Archived edition 21 November 2026 · Week 47

KI PowerUpdate

Your weekly AI briefing. From hype to need.

Go to the latest edition →

The week in two and a half minutes, read by a synthetic voice.

Duration: 2:21 min Download MP3 Podcast feed Read the script

01🧠 AI News of the Week

TechCrunch

Anthropic IPO raises nearly $105 billion in new capital

Anthropic priced its IPO at $83 per share, raising close to $105 billion, with each of the seven founders receiving about $280 million. A leading AI lab now answers to the quarterly scrutiny of public markets. Customers and partners will watch whether pricing and development pace shift under shareholder expectations.

Read the original article →

t3n

Goldman Sachs restricts AI tools as costs climb

Goldman Sachs is tightening access to AI tools, reserving Claude Code and GitHub for developers only. The move suggests that the cost of company-wide AI use is reaching limits even at large banks. Firms should allocate licences by expected value rather than rolling them out broadly.

Read the original article →

02🚀 Tools & Startups

03⚖️ Regulation & Ethics

04🎙️ Voices & Perspectives

05🏭 Business & Society

📄 Read the script

What does it mean when a record-setting IPO is followed one day later by a global sell-off? Anthropic raised almost 105 billion dollars at 83 dollars a share, and each of its seven founders took home roughly 280 million. By the second trading day the stock was down ten percent, the Nasdaq was selling off, and the S&P 500 had slipped into correction.

The market is asking an uncomfortable question: who carries the weight of enormous compute commitments? Anthropic has taken them on, while partners such as Amazon and Google could keep profiting even if Anthropic loses its lead. The industry's financing model is circular, and an IPO makes that circularity visible. Dario Amodei promised at the opening bell to pace development regardless of pressure. That is a sentence that only becomes worth something when the share price is under strain.

How expensive company-wide AI use really is shows up in the daily life of a large bank. Goldman Sachs now limits Claude Code and GitHub to developers. If even a bank with that financial strength rations access, licensing turns into a governance question: who gets which tool, and for what expected return?

The cost side has a physical counterpart. A new Harvard study sharpens the debate over data centre electricity and water use in the United States, where resistance is growing. Politics responds with contradictions. President Trump wants data centres at home but threatens companies that build abroad with enormous penalties. Google and Blue Origin, meanwhile, are thinking about data centres in orbit, which sounds less like a solution than evidence of how scarce power and land have become on the ground.

As the cost question grows, so does the control question. Google will let Gemini operate Macs, organising files, filling out forms and running multistep workflows. The agent moves from the browser into the operating system. That is convenient until you ask what happens when it is wrong. Meta's Muse agent reportedly mixed up conversations and showed other people's account balances, and GitLab now logs agent actions in its audit log. These are two sides of one coin: autonomy requires traceability.

Google itself has built an artificial internet with thousands of websites to expose agents to realistic attacks. Anyone deploying agents in production should test them in a similar spirit, because the question of how an error gets corrected before it causes harm cannot be answered after rollout.

That capabilities still impress is clear from mathematics. An AI-assisted result on the Riemann hypothesis is occupying the field, and a human approach exists alongside it. Verification is under way. If it holds, it would be more than a record; it would signal that models can contribute substantively to fundamental research.

Which leaves the question of rules. UN talks on autonomous weapons ended without a binding commitment, even though many countries wanted limits. The technology moves, regulation waits, and between them sit companies that must plan under uncertainty.

The week in one line: AI is becoming more expensive, more autonomous and more political at the same time, and anyone deploying it has to decide on cost, control and location together.

Sources in this edition: Handelsblatt, TechCrunch, heise, t3n