This was the week the hyperscalers flexed their in-house muscle. Microsoft Build and Computex dropped within hours of each other, and the message from both was the same: the era of relying solely on third-party models is over. Microsoft unveiled seven homegrown MAI models, NVIDIA released the most intelligent US open-weights model to date, and Anthropic quietly filed its IPO paperwork at a valuation that would have been unthinkable 18 months ago. On the operations side, Google, Microsoft, and AWS all shipped agentic SRE tools in the same week – the race to automate incident response is no longer theoretical.

🤖 AI & Machine Learning

Microsoft launches seven in-house MAI models at Build 2026 – Microsoft announced the MAI model family: seven models spanning reasoning, coding, image generation, transcription, and voice – all developed in-house without distillation from third-party models. The standout is MAI-Thinking-1, a reasoning model that matches leading models in its weight class on software engineering benchmarks and is preferred over Sonnet 4.6 in blind human evaluations. MAI-Code-1-Flash, with just 5 billion active parameters, scores 51% on SWE Bench Pro and is rolling out as a default model in VS Code and GitHub Copilot. MAI-Transcribe-1.5 claims SOTA accuracy across 43 languages. This is Microsoft’s clearest signal yet that it’s building its own model stack, not just reselling OpenAI. Source

NVIDIA releases Nemotron 3 Ultra – the most intelligent US open-weights model – Announced during Jensen Huang’s Computex keynote and published on June 4, Nemotron 3 Ultra is a 550B-parameter Mixture-of-Experts model with 55B active parameters, purpose-built for orchestrating long-running agent workflows. It uses a hybrid Mamba-Transformer architecture for efficient million-token context handling, NVFP4 quantization for up to 5x higher throughput, and multi-teacher on-policy distillation from over ten domain-specific teachers. It scores 48 on the Artificial Analysis Intelligence Index – well ahead of Gemma and gpt-oss-120b, though still behind Chinese open-weights leaders. Fully open: weights, data, recipes, and available everywhere from Hugging Face to NIM to AWS JumpStart. Source

Anthropic confidentially files for IPO at nearly $1 trillion valuation – Anthropic submitted a draft S-1 to the SEC on June 1, setting up what could be the most significant AI IPO to date. The company last raised $65 billion at a post-money valuation of $965 billion – ahead of OpenAI – and reports annualized revenue of $47 billion. Anthropic is getting ahead of OpenAI, which is also preparing its own confidential filing. This is no longer a startup story; it’s an infrastructure-of-the-economy story. The confidential filing doesn’t lock in a timeline, but the IPO market is watching closely. Source

Anthropic ships Claude Opus 4.8 with Dynamic Workflows – Anthropic released Claude Opus 4.8 on May 28, alongside Dynamic Workflows – a new Claude Code feature that lets the model orchestrate hundreds of parallel subagents in a single session. The system dynamically plans work, fans it across subagents, cross-checks results, and converges on a verified output before reporting back. Anthropic says Opus 4.8 with Dynamic Workflows can carry out codebase-scale migrations across hundreds of thousands of lines of code, from kickoff to merge, using the existing test suite as its acceptance bar. Also shipping: effort control in claude.ai (choose how hard Claude thinks), and a 3x price reduction on fast mode. Source

🔧 Site Reliability Engineering

Google publishes “AI in SRE” whitepaper – moving beyond automation to agentic operations – Google SRE released a detailed whitepaper titled “AI in SRE Practice: Moving Beyond Automation at Google,” detailing how the team that invented SRE is integrating agentic AI across the entire service lifecycle. The paper goes far beyond RCA and troubleshooting. It covers autonomous mitigation agents (AI Operator) that handle complex on-call responsibilities, strict execution guardrails (Actus) for governing agent actions, and continuous evaluation pipelines (IRM Analyzer) grounded in human operational memory. The key insight: SRE is no longer just automating runbooks – it’s building autonomous control planes for high-velocity software development. The whitepaper is publicly available and worth reading in full. Source

Azure SRE Agent gets enterprise features at Microsoft Build 2026 – Since going GA in March, Azure SRE Agent has been diagnosing live issues and automating response workflows in production. At Build 2026, Microsoft announced five releases aimed at enterprise-scale adoption: VNet integration (run the agent inside private networks), Managed Connectors using the Model Context Protocol (MCP) for governing connections to Jira, GitLab, Slack, PagerDuty, and more, a granular permissions model with allow/ask/deny rules on individual tools, native GitHub Enterprise support, and expanded SaaS integrations. The MCP-based connector architecture is notable – every tool the agent touches goes through the same wire format, and the trust boundary between “what the model decided” and “what was actually sent” is explicit and auditable. Source

AWS launches next-gen Resilience Hub with GenAI-powered failure mode analysis – AWS announced the general availability of the next generation of Resilience Hub on May 28, replacing static rule-based resilience checks with generative AI-powered failure mode assessments. The new version automatically discovers dependencies (including cross-region calls and third-party endpoints you didn’t know about), evaluates your architecture against AWS Well-Architected best practices, and generates actionable failure mode findings specific to your setup. It integrates with AWS Organizations for portfolio-wide reporting, so platform teams can set resilience policies centrally and track compliance across hundreds of applications. For SRE teams that have been manually auditing blast radius and failover coverage, this is a meaningful step toward automated resilience posture management. Source

🔭 Observability

OpenTelemetry launches “Blueprints” initiative for enterprise adoption – OpenTelemetry introduced Blueprints, a new initiative providing prescriptive, opinionated deployment guidance for common observability scenarios. Each blueprint addresses a specific operational challenge – SDK configuration, Collector deployment patterns, context propagation, semantic conventions – with architecture patterns, best practices, and actionable implementation steps bundled together. Organizations like Adobe, Mastodon, and Skyscanner have already contributed reference implementations showing how they deployed OTel at scale. The project is responding to mounting feedback that OTel’s flexibility has become its own adoption barrier: teams want proven patterns, not infinite configuration options. Source

Azure SRE Agent’s Managed Connectors bring MCP-governed observability tool integration – One of the more quietly significant announcements at Build 2026: Azure SRE Agent’s new Managed Connectors use the Model Context Protocol (MCP) to provide a unified, governed interface between AI agents and observability tools. The connector catalog now includes Datadog, PagerDuty, ServiceNow, Splunk, and more – each exposed to the agent as curated MCP operations with explicit permission boundaries. Admins select which operations the agent can see, pin parameter defaults, and require human approval on sensitive actions. The agent never sees tools you didn’t expose and can’t bypass your approval rules. This is a practical blueprint for how AI agents should interact with production observability systems: transparent, auditable, and admin-controlled. Source

OpenTelemetry Profiles reaches public alpha – profiling becomes the fourth signal – The OpenTelemetry Profiling SIG officially moved the Profiles signal to public alpha, establishing profiling as a first-class signal alongside traces, metrics, and logs. The alpha includes the OTel eBPF Profiler, a production-scale, cross-language profiler that observes applications without code changes. Contributors from Google, Datadog, Elastic, Grafana Labs, and Splunk are all backing the standard. For years, observability answered what happened, where, and how often. Profiling adds why – continuous CPU and memory profiling correlated with your existing telemetry pipeline. The alpha is explicitly not production-ready yet, but the vendor alignment behind it suggests this will move fast. Source

🔗 Quick Links

  • NVIDIA Vera Rubin in full production – Jensen Huang declared the next-gen multi-rack pod-scale AI supercomputer is shipping, along with RTX Spark, an Arm-based Windows PC with a Grace CPU and Blackwell GPU. The Next Web
  • Anthropic expands Project Glasswing to 150+ orgs in 15+ countries – Claude Mythos, Anthropic’s most powerful model, is now scanning critical infrastructure codebases for zero-day vulnerabilities across power, water, healthcare, and communications sectors. TechCrunch
  • Holo3.1 computer-use models ship with quantized weights – Four sizes from 0.8B to 35B, with FP8/NVFP4/Q4 GGUF checkpoints for local inference. First release optimized for edge deployment. Hugging Face
  • OpenAI models now GA on Amazon Bedrock – GPT-5.5, GPT-5.4, and Codex are generally available with IAM, PrivateLink, and CloudTrail integration. AWS Blog
  • Anthropic publishes agent containment engineering post – How Anthropic caps blast radius across claude.ai, Claude Code, and Cowork as agents grow more capable. Anthropic
  • AI agents generating untracked chaos engineering failures – VentureBeat reports 79% of orgs have agents in production, but most lack incident classification for autonomous agent-initiated cascades. VentureBeat

💬 My Take

The through-line this week is convergence on a single question: who controls the agents? Not “should we build agents” – that ship sailed months ago. The question now is governance at scale.

Google’s SRE AI whitepaper lays out the most mature framework I’ve seen: autonomous mitigation agents with explicit guardrails and continuous evaluation pipelines. Microsoft’s Azure SRE Agent takes a different but complementary approach with MCP-based connectors that make every tool interaction auditable and admin-governed. AWS Resilience Hub automates the failure mode analysis that SRE teams have been doing manually with spreadsheets and tribal knowledge. All three arrived in the same seven-day window.

Meanwhile, the model layer keeps compressing. Microsoft now has its own reasoning model that competes with Sonnet. NVIDIA’s open-weights model matches frontier intelligence in a package anyone can deploy. Anthropic is filing for an IPO while simultaneously giving Claude Code the ability to orchestrate hundreds of parallel agents autonomously. The gap between “impressive demo” and “production system that makes unsupervised decisions” is narrowing faster than governance frameworks can keep up. The VentureBeat piece on untracked agent-initiated failures is the canary – 79% of organizations have agents in production, but almost none classify agent actions as potential incident causes. That’s the gap that will generate the next wave of high-profile postmortems.

Thanks for reading this week’s roundup. If something here caught your eye or I missed a story you think deserves attention, I’d love to hear about it – reach out on LinkedIn. See you next week.