Claude's privacy leak, AI power crunch, and a self-improving startup bets $410M on compute
Anthropic's shared chats ended up on Google, the US grid is threatening data center blackouts, and a superintelligence startup just dropped $410M on AWS.
Subscribe free →Your Claude Shared Chats Were Indexable by Google
Anthropic's "share chat" feature — which lets users generate a public link to any conversation or Artifact — turned out to be a little too public. Those links were getting crawled and indexed by Google, meaning sensitive prompts and outputs could surface in search results. Anthropic appears to have quietly addressed the issue, but it's unclear how long the exposure window was open.
Why it matters: It's a textbook reminder that "share via link" and "private" are not the same thing — and enterprises dumping sensitive workflows into Claude should audit what their teams have been sharing.
The US Grid Is About to Start Cutting Power to Data Centers
PJM Interconnection — the largest grid operator in the US — is moving toward a policy that would let it temporarily cut power to data centers during peak demand to prevent wider blackouts. The data center construction boom has outpaced the grid's ability to generate and distribute power, and operators are out of easy options.
Why it matters: This is the infrastructure ceiling the AI industry has been racing toward — if compute capacity can be switched off by a grid operator, the "infinite scale" narrative hits a hard physical wall.
Recursive Superintelligence Signs $410M Compute Deal With Amazon
Recursive Superintelligence — a startup building self-improving AI systems — has committed $410M to AWS for compute. The company's entire thesis is that AI should automate its own development, so rather than hiring engineers, it's essentially buying horsepower and pointing it at itself.
Why it matters: This is what an AI-native company actually looks like in 2026: near-zero headcount, nine-figure cloud bills, and a bet that recursive self-improvement eventually compounds faster than human R&D.
Dario Amodei: Fine With Open Weights, Very Not Fine With China
In a public response to ongoing debates about open-source AI, Anthropic CEO Dario Amodei clarified he doesn't oppose open-weight models in principle. His actual concern is China's rapidly advancing AI capabilities — and the risk that open models accelerate adversaries more than they benefit the open-source community.
Why it matters: Amodei is drawing a distinction that most of the open-vs-closed debate ignores: the geopolitical asymmetry of who benefits from freely available frontier models.
Satya Nadella: Single-Model AI Dependency Is a Business Risk
Microsoft's CEO warned that companies routing everything through one AI provider — without their own models or an AI gateway layer — are making a strategic mistake. An AI gateway sits between your application and the underlying model, abstracting away vendor lock-in and keeping your prompts out of the model provider's hands.
Why it matters: Coming from the CEO of a company that sells both Azure AI and Copilot, this is notable self-aware advice — and a subtle pitch for Microsoft's own AI infrastructure stack.
Hugging Face Is a Deepfake Factory and Isn't Doing Much About It
A report from European nonprofit AI Forensics found that Hugging Face is hosting models specifically designed to generate nonconsensual sexual deepfakes of women and children. The models are easy to find, easy to run, and the platform's enforcement is minimal at best.
Why it matters: Hugging Face has positioned itself as the GitHub of AI — but GitHub has content policies with teeth; the deepfake problem here is a direct test of whether the open model hosting ecosystem can self-regulate before regulators force the issue.
Quick Hits
- →Cursor, now the third-biggest AI coding tool market globally in India, is rolling out localized pricing ahead of its pending SpaceX acquisition — a rare combo of aggressive growth and looming ownership change. TechCrunch
- →Fish Audio raised a $52M seed round with $21M ARR and 8M users — AI voice cloning is quietly becoming a real business. TechCrunch
- →Perplexity's Personal Computer agentic tool expanded to Windows, letting the AI control your local apps and files. The Verge
- →Lyft and Baidu's Apollo Go are starting robotaxi tests in London via the Freenow network — the autonomous vehicle battle is going transatlantic. TechCrunch
- →Microsoft launched its first dedicated AI cybersecurity model alongside a new agentic security platform, claiming it outperforms competitors on cost and accuracy. TechCrunch
AI TLDR