Big 5 AI Vendor Roundup: Week of November 9, 2026

Technology Note By: Mark Tauschek, Bill Wong, Infotech

Special edition: Resilience in an increasingly interconnected world

Four unrelated events made resilience the week's defining issue. If you’re thinking that resilience is becoming every week’s defining issue, it is, which is why we cover it in every weekly roundup. This week was particularly wild though, as Nvidia lost its founder and CEO after a highly publicized personal and national security controversy. A worldwide Microsoft 365 failure disabled Outlook, Teams, Copilot, and other services for more than a day. The FAA imposed a nationwide ground stop after its air traffic control communications became unstable, and a tanker struck a docked cruise vessel in New Orleans. At the same time, AI vendors kept releasing more capable models and extending agents into payments, business systems, and software development.

There is no evidence that AI caused the Microsoft outage, the FAA interruption, or the maritime collision. Speculation about AI hacking and AI-discovered vulnerability patches moved faster than verified facts, and that’s why we’re including those stories this week. The enterprise lesson is broader and more useful: organizations must be able to absorb failures they didn't predict, keep critical work going without a dominant supplier, and distinguish an actual technical cause from a plausible sounding narrative.

Four shocks put resilience in focus

  • Nvidia CEO Jensen Huang resigns after an admitted affair and additional allegations. Nvidia announced on November 12 that Huang had resigned as CEO, president, and director, effective immediately, and named CFO Colette Kress interim CEO. Reporting by WIRED connects the departure to Huang's admitted infidelity involving a woman US authorities suspect of Chinese intelligence ties, and a reported relationship with a direct subordinate. The espionage suspicions and subordinate allegation have not been established as findings of wrongdoing. Nvidia's board is investigating, and its SEC filing confirms the transition without assigning a cause. Nvidia shares fell 4.3% on November 12, according to The Guardian, and 23.8% since the October 29 New York Times report about the alleged affair with a woman who is a suspected Chinese spy. The event creates a sudden leadership and governance risk at the company on which much of the AI market relies. Enterprise buyers should review critical supplier succession, contractual continuity, supply allocation, and concentration, not speculate about personal conduct.
  • A historic Microsoft outage disrupts work and Copilot worldwide. Beginning November 12, a major Microsoft 365 incident affected Exchange Online and Outlook, Teams, SharePoint, Microsoft 365 Copilot, and related services across regions. Some government and business users couldn't send email, hold meetings, or access normal workflows. Microsoft declared broad restoration on November 14 after more than 24 hours. Its service-health updates for incident MO1550821 say an issue introduced during an earlier network capacity upgrade surfaced after a subsequent network change. Overloaded or unresponsive servers and large email queues prolonged recovery. Microsoft says neither a cyberattack nor AI caused the incident. The impact spread across products because they depended on shared infrastructure, and external assistants using Microsoft 365 connectors were also disrupted. A second AI provider doesn't solve a problem when both need the same underlying email, identity, and collaboration services.
  • FAA orders nationwide ground stop amid air traffic control network instability. On November 13, unstable radar and communications information prompted an exceptional nationwide halt to departures for roughly five hours. Transportation officials said aging routers and switching equipment were overloaded during modernization work, and they reported no evidence of a cyberattack or connection to the Microsoft outage. CBS News reported that a network update appeared to overwhelm older equipment. Early online speculation connected the incident to AI, hacking, or vulnerability patches, but no credible finding supports those explanations. The preliminary facts point instead to a familiar failure mode: changes to interconnected legacy systems can exceed the capacity of components that remain in service.
  • A tanker strikes a docked cruise vessel in New Orleans. On November 13, the Atlantic Asphalt struck the American Heritage on the Mississippi River, causing the passenger vessel to capsize and partially sink. Everyone aboard was accounted for, with dozens treated for injuries. The US Coast Guard said investigators were looking at a possible loss of steering, while the National Transportation Safety Board opened its own investigation. Speculation about an AI or cyber cause is unsupported. This is not an AI incident on current evidence, but it belongs in a resilience note because physical operations still need independent safety barriers, trained people, emergency procedures, and the ability to respond when automation or machinery fails.

OpenAI puts shutdown controls outside the agent and enters payments

  • Control Room can stop frontier research workloads independently of the agent. On November 9, OpenAI detailed a centralized system that tracks research inference jobs and denies further model output after a stop decision. OpenAI says the mechanism stops output within 35 seconds, even if the agent's own code keeps running, and is designed to fail closed if the control service goes down. Its current scope is research infrastructure, not a customer-facing enterprise kill switch. The design offers an important pattern: critical controls must run outside the workload they govern, with no way for an agent to disable or bypass them.
  • ChatGPT Wallet lets agents buy with explicit user authorization. Launched November 13, the early rollout lets eligible US Pro subscribers authorize a dot to make a purchase using a payment token rather than exposing the actual card number. Users must review and approve each purchase, including a biometric check. OpenAI is working with card networks and merchants on acceptance. It is not yet an unrestricted spending agent, but enterprise IT and procurement teams should expect similar delegated purchases and require verified identity, transaction limits, approval evidence, and refund and dispute handling.
  • Businesses can claim and manage how they appear in ChatGPT. Business Hub, announced November 10 in beta, lets verified businesses update information, review engagement, and connect product feeds. Eligible US businesses can publish customer-facing business agents. This adds a new customer channel and another dependency on a platform's identity verification, distribution, accuracy, and availability. Organizations should designate an owner for AI channel information and maintain an independent customer contact path.
  • ChatGPT Enterprise arrives in AWS Marketplace. Starting November 9, eligible customers can procure ChatGPT Enterprise through AWS and apply eligible spending toward existing cloud commitments. The application retains OpenAI's features while procurement and billing run through AWS. The convenience is real, but buyers should document support responsibility, termination rights, data processing terms, and how they would operate if either partner had a prolonged outage.

Anthropic releases Fable 5.2 and begins its IPO roadshow

  • Claude Fable 5.2 and Mythos 5.2 raise technical capability with different safeguards. Released November 9, the two models share an underlying system, but Fable applies stronger cyber and biology restrictions, while Mythos remains available to vetted organizations. Anthropic reports up to 30% faster generation than the 5.1 versions and pricing of $8 per million input tokens and $40 per million output tokens. Its system card describes lower rates of sandbox boundary violations in several tests, though it acknowledges that no evaluation suite can catch every failure. Fable is available across Anthropic's cloud partners; enterprises should assess safety, model routing, cost per accepted task, and authorization for restricted tools separately.
  • Anthropic begins investor marketing for its proposed IPO. On November 10, Anthropic announced a proposed offering of 825 million shares at $70 to $78 each, implying gross offering proceeds of approximately $57.8 billion to $64.4 billion before any additional allotment and fees. The company intends to list as ANT on Nasdaq. The offering remains subject to pricing and completion. Its transition toward public ownership makes revenue quality, cloud concentration, large infrastructure obligations, governance, and operational continuity increasingly visible to enterprise buyers.

Microsoft expands agent permissions while its own services struggle

  • Copilot connectors move from reading information to changing it. Microsoft announced write capabilities on November 9 that let eligible Copilot connectors create, update, and delete content in third-party applications. Write actions require user approval by default, but users can choose an option to allow a tool without repeated confirmation. Administrators can review which tools provide read or write access. The move increases productivity but also makes connector permissions, audit records, recovery, and protection against repeated actions material governance decisions.
  • Microsoft adds Fable 5.2 to Copilot and widens model choice. The November 11 update gives eligible users another frontier model inside Microsoft's environment. GitHub Copilot also added Claude Fable 5.2 and Google Gemini 4 variants, while announcing that six older models will be retired on December 14. GitHub's deprecation notice shows why buyers need pinned configurations and routine regression tests: the platform can change its underlying models even when the Copilot product name stays the same.
  • GitHub Copilot app supports GitHub Enterprise Server repositories. Support became generally available November 13 for connecting to remote repositories, issues, and pull requests in GHES. Cloud sessions aren't available for GHES. This improves reach for enterprises with self-managed source code but requires separate checks of account linking, repository permissions, where code is processed, and whether the required functions still work during a Microsoft or GitHub disruption.

Google releases Gemini 4 Hydrogen and reusable Workspace skills

  • Gemini 4 Hydrogen targets faster coding and everyday agent work. Announced November 12, Hydrogen is available through the Gemini API, AI Studio, Gemini for business, and consumer products at $1.50 per million input tokens and $7.50 per million output tokens. Google introduced Orchestrator in Antigravity to switch between Hydrogen and its larger Carbon model during a task. The company's speed, accuracy, and safety results are vendor tests. Buyers should measure the whole workflow, including automatic routing, retries, security restrictions, and recovery after a failed tool call.
  • Gemini 4 Carbon and Agent Gateway move into more enterprise workflows. On November 10, Carbon became generally available in Gemini Enterprise apps and workflow tools, with enterprise application access off by default. On November 12, Agent Gateway for GKE entered preview to centralize traffic and policy for agents running on Kubernetes. These are meaningful controls, but default settings, identity, cross-agent communication, and logs must be tested before enabling broader autonomy.
  • Workspace skills roll out for shared, repeatable AI tasks. Google began rolling out Markdown-based skills through Workspace Studio and apps on November 10, allowing teams to package instructions, templates, and reference material. Skills can improve consistency but also spread incorrect instructions or excess access. IT leaders should review ownership, change control, permission scopes, and testing before skills are shared widely.

AWS emphasizes recovery, isolation, and multiple access paths

  • AgentCore's next runtime becomes the default for new agents. Starting November 12, new AgentCore runtimes use an architecture that restores initialized agents from snapshots, adjusts resources to demand, and scales idle sessions to zero. AWS reports substantially faster startup and lower billed memory for some workloads. Sessions remain isolated in Firecracker microVMs. Existing runtimes don't move automatically. Buyers should test recovery of task state, restart behavior, credential isolation, and whether a stopped agent can repeat external transactions.
  • Secrets Manager now replicates secrets across AWS accounts. The November 13 release creates encrypted replicas in separate accounts that can be promoted during recovery. This helps isolate a backup access path from a compromised primary account, although both copies still depend on AWS. Organizations should test failover permissions, key availability, revocation, and the ability to restore operations when central administration is unavailable.
  • AWS adds direct web search and more restricted cyber model access. The November 10 Web Search API lets applications request ranked results from AWS's web index, with domain and date filters and IAM controls. On the same day, Bedrock added standard runtime access for eligible Daybreak cyber models. These services reduce integration work, but concentration in one provider's identity, network, and runtime layers can undermine multi-model resilience.

Our Take

This week shouldn't be described as an AI-driven series of disasters. The confirmed causes and ongoing investigations don't support that. The common factor is dependence: a sudden change at Nvidia puts a critical supplier's governance under scrutiny; Microsoft shows how many workflows can fail through shared infrastructure; the FAA demonstrates how new systems can overload old ones; and the Louisiana collision shows the value of emergency procedures when physical control may be lost. False claims about AI hacking are another operational problem because they can distract response teams from the failure actually in front of them.

More capable AI will add dependencies faster. Agent wallets, connectors with write access, reusable skills, cloud runtimes, and automated model selection all make work easier but create more ways for systems to affect one another. Microsoft's outage is especially instructive: switching Copilot to a different model would not restore Outlook, Teams, or the underlying connectors. Resilience must cover people, suppliers, identity, networks, applications, task state, and manual operations, not simply model failover.

What IT leaders should be doing

  • Map shared failure points before adding another AI service. Identify identity providers, DNS, email, collaboration, cloud regions, gateways, model hosts, data stores, connectors, and payment systems behind each critical workflow. An alternative model that uses the same failed services isn't a true backup.
  • Maintain an independent communications and productivity path. Provide emergency contact channels and status pages outside the primary Microsoft 365 tenant. Prepare approved alternatives for email, collaboration, development, search, and customer support so users can keep working during extended outages.
  • Test controlled degradation and restart. Queue work, preserve drafts and agent state outside the model provider, cap retries, use circuit breakers, and prevent duplicate purchases or production changes when an agent resumes after failure. Practice recovery with the actual business workflow, not just a health-check endpoint.
  • Treat technology changes as resilience events. Test modernization projects against legacy capacity, stage rollouts, define rollback triggers, and validate both old and new components under realistic peak loads. Never assume that installing a faster component makes the full system more stable.
  • Keep independent safety and shutdown controls. Agents that can change files, send messages, run code, or spend money need named owners, scoped credentials, approval gates, immutable action logs, and an externally enforced stop mechanism. Test what happens if the control service itself fails.
  • Prepare for supplier and leadership disruption. Review concentration in Nvidia, cloud providers, model partners, and critical integration firms. Require documented continuity and escalation processes, assess alternatives ahead of time, and avoid critical roadmaps that depend on one executive or supplier relationship.
  • Separate incident facts from theories. Use incident command, clear evidence thresholds, named spokespeople, and dated updates. State what is known, unknown, and ruled out. Don't allow unverified claims of AI compromise to drive operational decisions.

Want to Know More?

Latest Technology Notes

All Technology Notes