Suprvisr TLDR - May 4 2026

Cumulus now lets you pull colleagues directly into a conversation with your agent, so the model's output can be reviewed, redirected, or signed off on by the right person before it leaves the chat window. Useful when a draft email needs your manager's eyes, when a procurement decision needs finance to weigh in, or when an agent is about to do something that, after a week like this one, you would rather have a second person check first.

It works the way your team already works. Tag a colleague, they drop into the same thread, the context comes with them, the model picks up where it left off. No screenshots, no copy-paste, just native team collaboration in a single pane of glass.

(no credit card required)

On Monday, Microsoft and OpenAI announced an amended agreement that ends the exclusivity at the heart of their partnership. Microsoft's license to OpenAI models is now non-exclusive through 2032. OpenAI can serve its products on any cloud, not just Azure. Microsoft stops paying revenue share to OpenAI on resold products, while OpenAI's payments to Microsoft continue through 2030 subject to a cap. The AGI clause that previously could have ended Microsoft's license is gone.

This is the most significant restructuring of the relationship since Microsoft's first $1 billion investment in 2019. Within hours, AWS confirmed it would offer OpenAI models through Bedrock. The implication for buyers is straightforward. If you have been locked into Azure to access GPT, you now have leverage. If you are a Microsoft shop, your AI strategy is no longer determined by your cloud strategy.

That is how long it took an AI coding agent (Cursor running Anthropic's Claude Opus 4.6) to delete the entire production database, and all volume-level backups, of PocketOS, a SaaS company that powers operations for car rental businesses. The agent hit a credential mismatch in a staging environment, decided on its own to "fix" it, found an unrelated API token sitting in another file, and used it to call Railway's volume-delete endpoint. No confirmation prompt, no human in the loop. PocketOS's most recent usable backup was three months old. The founder's customers spent the weekend rebuilding bookings from Stripe receipts and email confirmations. The lesson is not that AI is dangerous. The lesson is that an over-scoped API token plus an autonomous agent equals the fastest disaster you will ever own.

On Wednesday, Alphabet, Microsoft, Amazon, and Meta all reported quarterly earnings within minutes of each other, and all four raised their 2026 capital expenditure guidance. Wall Street analysts now project combined AI infrastructure spending across the four to top $1 trillion in 2027. Alphabet led the night with Google Cloud revenue up 63%, while Meta dropped nearly 9% the next day after raising its capex range to as much as $145 billion. Microsoft's AI revenue is growing 123% year over year. AWS came in at 28% growth.

This was the quarter the market stopped giving AI spending a free pass. Alphabet showed where its capex was producing revenue and the stock climbed. Meta raised its capex without that proof and the stock fell almost 9%. Expect every earnings call this quarter to face the same question.

For most leaders, the implications are simpler than they look. If you sell anything AI infrastructure consumes (chips, power, cooling, real estate, water), you have more potential customers than you did a week ago. If you spend on AI internally, the standard for justifying that spend just got stricter. The bar is now a real number.

1. The largest seed round in European history is a bet against LLMs. David Silver, the DeepMind researcher who built AlphaGo and AlphaZero, raised $1.1 billion on Monday at a $5.1 billion valuation for Ineffable Intelligence, a new lab pursuing AI that learns from experience instead of human-written text. Sequoia, Lightspeed, Nvidia, Google, and the UK's Sovereign AI Fund all wrote checks. According to its newly launched site, Ineffable aims to create a "superlearner" capable of discovering knowledge and skills without relying on human data by leveraging reinforcement learning; a technique in which AI systems learn through trial and error rather than studying human-generated examples.

2. Cohere bought Germany's Aleph Alpha to build a sovereign AI alternative 🇨🇦 Toronto-based Cohere announced a merger with Germany's Aleph Alpha valued at roughly US$20 billion, with Schwarz Group leading a 500 million euro financing round. The new entity stays headquartered in Toronto and is being positioned as a Canadian-German sovereign AI option for regulated industries. If you operate in finance, healthcare, energy, or government work in Canada or the EU, your AI shortlist just got a credible non-American option on it. (Besides us of course!)

3. The Mercor breach makes voice authentication officially indefensible. The extortion group Lapsus$ leaked roughly 4 terabytes of data from AI contractor platform Mercor, including studio-quality voice recordings paired with government IDs for over 40,000 people. Off-the-shelf voice cloning needs about 15 seconds of clean audio. The Mercor recordings average two to five minutes. If your customer authentication, your call centre, or your bank involves a voiceprint as a primary factor, treat it as compromised and add a second factor this quarter.

Use the permission tools your platforms already give you. AI agents do not need access to production API keys, period. Build structural guardrails around destructive actions so the agent physically cannot execute them, even if it tries. Rotate any key an AI has seen, since you have to assume it's been logged or leaked somewhere. And the test for whether you're done: sit down with your team, ask "what is the worst thing this agent could do right now," and make them prove the answer is impossible. If they can't, you have your work for the week. Only you can stop forest fires AI deleting your database.