Welcome to episode 345 of The Cloud Pod, where the forecast is always cloudy! Justin, Ryan, and Matt are in the studio this week and are ready to bring you all the latest in cloud and AI news, including what’s going on between Anthropic, the DOD, and OpenAI, what the war means for Middle East data centers (Spoiler – I hope you have a good Disaster Recovery plan), and Transit Gateway pricing changes that are enough to make a grown man cry. And don’t bother waiting: Matt has completely forgotten almost two years of “bye everybody” and now claims full amnesia as to what his outtro is. Oh well. Let’s get into today’s show.

Titles we almost went with this week

Claude Learned to Use a Computer Better Than Your Dad **OpenAI
Amazon and OpenAI’s $138 Billion AI Bromance
When Two AZs Go Dark the Cloud Gets Crispy
Fifty Billion Reasons AWS Loves OpenAI Now **Anthropic
Azure Still Wins Even When AWS Thinks It Did
Fire, Water, and a Multi-AZ Assumption Goes Up in Smoke
Claude Refuses to Go Full Skynet for the Pentagon
GPT-5.3 Instant Finally Stops Lecturing You
No Killer Robots Without Human Approval Please
Terraform Finally Sees Your Forgotten Cloud Resources
Stage Before You Rage Deploy Azure Firewall
CrowdStrike to Zscaler AWS Wants Your Security Tab
One Hub to Rule Your API Sprawl
Transit Gateway Attachments Just Got Surprisingly Expensive
Azure Container Registry Finally Has Room for Your AI Hoarding
Bedrock Gets a Roommate OpenAI Moves In
Azure Firewall Gets a Safety on the Trigger
Stop Writing Scripts, Just Import the Dang Infrastructure
Audit Your APIs Before March 2026 Bites You
Damn it… my excuse not to DR is gone
I’m Epically Furious about DR

AI Is Going Great – Or How ML Makes Money

03:34 Anthropic acquires Vercept to advance Claude’s computer use capabilities

Anthropic acquired Vercept, a team specializing in AI perception and interaction, to strengthen Claude’s computer use capabilities.
The Vercept founders, including Ross Girshick, bring deep expertise in how AI systems visually interpret and interact with software interfaces.
Claude Sonnet 4.6 shows substantial improvement in computer use benchmarks, jumping from under 15% on the OSWorld evaluation in late 2024 to 72.5% today.
The model is now approaching human-level performance on tasks like navigating spreadsheets and completing multi-tab web forms.
Computer use enables Claude to operate inside live applications the way a human would, handling multi-step workflows across tools that cannot be automated through code alone.
This is relevant for enterprise use cases involving document processing, browser-based workflows, and cross-application task management.
This is Anthropic’s second acquisition in a short period, following the purchase of Bun, which was tied to the Claude Code milestone. The pattern suggests Anthropic is actively acquiring specialized engineering teams rather than just technology assets.
For developers and businesses building agentic workflows on Claude, the improved computer use performance means more reliable automation of complex, real-world software tasks without requiring custom integrations or APIs for every application involved.

05:18 Justin – “It seems like every day I have to update Claude Code because they released a new feature or a new capability.”

12:34 Improving skill-creator: Test, measure, and refine Agent Skills

Anthropic has updated its skill-creator tool for Claude Agent Skills, now available on Claude.ai, Cowork, and as a plugin for Claude Code.
The update brings software development practices like testing, benchmarking, and iterative refinement to skill authoring without requiring users to write code.
The core addition is an eval framework that lets skill authors define test prompts, describe expected outputs, and verify skill behavior across model updates.
A practical example given is the PDF skill fix, where evals isolated a positioning failure on non-fillable forms and guided a targeted fix.
A new benchmark mode tracks eval pass rate, elapsed time, and token usage, and can be integrated into CI systems or local dashboards. Multi-agent parallel eval execution is also included to reduce test time and prevent context bleed between runs.
Comparator agents enable A/B testing between two skill versions or between a skill and no skill, with blind judging to reduce bias in assessing whether a change improves output quality.
Anthropic notes that as base-model capabilities improve, some capability-uptake skills may become unnecessary, and the eval framework is positioned as a step toward skills being defined by natural-language descriptions of desired outcomes rather than detailed implementation instructions.

13:54 Justin – “For things that are actually in pipelines or agentic capabilities where you want things to be specific, this is great.”

14:35 Statement on the comments from Secretary of War Pete Hegseth

Anthropic has publicly refused to allow Claude to be used for mass domestic surveillance of Americans or fully autonomous weapons, citing concerns about current AI reliability and civil liberties.
These two exceptions led to a breakdown in negotiations with the Department of War after months of discussions.
The Department of War is moving to designate Anthropic as a supply chain risk under 10 USC 3252, a designation Anthropic states would be the first time applied to a US adversary. Anthropic has indicated it will challenge any such designation in court.
From a practical standpoint, the legal scope of a supply chain risk designation is narrow. It would only affect the use of Claude on Department of War contract work, leaving commercial API customers, Claude.ai users, and non-DoW contractor use cases completely unaffected.
This situation raises a broader question for cloud and AI vendors about the terms under which they can negotiate acceptable use policies with government customers.
The outcome could set a precedent for how American companies handle government contracts that conflict with their own usage restrictions.
Anthropic notes it has been deployed in US government classified networks since June 2024, making this dispute notable for the AI industry as more frontier model providers pursue federal contracts through programs like FedRAMP and classified cloud environments.

Statement from Dario Amodei on our discussions with the Department of War

Anthropic has publicly refused the Department of War’s requests to remove two specific safeguards from Claude: restrictions on mass domestic surveillance use cases and on fully autonomous weapons systems.
This is notable because Anthropic was already the first frontier AI company to deploy models in US classified networks,

345: Damn It… my excuse is now gone for Disaster Recovery