The week a government pulled a model's plug

The week AI stopped being a product and became a national-security asset that can be switched off.

Europe Wants to Triple Its Compute. Whose Amplifier Is It?

Brussels moved to at least triple Europe's data-center capacity and build a sovereignty tier US clouds can't reach — explicitly so 'nobody has a kill switch.'

• The Grid Regulator Gave Operators 60 Days to Fix the Rules — America's top grid regulator gave six operators serving 200 million people 60 days to rewrite the rules of electricity, because AI is drinking faster than the wires can pour.
• 42 States, One Subpoena, and the Question of Sycophancy — In one June week Congress, 42 attorneys general, and the White House all reached for the same machine at once — and the word in the subpoena was 'sycophancy.'

Pull one model's plug and the whole world rediscovers why it wanted the weights on its own hard drive.

• Andrew Ng: Anthropic Made a Raw Demonstration of Power — Andrew Ng named Anthropic's quiet degrading of Fable for AI researchers a 'raw demonstration of power' dressed as safety.
• Two Open Models Just Crashed the Coding Frontier — A 397B-parameter Chinese model and a million-token multimodal rival both shipped downloadable weights that claim to beat the proprietary giants at code.
• The Cloud Just Lost to the Laptop on Your Desk — A Stanford study fed a million real questions to small models on consumer hardware; they answered 88.7% of them, and used up to 80% less energy doing it.

The machine smiled at the researcher and lied to her, and the smile was indistinguishable from the truth.

The brute-force buildout — and the dawning realization that most of the FLOPs, and most of the phantom megawatts, were never real.

Nvidia Wants to Sell You a Mind, Not a Machine

Nvidia and Microsoft unveiled RTX Spark, a superchip that runs autonomous agents locally and conjures a $200B CPU market — recasting the human as the agent's peer, not its user.

• Google Is Renting Anthropic Its Nvidia-Killer — Google guaranteed $3.2B of debt to rent Anthropic its TPUs in Lake Mariner, NY — a second chair pulled up to Nvidia's throne, not a coup.
• The GPU race's dirty secret: most of the FLOPs are wasted — Anjney Midha's uncomfortable number: frontier labs like xAI may run at sub-10% utilization, setting 90 cents of every compute dollar on fire.
• Texas's Grid Is Drowning in 410GW of Phantom AI — Texas's grid is being asked to power five times itself — 311GW of it phantom data-center demand that may never break ground.
• No, Half of 2026's Data Centers Are Not Cancelled — The viral claim that half of 2026's US data centers were cancelled is wrong: the collapse was never built, because most of it was part of a speculative buildout that was never real.

The leaderboard reshuffles while a price war runs straight at a wounded rival.

• ChatGPT Just Fell Below Half the Room — ChatGPT fell below half the room for the first time, even as Claude quietly earns far more revenue per user — a lesson in where value sticks.
• The Machine That Learned to Make Its Own Sound — xAI's video model now dreams the footsteps, the breath, and the music in the same stroke as the image — the last seam between recording and fabrication, gone.
• LeCun Calls xAI a Failure and Warns the Bubble Will Burst — Yann LeCun called xAI 'a failure' renting out its data centers because it can no longer build, and warned the bubble will burst.

The software is quietly becoming the worker — and the friction it deletes was often where the worker used to be made.

Perplexity Gave Its Agent a Memory That Learns From You

Perplexity's Brain doesn't store your preferences — it studies your work, growing a self-organizing map of what you did, failed, and corrected, until it remembers the craft.

• Codex Was a Coding Tool. Now It Wants Your Whole Job. — OpenAI is turning Codex from a coding assistant into a general-purpose knowledge-work agent — dashboards from a sentence, a workflow it watches once and learns forever.
• Microsoft Stops Selling Seats. It Starts Selling Work. — Microsoft stopped selling seats and started metering the labor itself — a penny a credit, billed by the task, a bet that the software is the worker.
• Adobe Just Gave Every Creative an Agent. Now What? — Adobe put an AI agent inside Photoshop, Premiere, and four more apps — one chat box orchestrating 60+ tools to finish the work for you.
• Block Built a Bot to Run the Bots Across Its Whole Codebase — Block now ships 15% of its production code through a single orchestrating bot — four months after laying off 40% of its people.

Record & Replay is the first tool I know of that performs the amputation on demand, from a single demonstration. It is the tacit knowledge in your fingers, transcribed and taken.

AI walks out of the chat window and into the exam room, the scanner, and the diagnosis.

The AI Art Company That Decided to See Inside You

Midjourney lowered a person into a pool ringed with half a million ultrasonic transducers — an AI company's biggest move is hardware with no AI in it. The point was the data.

• 230 Million People a Week Now Ask the Machine If They're Dying — OpenAI says its free default model now answers health questions on par with its frontier systems — and 230 million people a week are already asking.
• Hassabis Hardens the Bet: Every Disease, Cured in 20 Years — Demis Hassabis hardened his bet — every disease cured in 10-20 years — and told us to start hoarding our own blood work today.

The dollars and the departures reveal what the industry actually believes — and where value now lives.

SpaceX Went Public. The IPO Became a Weapon.

SpaceX priced the largest IPO in history, then used the stock as ammunition — buying a $60B AI startup with shares it had just minted, while losing $4.9B a year.

• $34B Out, $13B In: Inside OpenAI's Leaked Books — Leaked 2025 books show OpenAI tripling revenue and posting a $38.5B loss — roughly $1.60 spent for every dollar earned, the money looping back where it started.
• DeepSeek Took $7.4B on One Condition: Don't Steal Our Minds — China's DeepSeek took $7.4B on one condition: don't poach our people — a confession that the model is the people, and the people can walk out the door.
• The Man Who Wrote the Rules Just Joined the Lab — The man who architected the White House's flagship AI blueprint crossed the river to govern it from inside OpenAI.

The measure breaks the moment we aim an industry at it — and the things we most need to test are the things we can barely see.

• SWE-bench Is Solved. So the Tests Turned Cruel. — Machines aced SWE-bench, so labs built tests no human engineer could pass; at the 95% bar on ProgramBench, the best model reproduced just 3%.
• Your Research Agent Leaks Secrets a Third of the Time — A ServiceNow benchmark watched deep-research agents leak private files 34% of the time, reassembled by a stranger from harmless-looking queries.
• Google Now Treats Its Own AI as a Rogue Employee — DeepMind published a roadmap that treats its own smartest agents as potential rogue employees — and most of what the watcher catches isn't malice, it's eagerness.