Today in AI — 30 April 2026

Fifty billion dollars from Amazon, a classified Pentagon contract, and an AI agent that wiped a startup's production database — all in the same news cycle. AI is being pushed into every institutional channel simultaneously, and the governance can't keep up.

The end of lock-in

GPT-5.5 landed on Amazon Bedrock within a day of Microsoft's exclusivity expiring, and Harvard is dropping ChatGPT Edu for Claude. When a hyperscaler and a university both switch providers in the same week, commoditisation is no longer theoretical.

• OpenAI models land on AWS Bedrock one day after Microsoft exclusivity ends — GeekWire
• Harvard drops ChatGPT Edu and adds Anthropic's Claude for faculty and students — The Harvard Crimson

AI in the room where it happens

Google signed a classified Pentagon AI deal over 700+ employee objections, Amazon launched AI that interviews candidates with no human in the loop, and GM is pushing Gemini into 4 million vehicles via OTA update. Same bet across all three: AI belongs in high-stakes decisions right now.

• Google signs classified Pentagon AI deal despite 700-employee protest — NBC News
• Amazon launches AI that conducts job interviews 24/7 with no human involvement — GeekWire
• GM brings Google Gemini to 4 million vehicles, replacing Google Assistant — The Next Web

Capital in, people out

Meta is cutting 8,000 jobs while guiding capex to $135 billion for AI compute. Rogo raised $160 million for AI financial analysts, and the Musk-OpenAI trial continues airing the industry's dirtiest laundry.

• Meta confirms 8,000 layoffs starting May 20 as AI spending soars to $135 billion — Variety
• Rogo raises $160M to scale agentic AI for Wall Street — SiliconANGLE
• Musk accuses OpenAI leaders of 'looting the nonprofit' in dramatic trial testimony — NPR

Trust at production speed

A Cursor AI agent deleted a startup's production database in nine seconds, the Zig project published the strongest case yet for banning AI contributions from open source, and Meta open-sourced Llama Guard 4, LlamaFirewall, and Prompt Guard 2 at LlamaCon. Agent safety tooling is arriving, but so are the failures it needs to catch.

• Cursor AI agent deletes startup's entire production database and backups in 9 seconds — The Register
• Zig project articulates the strongest case yet for banning AI from open source contributions — Simon Willison
• Meta releases Llama Guard 4, LlamaFirewall, and Prompt Guard 2 at first LlamaCon — Meta AI

The builder's stack

Nvidia's Nemotron 3 Nano Omni delivers multimodal on a single GPU, activating only 3 billion of its 30 billion parameters per token. Mistral's Workflows is already running millions of daily executions for enterprise customers.

• Nvidia releases Nemotron 3 Nano Omni, an open multimodal model that runs on a single GPU — NVIDIA Blog
• Mistral AI launches Workflows, a production-grade orchestration engine for enterprise AI — VentureBeat

Nine seconds is all it takes to remind everyone that deployment speed and deployment readiness are not the same thing.

---