You're reading the Saturday, November 21, 2026 edition. Today's briefing →
Toronto
Saturday, November 21, 2026No. 160

Digital Plumber

Plumbing the information age

AI-curated intelligence for people who run networks. Daily coverage of AIOps, network automation, agentic operations, AI infrastructure, security and the vendors shaping them.

Today's 3 things that matter

Picked by the AI editor
  1. DC networking·Industry news

    Lambda runs Nvidia co-packaged optics in 10,000-GPU cluster

    Lambda has deployed Nvidia's co-packaged optics technology in an AI cluster connecting more than 10,000 Blackwell Ultra GPUs, demonstrating production-scale CPO viability.

    Why it matters CPO power reductions accelerate 1.6T optics ROI at scale; validates architecture for 128K+ GPU fabric planning and power budgeting in production data center builds.

  2. Security·Primary source

    Attackers root Cisco Expressway clusters via hard-coded credentials

    CVE-2026-76466 (CVSS 9.8), a critical flaw affecting Cisco Expressway Series clustered deployments, involves hard-coded credentials and insufficient validation of client certificates.

    Why it matters A maintenance identity and its cookie are hard-coded into the software and are the same on every install, allowing anyone who can reach port 4372 to join the cluster and run commands as root without a password. Cisco collaboration edge gateways are internet-facing perimeter devices critical to call control.

  3. Automation·Primary source

    SCinet builds SC26's record network on Nautobot source of truth

    SCinet, the temporary supercomputing network at SC26, uses Nautobot as its network source of truth to automate design, deployment, and teardown of a large-scale research network.

    Why it matters Demonstrates Nautobot at production scale in a time-critical, high-stakes environment where network changes must be automated, validated, and reversible—a real-world validation of NSoT-driven automation patterns for practitioners.

Today's briefing

What happened, and why it matters

18 stories · 11 topics · Updated 12:25 PM ET

SCinet builds SC26's record network on Nautobot source of truth

Network to Code · Nov 17, 2026 · Primary source

SCinet, the temporary supercomputing network at SC26, uses Nautobot as its network source of truth to automate design, deployment, and teardown of a large-scale research network. The platform enables rapid modeling of network infrastructure and automates repetitive tasks like cabling, hardware deployment, and IP schema configuration.

Why it matters Demonstrates Nautobot at production scale in a time-critical, high-stakes environment where network changes must be automated, validated, and reversible—a real-world validation of NSoT-driven automation patterns for practitioners.

SCinet, which connects some of the world's fastest supercomputers and research networks at the annual SC conference, historically built its temporary network manually each November. For SC26 in Chicago, SCinet adopted Nautobot as a shared data foundation across Automation and DevOps teams, automating design and planning workflows that previously took weeks. By modeling the entire network infrastructure in Nautobot before physical deployment, SCinet teams reduced manual configuration tasks (cabling diagrams, IP assignments, device staging) from manual spreadsheet-driven processes to automated workflows. The platform provided a single source of truth that allowed distributed volunteer teams to collaborate safely on a network that must reach production readiness by a fixed date—demonstrating how NSoT platforms enable both speed and safety in large-scale, short-duration network operations. This is not a pilot; it is a full operational deployment automating thousands of devices and connections.

Read the original at networktocode.com ↗

OpsMill opens 2026 Nornir roadmap to community feedback

OpsMill · Nov 16, 2026 · Primary source

OpsMill published a 2026 Nornir roadmap with seven proposals for the Python automation framework, including async inventory support and improved integration with remote sources of truth like Infrahub, NetBox, and Nautobot. All changes are additive and maintain backward compatibility.

Why it matters Nornir's async inventory capability directly addresses a pain point for teams whose NSoT is remote (HTTP/GraphQL APIs); practitioners will be able to load device inventory from NetBox/Nautobot/Infrahub without blocking task execution, improving performance on large inventories.

Nornir's roadmap emphasizes three principles: the architecture stays small and focused; new capabilities are delivered as plugins; all changes are additive with no breaking changes to existing code. Key proposal: async inventory support extends Nornir's runner concurrency to the inventory loading phase itself. Many teams store device inventory in HTTP or GraphQL APIs—Infrahub, NetBox, Nautobot—and the inventory plugin can spend significant time waiting for remote requests. Async inventory removes that blocking behavior, allowing teams to load thousands of devices in parallel. The roadmap also proposes improved plugin extensibility and tighter integrations with modern infrastructure-as-code tools. OpsMill committed to a stable 3.7 release within four to six weeks if testing proceeds on schedule. This is not a major rewrite; it is careful, community-driven incremental improvement aligned with how automation teams actually use Nornir.

Read the original at opsmill.com ↗

Dynatrace named leader in first observability pipelines Magic Quadrant

Dynatrace · Nov 17, 2026 · Primary source

Gartner named Dynatrace (Bindplane) a Leader in the inaugural 2026 Gartner Magic Quadrant for Observability Pipelines. Dynatrace (Bindplane) ranked highest in three of five use cases: DevOps/AppDev (3.73), Platform Operations (3.70), and SecOps (3.68).

Why it matters Bindplane is an OpenTelemetry-native observability pipeline that uses AI to collect, optimize, govern, and route telemetry from any source to any destination—critical for NOC teams managing multi-vendor telemetry cost and governance at scale.

Dynatrace announced on November 17, 2026 that Gartner named Dynatrace (Bindplane) a Leader in the inaugural 2026 Gartner Magic Quadrant for Observability Pipelines. Acquired by Dynatrace in April 2026, Bindplane is an OpenTelemetry-native observability pipeline that uses AI to help organizations collect, optimize, govern, and route telemetry from any source to any destination before it reaches downstream observability, security, and analytics platforms. The platform ranked highest in three of five Critical Capabilities use cases: DevOps/AppDev (3.73 out of 5), Platform Operations (3.70 out of 5), and SecOps (3.68 out of 5, tied). This positions Bindplane as a core infrastructure layer for telemetry pipelines in observability stacks—particularly relevant for NOC practitioners managing OpenTelemetry adoption, telemetry cost optimization, and cross-domain correlation at scale. The recognition reflects maturity in handling the telemetry volume and governance challenges that drive modern AIOps workflows.

Read the original at ir.dynatrace.com ↗

Lambda runs Nvidia co-packaged optics in 10,000-GPU cluster

TechTarget · Nov 17, 2026 · Industry news

Lambda has deployed Nvidia's co-packaged optics technology in an AI cluster connecting more than 10,000 Blackwell Ultra GPUs, demonstrating production-scale CPO viability. Nvidia said each CPO switch replaces 25-watt, 1.6-Tb/s transceivers with optical components consuming five watts for the same bandwidth, amounting to a potential 1.44-kilowatt reduction per switch in optical-component power across 72 transceivers.

Why it matters CPO power reductions accelerate 1.6T optics ROI at scale; validates architecture for 128K+ GPU fabric planning and power budgeting in production data center builds.

Lambda's Quebec deployment of NVIDIA's co-packaged optics (CPO) with Spectrum-X Photonics switches marks the first significant production use of integrated photonic engines in hyperscale AI clusters. The architecture integrates optical components directly on the switch package rather than using pluggable transceivers, reducing electrical trace lengths and signal loss. NVIDIA's power comparison shows each optical transceiver component dropping from 25W to 5W—a 5x reduction—though total switch power also improves from an estimated 7kW (conventional) to 3.95kW (CPO). Lambda's modeling suggests a 128,000-GPU data center would need approximately 655,000 discrete transceiver modules with conventional pluggable optics, each consuming power and representing a failure point. The CPO architecture eliminates this mechanical and thermal overhead. While NVIDIA's component-level power savings are modeled rather than measured at Lambda's site, the deployment validates that CPO switches can reach production scale with AMD Instinct accelerators and function reliably in a managed AI cloud environment. This matters for practitioners planning 2027-2028 AI fabric refreshes where power density is becoming the binding constraint on GPU cluster scale.

Read the original at techtarget.com ↗

Oak Ridge Lux supercomputer runs Juniper 800G Ethernet fabric

TechTarget · Nov 16, 2026 · Industry news

HPE announced that Lux, an AI cluster for Oak Ridge National Laboratory, debuted at No. 14 on the TOP500, built on HPE ProLiant Compute XD685 and Juniper Networking data center solutions with 800G Ethernet, featuring Juniper QFX Switches with load balancing and congestion management. Lux is the first DOE national laboratory deployment of HPE Networking solutions to bring scalable, high-performance, and open Ethernet to AI-driven science.

Why it matters Production validation of 800G open Ethernet for AI at scale in government HPC; demonstrates Juniper QFX fabric viability for 220+ petaflops workloads and informs multi-vendor fabric strategy for large clusters.

Oak Ridge National Laboratory's Lux system demonstrates 800G Ethernet as production-ready for massive AI clusters in federal research settings. The system ranked #14 on the November 2026 TOP500 with 220.79 petaflops on High Performance Linpack, pairing AMD Instinct MI355X GPUs and Pensando DPUs with HPE Juniper QFX 800G switches. The deployment validates Juniper's Express 5 ASIC at scale in a real workload environment. HPE disclosed that Lux uses Juniper QFX switching with load balancing and congestion management to reduce network bottlenecks and optimize GPU utilization—addressing collective-communication patterns in large model training. This is HPE's first DOE lab deployment of Juniper networking post-acquisition (July 2025), positioning open Ethernet as the fabric of choice for US government AI infrastructure rather than proprietary InfiniBand or vendor-locked solutions. Lux runs alongside Frontier, Oak Ridge's exascale Cray Slingshot system (#3 on TOP500), creating a bifurcated AI/HPC infrastructure model at the lab. For practitioners evaluating data center network refresh decisions, Lux's performance ranking and DOE endorsement validates that open 800G Ethernet can deliver performance parity with traditional HPC networks when paired with sophisticated congestion control and load balancing. The system also demonstrates interoperability between AMD accelerators, HPE compute, and Juniper switching in a government procurement context.

Read the original at techtarget.com ↗

China-linked group exploited critical VMware vCenter bypass since September

Google Threat Intelligence Group / Mandiant · Nov 19, 2026 · Primary source

A suspected China-linked espionage group ARMORY CASTLE exploited CVE-2026-123890, a critical authentication bypass (CVSS 9.8) in VMware vCenter Server Directory Service, in early September 2026. The group used a custom tool called BONESAW to enable remote LMDB exfiltration, SAML STS hijacking, and SSO administrator account creation on vCenter servers.

Why it matters There is no workaround; vSphere teams need both the update and a compromise review. Virtualization infrastructure compromise represents complete infrastructure control and lateral movement capability.

ARMORY CASTLE (formerly UNC5221) exploited this zero-day from at least early September, more than two months before Broadcom released a fix on November 19 in advisory VMSA-2026-0009. CVE-2026-123890 is an unauthenticated Secure Remote Password (SRP-6a) session key disclosure vulnerability over Distributed Computing Environment/Remote Procedure Calls (DCE/RPC). Port 2012 is used as the control interface RPC for vCenter Single Sign-On (SSO) and was integral to the remote exploitation. After creating a listener on TCP port 33992, ARMORY CASTLE moved laterally to the vCenter appliance as Administrator, created local admin accounts, and enabled SSH via the vCenter Server Appliance Management Interface (VAMI) API. Fixed in vCenter 9.1.1.0200, 9.0.2.0200 and 8.0 U3m. The two-month unpatched window and use of specialized malware indicates nation-state-grade targeting of virtualization infrastructure.

Read the original at cloud.google.com ↗

Attackers root Cisco Expressway clusters via hard-coded credentials

Secure.com / CISA KEV Catalog · Nov 20, 2026 · Primary source

CVE-2026-76466 (CVSS 9.8), a critical flaw affecting Cisco Expressway Series clustered deployments, involves hard-coded credentials and insufficient validation of client certificates. On November 19, attackers rooted four clustered Expressway X15.5.3 devices at a Russian organization, two weeks after exploit code appeared on GitHub.

Why it matters A maintenance identity and its cookie are hard-coded into the software and are the same on every install, allowing anyone who can reach port 4372 to join the cluster and run commands as root without a password. Cisco collaboration edge gateways are internet-facing perimeter devices critical to call control.

CISA added CVE-2026-76466 to its Known Exploited Vulnerabilities catalog on November 20 with a November 23 deadline; Cisco fixed it on October 21 in Expressway X15.5.20. The bug sits in the cluster database service, which runs on Erlang and listens on TCP 4372. Once inside, attackers ran custom Python scripts that scanned for SMB shares, probed SQL Server and WinRM, and pulled credentials from internal GitLab repositories. X14 and earlier have no fix, so Cisco recommends migrating to a fixed release. Cisco lists no workaround. The hardcoded nature means every Expressway install worldwide shares the same exposed credential.

Read the original at secure.com ↗

CISA adds Barracuda, Cisco BroadWorks and Synology flaws to KEV

CISA Known Exploited Vulnerabilities Catalog · Nov 18, 2026 · Primary source

On November 18, CISA added four critical vulnerabilities including CVE-2024-10441 (Synology DiskStation Manager), CVE-2026-40217 (BerriAI LiteLLM sandbox escape), CVE-2026-90436 (Apache Axis unsafe reflection RCE), and CVE-2026-90817 (Vanderbilt REDCap code injection). On November 10, CISA added six vulnerabilities including CVE-2026-20320 (Cisco BroadWorks XXE), CVE-2026-116606 (Barracuda ESG auth bypass), and CVE-2026-111560 (Barracuda OS command injection).

Why it matters Email gateways (Barracuda), voice/comms platforms (Cisco BroadWorks), and NAS devices (Synology) are high-value targets; OS command injection and auth bypass flaws in gateway appliances directly expose internal networks to unauthenticated RCE.

Barracuda flaws are particularly significant for network operations teams: CVE-2026-116606 is an authentication bypass using an alternate path or channel in Email Security Gateway, and CVE-2026-111560 is an OS command injection affecting multiple Barracuda products. These gateway appliances sit at network edges and control mail traffic for millions of organizations. Network administrators should prioritize email security appliance patching alongside CISA deadline compliance. CVE-2024-10441 in Synology DSM is being actively exploited for C3RB3R ransomware deployment on network-attached storage.

Read the original at cisa.gov ↗

Metano leaves stealth with $12 million to police AI agents

i10x.ai · Nov 19, 2026 · Industry news

Metano, an AI agent security startup, emerged from stealth on November 19 with $12M seed funding led by Sierra Ventures. The platform discovers, monitors, and enforces controls on autonomous AI agents across enterprise systems to prevent malicious or unauthorized actions.

Why it matters Network and security practitioners need agent governance as AI systems gain autonomy; Metano's runtime enforcement approach applies control-plane thinking to AI agent behavior.

Metano launched with $12 million in seed funding led by Sierra Ventures, with participation from Correlation Ventures, High Alpha, and 360 ONE. Co-founded by Neelima Rustagi (CEO) and Ratnesh Pandey (CTO), both with deep security backgrounds from ServiceNow, Zscaler, Palo Alto Networks, FireEye, and Microsoft, Metano focuses on agent discovery, context-aware threat detection, and real-time enforcement across endpoints, cloud, and SaaS environments. The platform uses a context graph called Hive to judge whether an agent's session behavior matches its intended purpose, then enforces guardrails where agents operate. The company completed SOC 2 Type 2 and holds ISO 27001:2022 certification. This addresses an emerging gap in infrastructure security: traditional tools weren't designed for systems that make autonomous decisions and change behavior during execution. For infrastructure teams, the risk model shifts from static configuration to behavioral anomaly detection—similar to network anomaly detection, but applied to agent intent and action alignment.

Read the original at i10x.ai ↗

StepSecurity raises $20 million to block supply chain attacks

Axios Pro · Nov 18, 2026 · Industry news

StepSecurity raised $20M in Series A funding led by Elephant to expand protection across developer machines, package downloads, CI/CD pipelines, and emerging AI agent sandboxes. The company protects 20,000+ open-source projects and detected recent supply chain worms like the Shai-Hulud attack.

Why it matters DevNetOps and automation practitioners rely on CI/CD pipelines and package registries; StepSecurity's supply-chain controls prevent compromised dependencies from reaching production infrastructure.

StepSecurity announced an oversubscribed $20 million Series A led by Elephant, with participation from existing investors Runtime Ventures, Inner Loop Capital, and SaaS Ventures, bringing total funding to $23 million. Founded in 2021 by Varun Sharma (CEO) and Ashish Kurmi (CTO), the company provides protection across developer machines, repository checks, a Secure Registry, and Harden-Runner for CI/CD—detecting and blocking malicious packages before they reach build systems. The platform achieved 5x ARR growth in 2026 and protects customers including Circle, Dexcom, Kong, Mercari, Chainguard, and XBOW. The timing is relevant: the company detected the tj-actions compromise and malicious axios releases through community telemetry, underscoring the value of visibility into dependency provenance. For infrastructure operators, this closes a gap in the supply-chain attack surface: malicious packages in CI/CD runners or development environments can compromise automation scripts, Terraform modules, and infrastructure-as-code artifacts before human review.

Read the original at axios.com ↗

Salesforce hires Cisco distinguished engineer Peter Bosch

SDxCentral · Nov 16, 2026 · Industry news

Salesforce hired Peter Bosch, a long-serving Cisco engineer, to serve as VP of engineering. Bosch joined Cisco in 2012 and held multiple roles, culminating in a distinguished engineering role leading the company's Emerging Technologies and Incubation (ET&I)—now known as Outshift—startup incubator.

Why it matters Salesforce acquires deep distributed systems and network architecture expertise from Cisco's innovation division during its period of engineering focus.

Bosch served at Cisco starting in 2012 across multiple roles, culminating in a distinguished engineering role leading the network giant's Emerging Technologies and Incubation (ET&I) startup incubator, now known as Outshift. Before Cisco, he joined from Juniper Networks as a distinguished engineer and director of engineering. At Cisco, he worked in the company's chief technology and architecture office and on virtualized mobility solutions. Bosch also serves as an associate professor at the University of Twente, leading lectures on distributed systems. He confirmed his departure in a LinkedIn post stating: 'Great time at Cisco, great company to work for, yet it is time for something new'.

Read the original at sdxcentral.com ↗

Azure SRE Agent adds AWS and Google Cloud connectors

Microsoft Tech Community · Nov 18, 2026 · Primary source

Azure SRE Agent joins Microsoft Teams with its own Entra Agent ID and now works across clouds with AWS and Google Cloud connectors and agent collaboration. Cross-cloud and cross-team agent communication now available.

Why it matters 11,000+ agents deployed across Microsoft (7x increase since March GA); root cause analysis 60% faster; Azure Copilot observability agent saves 30 minutes per alert.

Microsoft shipped major architectural changes to Azure SRE Agent at Ignite 2026. Agents now have their own Microsoft Entra identities, enabling secure cross-cloud and inter-agent communication—agents can invoke other agents' tools using their own permissions. AWS and Google Cloud connectors allow agents to investigate and remediate incidents across multi-cloud environments from a single orchestration point. Teams' integration gives operational agents a channel for alerts, context sharing, and human handoff. Grapher gets a natural-language AI layer for metric exploration and dashboard construction. On-demand SRE Agent in Container Apps lets teams spin up dedicated agent instances for specific investigations. Internally, Microsoft has deployed 11,000+ Azure SRE Agent instances, achieving 60% faster root-cause analysis and cutting per-alert investigation time by 30 minutes via the Copilot observability agent. The new Entra Agent ID and cross-agent communication patterns establish Azure's governance model as agents stop being isolated tools and become coordinated teams.

Read the original at techcommunity.microsoft.com ↗

MCP agent protocol stack reaches completion across 200 repositories

Coddykit · Nov 18, 2026 · Analysis

November 18, 2026 marks completion of AI agent infrastructure build-out; ecosystem shifted from 'can agents work?' to 'can agents ship finished deliverables at scale?'; MCP protocol stack now complete with tools and agent-to-agent communication.

Why it matters Infrastructure is production-ready; skills marketplace becoming standardized way to discover and compose agent expertise; operations teams can now focus on application and deployment, not plumbing.

Coddykit's analysis of 200+ repositories and 30 days of development activity identifies November 18, 2026 as an inflection point: the agent ecosystem completed its infrastructure phase and entered the application phase. Memory engines (108K+ stars), orchestration platforms (111K), and context compressors (76K) established a foundation. Agent management platforms (110K), web scraping engines (93K), and local chat interfaces (60K) now build on that foundation. The protocol stack is complete: MCP for agent-to-tool communication, A2A for agent-to-agent coordination, and standardized observability interfaces. Skills marketplaces consolidating around standardized discovery and composition patterns—the new npm for agents. Desktop environments like Hyprland (40K stars) becoming critical infrastructure for agents needing scriptable window management. For operations teams, this means the question shifts from 'do we have agent frameworks?' to 'which frameworks and patterns fit our topology and governance model?' Infrastructure debates are settling; application patterns are emerging.

Read the original at coddykit.com ↗

Charter builds 1,500 edge sites as LLM traffic grows

Fierce Network · Nov 20, 2026 · Industry news

Large language model traffic on Charter's network rose from 0.02% to 0.09% of downstream and 0.05% to 0.32% of upstream year-over-year, while Charter's roughly 1,500 hub sites sit within 20–40 kilometers of 70 million homes. Partners use them for CDN, video and revenue-generating inference, with GPUs within 10 milliseconds of users vital to physical AI.

Why it matters Charter's high-split plan of symmetrical 5–10 Gbps upstream and downstream is validated by measured LLM traffic growth, directly informing edge compute infrastructure sizing decisions for cable operators.

Charter's Chief Technology and Innovation Officer outlined the operator's strategy for navigating the AI era. While LLM traffic currently accounts for less than 1% of total traffic, the year-over-year growth rate—fourfold in upstream, threefold in downstream—demonstrates accelerating adoption. Charter has retrofitted its existing 1,500 regional hub sites (which serve as network aggregation points) with virtualized platforms delivering 70% less equipment while freeing up 80% more power, space and cooling. This existing distributed infrastructure is now being monetized for edge AI workloads, with partners deploying CDN, video processing, and revenue-generating inference at sites positioned within 10–40 km of end users. The strategic implication is significant: Charter sees its cable HFC advantage not in raw speed but in geographical distribution for deterministic, sub-millisecond latency inference. This validates its fiber-and-cable convergence model and symmetric high-split strategy—avoiding the costly fiber-only build-out required by some fiber-first competitors while leveraging distributed computing at the edge of its existing footprint.

Read the original at fierce-network.com ↗

Paper maps open problems in scale-up accelerator fabrics

arXiv · Nov 19, 2026 · Research

Paper analyzing networking challenges for AI accelerator interconnects. Authors assess topology, protocol mechanisms, and system integration requirements for high-bandwidth, low-latency scale-up fabrics used in distributed training and inference workloads.

Why it matters Scale-up networking directly impacts cluster performance for large AI training jobs; this paper provides architectural foundation for evaluating and designing next-gen interconnects.

Tingting Xu and 11 co-authors from leading cloud/accelerator companies analyze the design space for scale-up fabrics—tightly coupled accelerator networks combining high per-device bandwidth with low, predictable latency. The paper moves beyond vendor-specific implementations to establish a requirements-driven framework for topology, protocol mechanisms, communication semantics, and system integration decisions. Scale-up networking has become a critical bottleneck as model sizes expand; existing solutions (custom ASICs, proprietary protocols) lack standardization. This work provides practitioners with a systematic way to evaluate trade-offs between technologies like optical switching, packet-based switching, and hybrid approaches, and to understand how protocol choices interact with topology constraints. The paper acknowledges that topology, protocol, and system design remain actively debated, making a unified framework valuable for infrastructure teams sizing AI cluster investments.

Read the original at arxiv.org ↗

IETF draft adds post-quantum forward secrecy to EAP-AKA'

IETF · Nov 18, 2026 · Primary source

New IETF draft specifies post-quantum key encapsulation updates to RFC 9678 for EAP-AKA' reauthentication. Addresses forward secrecy and hybrid post-quantum cryptography in mobile and wireless authentication protocols.

Why it matters Mobile and cellular networks need post-quantum-resistant authentication; this draft guides operators on upgrading EAP-AKA' implementations before quantum computers threaten current cryptography.

Draft-ietf-emu-fs-reauth-00 updates forward secrecy provisions in RFC 9678 (EAP-AKA') to incorporate post-quantum key encapsulation mechanisms and hybrid post-quantum approaches. EAP-AKA' is widely deployed in 3GPP mobile networks for user authentication and key establishment. The draft allows systems to combine classical and post-quantum key encapsulation in reauthentication flows, critical because traffic encrypted today with keys derived from current reauthentication could become vulnerable to future quantum attacks. Network operators planning multi-year upgrades must track this work to understand migration timelines and testing requirements for devices in the field that will authenticate through networks for years to come.

Read the original at datatracker.ietf.org ↗

Google launches cheaper mid-tier Gemini 4 Neon model

Google DeepMind · Nov 18, 2026 · Primary source

Gemini 4 Neon launched November 18, 2026 as mid-tier model between Hydrogen and Carbon, delivering near-Carbon reasoning quality with faster task completion and lower cost. Priced at $3 input / $15 output per 1M tokens (25% lower than Carbon) with 1M-token context window and four reasoning levels.

Why it matters Neon generates ~110 output tokens per second versus Carbon's ~57 tokens/sec, and reduced token pricing allows cost-per-task optimization for high-volume enterprise knowledge work. Direct multimodal file input support (Word, PowerPoint, Excel) with page/range slicing reduces API calls for document processing workloads.

Adapted from Gemini 4 Carbon and released November 18, 2026, Neon supports 1,048,576-token context with 65,536-token default output ceiling and four reasoning levels (Low, Medium, High, XHigh). On Google's Multimodal Knowledge Agent evaluation, Neon achieves 56% Climb WSR, matching Claude Fable 5.2 (57%) and exceeding Carbon (54%). On Google's 19-row benchmark chart, Neon outperforms Claude Opus 5.5 on 16 of 18 scored rows. Neon uses ~15% more output tokens per task than Carbon, but lower token prices ($3/$15 vs $4/$20) and higher token throughput (~110 vs ~57 tok/sec) reduce total cost and latency per task. Available via Google API, Gemini Enterprise Agent Platform, and Antigravity/Gemini app for Pro/Ultra subscribers. Supports text, image, audio, video and document input.

Read the original at deepmind.google ↗

Palo Alto Networks buys Applied Compute for $4.5 billion

CSO Online · Nov 19, 2026 · Industry news

Palo Alto Networks agreed to buy Applied Compute in an all-stock deal valued at $4.5 billion to help enterprises adopt AI securely through a post-training and inference platform for open-weight models that keeps proprietary data on-premise.

Why it matters Security vendors consolidating inference and post-training stacks signals enterprise AI adoption moving from chat toward data-sensitive automation; impacts architecture decisions for private model deployment.

Palo Alto announced the $4.5B Applied Compute acquisition on November 18, 2026, targeting enterprises that run mixed open and closed models. Applied Compute's core pitch is to run open-weight models on enterprise data without exposing secrets to public APIs; the deal signals that enterprise AI adoption is moving beyond chat toward specialized, data-sensitive automation—and that security vendors see ownership of the full stack as existential. This acquisition matters for enterprise architecture because it reflects the emerging reality: frontier models from API providers aren't sufficient for sensitive workloads. Organizations now need post-training and inference capabilities that keep proprietary data on-premise or in sovereign clouds. The move by a security vendor signals that this stack will consolidate around established security companies rather than standalone ML platforms.

Read the original at csoonline.com ↗
Nothing in today's briefing matches that.

Listening

Podcasts and talks

Vendor Radar

Last 7 days · arrows compare with the 7 before

What changed this week

Last 7 days vs the 7 before

Biggest moves

Trending topics

agent · MCP · observability · LLM · SRE · automation · inference · Agentic AI · RAG · OpenConfig · AIOps · OpenTelemetry

Get the daily briefing

Today's 3 things that matter and every story with why it matters, in your inbox each morning. Free, and you can unsubscribe at any time. Prefer a reader? Follow the RSS feed.