectures; a 12.8 Tb\/s near-package optics implementation alongside co-packaged optics; multi-core fibre at both 500 metres and 50 kilometres; and a new Compute Optics Interface framework aimed squarely at optical connectivity inside AI scale-up systems. OIF president Nathan Tracy&#8217;s summary &mdash; that members will show how collaboration turns complex specifications into multi-vendor solutions spanning optics, electrical interfaces, management and energy efficiency &mdash; is the whole value of the exercise. A spec says what should work; an interop event says what did, between named implementations. That the Compute Optics Interface and scale-across strands now sit beside traditional transport work is the clearest available signal of how far AI-cluster requirements have pulled the optics agenda. If you have 2028-and-beyond fabric decisions in front of you, the ECOC results are the evidence base worth waiting for before committing to a reach assumption.<\/p>\n<p>Two CEOs supplied the demand-side framing, and both had substance behind the rhetoric. <strong>Calix<\/strong> CEO Michael Weening &mdash; &ldquo;I&#8217;m a firm believer AI changes everything, absolutely flat-out everything&rdquo; &mdash; is making an argument about what broadband service providers can operate and sell, not about data-centre fabrics. Calix has put more than $200 million into platform development since 2023, runs its agentic broadband platform on Google Cloud&#8217;s AI infrastructure alongside an Agent Workforce Cloud, and points at concrete uses among providers such as RTC Networks and Tombigbee: identifying subscribers likely to churn, correlating behavioural and warranty data to find households without outdoor Wi-Fi worth upselling, and listening to support calls to shorten resolution. The company expects customers to see a 50&ndash;60% velocity gain in capability delivery during 2027. <strong>Graphiant<\/strong>&#8216;s Ali Shaikh, CEO since October 2025 and previously its chief product officer, reports telcos moving aggressively rather than defensively &mdash; a notable shift from the disintermediation anxiety that has dominated operator commentary for most of this year. His is a roughly 50-person, $120-million-funded network-as-a-service company that rebuilt the protocol stack at the IP layer with routing software written in Rust and charges by capacity rather than per user, feature, box or port, aimed at circuit provisioning that still takes six months; AT&amp;T, Sony Pictures, Canva and Peraton are named among its users, and stc&#8217;s Tali Ventures co-led its most recent funding extension. Both readings can be true: the largest carriers are worried about hyperscaler encroachment on long-haul and platform economics, while the operators closer to the customer see AI as the first genuinely new service category in a long time.<\/p>\n<p style=\"font-size:13px;color:#6b7280;margin:0;\">Sources: <a href=\"https:\/\/www.nokia.com\/blog\/how-do-optical-networks-move-beyond-cruise-control-to-agentic-operations\/\">Nokia Blog (beyond cruise control to agentic operations)<\/a> &middot; <a href=\"https:\/\/convergedigest.com\/oif-448g-800zr-cmis-ai-interoperability-ecoc-2026\/\">Converge Digest (OIF at ECOC 2026)<\/a> &middot; <a href=\"https:\/\/www.lightreading.com\/ai-machine-learning\/calix-ceo-is-a-firm-believer-that-ai-changes-everything-\">Light Reading (Calix CEO on AI)<\/a> &middot; <a href=\"https:\/\/www.fierce-network.com\/cloud\/telcos-move-aggressively-seize-ai-opportunity-graphiant-ceo\">Fierce Network (Graphiant CEO on telco AI)<\/a><\/p>\n<\/p><\/div>\n<p>            <!-- Foundational reading --><\/p>\n<h2>Foundational reading<\/h2>\n<div class=\"article\">\n<h4>Harnesses, APIs, MCP, self-aware RAN and the layer-2 problem nobody budgets for<\/h4>\n<p class=\"meta\">Ubuntu (Canonical) &middot; TechTarget &middot; Versa Networks &middot; VIAVI Perspectives &middot; The New Stack &middot; August 8&ndash;28, 2026<\/p>\n<p>Five pieces that sit underneath this week&#8217;s news. <a href=\"https:\/\/ubuntu.com\/blog\/ai-harnesses-for-telco-autonomous-networks\">Canonical on AI harnesses for telco autonomous networks<\/a> uses a word worth adopting and then defines it properly: an end-to-end operational AI harness is the whole apparatus around the model &mdash; open cloud substrate, interoperable agent protocols, trusted operational context, declarative GitOps workflows, confidential inference, intent-driven orchestration, policy enforcement and closed-loop assurance. Its central architectural rule is the one to take away: separate the agentic read path, where observation and cognition happen, from the network&#8217;s write path, where governance and execution live, rather than wiring a model directly into critical network functions. The stack it describes is concrete &mdash; MAAS, Canonical Kubernetes, MicroCloud, Juju and charmed operators, Ubuntu Core and Charmed MLOps, with confidential inference resting on AMD SEV-SNP, Intel TDX or Nvidia confidential GPUs, and MCP and the Agentic AI Foundation supplying the agent protocols &mdash; but the read\/write separation survives independent of the vendor. It is the same insight TM Forum&#8217;s governance piece reaches from the process side, expressed as engineering. The model is increasingly the commodity; the harness is the differentiated, operator-specific asset, and it is also the part that cannot be procured whole. <a href=\"https:\/\/www.techtarget.com\/it-infrastructure\/tip\/Why-agentic-AI-needs-network-APIs-to-succeed\">TechTarget&#8217;s argument that agentic AI needs network APIs to succeed<\/a> is the concrete version of the same claim: APIs are the gateways that let an agent retrieve operational data, orchestrate workflows, implement changes and verify outcomes while security controls stay intact. It is specific about which ones matter &mdash; NETCONF, RESTCONF and REST underneath, OAuth 2.0 and mutual TLS around them, MCP and the AGNTCY effort above &mdash; and the corollary is a procurement rule: ask what an agentic platform does when it must act on a device with no usable API, and treat &ldquo;it drives the CLI&rdquo; as the answer it is.<\/p>\n<p><a href=\"https:\/\/versa-networks.com\/blog\/mcp-enterprise-infrastructure-zero-trust-governance\/\">Versa&#8217;s piece on MCP as a new operating model for enterprise infrastructure<\/a> extends that into the interface layer that has emerged over the past year. The Model Context Protocol standardises how an agent discovers and invokes the tools available to it, and Versa&#8217;s argument is that a per-vendor MCP server solves the one-tool-per-vendor problem while leaving the one-enterprise problem untouched: what infrastructure teams need is a routing and governance layer above every MCP server, where each request is authenticated, scoped and audited and access control decides which servers and agents may run at all. Its figures are worth testing against your own estate &mdash; NOC and SOC teams losing up to 70% of incident response time simply gathering information across consoles, and a claimed 45% reduction in mean time to resolution among customers running its MCP server. Vendor numbers, both of them, but the first will feel familiar to anyone who has watched an incident bridge. Either way, an MCP server in front of network infrastructure is a privileged access path and deserves the review a jump host would get. <a href=\"https:\/\/blog.viavisolutions.com\/2026\/08\/24\/the-journey-to-a-self-aware-autonomous-ran\/\">VIAVI on the journey to a self-aware autonomous RAN<\/a> makes the complementary point about the input side, and supplies the week&#8217;s most sobering statistics: 54% of operators already use AI for network planning and optimisation, yet more than half remain stuck in pilots, while energy accounts for as much as a quarter of a mobile operator&#8217;s costs. VIAVI&#8217;s answer is validation infrastructure &mdash; its TeraVM AI RAN Scenario Generator and RAN simulation digital twin exist to baseline performance and train against modelled conditions before anything touches a live network, with TM Forum&#8217;s six autonomy levels as the yardstick and Level 3 as the realistic target. Its framing is the right one: the operators who get there first will not be the ones with the most AI, but the ones who can trust it. Self-aware precedes autonomous, in that order, and the gap between the two is usually measurement.<\/p>\n<p>Finally, <a href=\"https:\/\/thenewstack.io\/kubevirt-evpn-vm-migration\/\">The New Stack on why KubeVirt VMs cannot move between clusters and how EVPN fixes it<\/a> is the most immediately actionable item here for anyone running virtualised workloads on Kubernetes, and it comes from Miguel Duarte Barroso, a Red Hat principal engineer on OpenShift Virtualization and a KubeVirt maintainer. The problem is an old one wearing new clothes, and it has two halves: live migration requires the VM to keep the same MAC and IP address on the destination cluster, which means a stretched layer-2 domain across what is normally a layer-3 boundary; and the memory-state transfer needs an isolated path of its own rather than competing with application traffic. The fix is EVPN with a BGP control plane over VXLAN, driven from Kubernetes custom resources through the open-source OpenPERouter &mdash; an Underlay resource establishing BGP peering with the top-of-rack switches, an L2VNI resource creating the stretched layer-2 segment with its VNI and VRF scoping, and an L3VNI resource handling inter-subnet and external routing. The worked example separates application traffic on VNI 110 from migration traffic on VNI 666 in different VRFs, which is the detail most improvised attempts miss. It is included here because it is a recurring pattern rather than a KubeVirt quirk: workload mobility gets designed on an assumption about layer 2 that the network team first hears about at the change window.<\/p>\n<p style=\"font-size:13px;color:#6b7280;margin:0;\">Sources: <a href=\"https:\/\/ubuntu.com\/blog\/ai-harnesses-for-telco-autonomous-networks\">Ubuntu (AI harnesses for telco autonomous networks)<\/a> &middot; <a href=\"https:\/\/www.techtarget.com\/it-infrastructure\/tip\/Why-agentic-AI-needs-network-APIs-to-succeed\">TechTarget (why agentic AI needs network APIs)<\/a> &middot; <a href=\"https:\/\/versa-networks.com\/blog\/mcp-enterprise-infrastructure-zero-trust-governance\/\">Versa Networks (MCP and enterprise infrastructure)<\/a> &middot; <a href=\"https:\/\/blog.viavisolutions.com\/2026\/08\/24\/the-journey-to-a-self-aware-autonomous-ran\/\">VIAVI Perspectives (self-aware autonomous RAN)<\/a> &middot; <a href=\"https:\/\/thenewstack.io\/kubevirt-evpn-vm-migration\/\">The New Stack (KubeVirt, VM migration and EVPN)<\/a><\/p>\n<\/p><\/div>\n<p>            <!-- Watch list --><\/p>\n<div class=\"watchlist\">\n<h2>Calls to action &amp; watch list<\/h2>\n<ul>\n<li><strong>Evaluate Cloud Control on multi-vendor reasoning, not single-vendor demos.<\/strong> If Cisco Cloud Control is on your shortlist, build the proof-of-concept around an incident that spans a non-Cisco device or a third-party observability source. The value of a cross-domain control plane is entirely in how far the domain extends, and that is the question a curated demo will not answer for you.<\/li>\n<li><strong>Give agents identities before they get write access.<\/strong> TM Forum&#8217;s governance-first argument reduces to one testable question: can you name, scope and audit every agent acting on your network today? If any of them run under a shared service account, that is this quarter&#8217;s first ticket, and it is work your existing IAM stack can start on now.<\/li>\n<li><strong>Scope the autonomy business case past the network domain.<\/strong> The enterprise-wide value argument is the one that survives a budget review. Pick one workflow that crosses from assurance into order management or field operations, and model the benefit there &mdash; a network-only case will be measured against a network-only cost and will lose.<\/li>\n<li><strong>Benchmark a domain-tuned open model before you commit to a frontier one.<\/strong> AT&amp;T&#8217;s OTel 2.0 is the GSMA-backed open telecom model, published on Hugging Face and running on-premises, and it already speaks 3GPP, ETSI, O-RAN and TM Forum vocabulary without the retrieval scaffolding a general-purpose model needs. Test it on your own standards and assurance questions before you accept a frontier model&#8217;s per-token bill as the price of agentic operations.<\/li>\n<li><strong>Ask your fabric vendors where they stand on multipath and out-of-order delivery.<\/strong> MetaRoCE and Nvidia&#8217;s multiplane architecture attack the same problem from the transport and topology sides. If your AI fabric still depends on ECMP hashing landing flows evenly, you have a tail-latency exposure that will not show up in link utilisation graphs.<\/li>\n<li><strong>Treat NVLink Fusion as an architectural commitment, not a feature.<\/strong> Pooled memory reached over a single vendor&#8217;s scale-up interconnect is genuinely powerful and genuinely sticky. Decide deliberately whether you want that interconnect to be structural in your rack, and write down what a migration would cost before you need the answer.<\/li>\n<li><strong>Interrogate new taxonomy before it reaches your RFP.<\/strong> The scepticism about &ldquo;scale-in&rdquo; is a useful habit generally: when a vendor introduces a new architectural pillar, check whether it names a problem you had before the category existed. Vendor-coined categories have a way of becoming sole-source requirements in a specification.<\/li>\n<li><strong>Wait for the ECOC interop results before fixing your reach assumptions.<\/strong> The OIF&#8217;s 448G, 800ZR, CMIS and AI interoperability work at ECOC 2026 is the evidence base for post-2028 fabric planning. Roadmap slides and interop demonstrations between named implementations are not the same category of information.<\/li>\n<li><strong>Review your MCP exposure like a privileged access path.<\/strong> If any part of your infrastructure is reachable through an MCP server, inventory the tools it exposes, the scope of each, and who can invoke them. It deserves the review a jump host would get, and most organisations have not given it one.<\/li>\n<li><strong>Put the fabric requirement into the platform design review.<\/strong> The KubeVirt and EVPN case is a specific instance of a general failure: workload mobility is designed on an assumption about layer 2 that the network team learns about at the change window. Get the network into the platform review while the assumption is still cheap to change.<\/li>\n<\/ul><\/div>\n<\/td>\n<\/tr>\n<p>        <!-- Footer --><\/p>\n<tr>\n<td class=\"footer\">\n<p class=\"brand\">Agentic NetOps<\/p>\n<p>A weekly intelligence bulletin from Security Radar LLC.<br \/>\n            Curated by Paul Davis &middot; <a href=\"mailto:paul.davis@security-radar.com\">paul.davis@security-radar.com<\/a><\/p>\n<p>&copy; 2026 Security Radar LLC. All rights reserved.<\/p>\n<p>Article titles and summaries are excerpted for review and commentary; all linked articles remain the copyright of their respective publishers and authors.<\/p>\n<p>*|LIST:ADDRESS|*<\/p>\n<p><a href=\"*|ARCHIVE|*\">View this email in your browser<\/a> &middot; <a href=\"*|UNSUB|*\">Unsubscribe<\/a><\/p>\n<\/td>\n<\/tr>\n<\/table>\n<\/td>\n<\/tr>\n<\/table>\n","protected":false},"excerpt":{"rendered":"<p>August 30, 2026 &middot; Weekly Edition Agentic NetOps Cisco Cloud Control reaches general availability and the agentic operations pitch stops being a roadmap slide. TM Forum answers with the uncomfortable half of the story &mdash; governance first, or the autonomy programme stalls at the pilot. Meanwhile the fabric people had&#8230;<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"closed","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[45,11],"tags":[],"class_list":["post-5775","post","type-post","status-publish","format-standard","hentry","category-ai-ml","category-trends"],"_links":{"self":[{"href":"https:\/\/www.cybersecurityinstitute.com\/blog\/index.php?rest_route=\/wp\/v2\/posts\/5775","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.cybersecurityinstitute.com\/blog\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.cybersecurityinstitute.com\/blog\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.cybersecurityinstitute.com\/blog\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.cybersecurityinstitute.com\/blog\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=5775"}],"version-history":[{"count":1,"href":"https:\/\/www.cybersecurityinstitute.com\/blog\/index.php?rest_route=\/wp\/v2\/posts\/5775\/revisions"}],"predecessor-version":[{"id":5816,"href":"https:\/\/www.cybersecurityinstitute.com\/blog\/index.php?rest_route=\/wp\/v2\/posts\/5775\/revisions\/5816"}],"wp:attachment":[{"href":"https:\/\/www.cybersecurityinstitute.com\/blog\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=5775"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.cybersecurityinstitute.com\/blog\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=5775"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.cybersecurityinstitute.com\/blog\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=5775"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}