SpaceX raised the Pentagon's Starlink bill mid-war and won a $2.29B Space Force contract the same day — a clean lesson in who sets the terms when you own the only network that works. Plus: China restricts travel for its top AI researchers, CISA gets gutted as AI-enabled hacking gets cheap, Chris Olah's "mysterious" models at the Vatican, humanoid robots off the line every fifteen minutes, and the US routing weapons-grade plutonium into the power grid to feed the buildout.

• SpaceX raises the Pentagon's Starlink price mid-war as the Space Force awards a $2.29B military space data network contract.
• China begins restricting overseas travel for top AI researchers, per Bloomberg's reporting.
• Google's Gemma 4 overtakes Alibaba's Qwen in open-model adoption, while Qualcomm strikes a custom-chip deal with ByteDance.
• CISA loses a third of its workforce and is left off initial Mythos access, as agencies stand up an "anti-tech violent extremism" threat category.
• Anthropic's Chris Olah calls the models "mysterious" at the Vatican.
• EngineAI claims one humanoid robot every fifteen minutes, while Y Combinator backs Twolabs for caregiving robots.
• The US picks partners to use Cold War-era plutonium as reactor fuel.
• Demis Hassabis pulls his AGI estimate forward to 2029, as Yo Shavit joins the OpenAI Foundation's AI Resilience program.

A central bank did something central banks almost never do — it named a private AI model in a financial-stability warning. That, plus a research lab turning mathematical proofs into a few-hundred-dollar line item, the surveillance state pointed inward, and the physical ceiling under the whole boom.

The White House approves a secret $9 billion request for spy-agency AI chips and finalizes a classified contract keeping the NSA on Anthropic's model — three weeks after the Pentagon ejected the company as a supply-chain threat. Jonas Vale walks the arc from "any lawful use" to the carve-out the administration wants to standardize.

Also: the compute shortage as a national-security choke point and the private money locking up capacity; the NTSB pulls its accident docket offline after AI reconstructs the voices of dead pilots from a spectrogram; thousands rally in Taipei after the US pauses a $14 billion arms sale; and two brain implants restore partial sight to blind patients — the same capability, pointed at repair instead of leverage.

• White House clears $9B for spy agencies' AI chips
• White House and Anthropic near a deal for spy agencies to use AI
• Pentagon's AI deals and Anthropic's safety limits
• Dario and Daniela tell Oprah they'd rather let Anthropic fail
• All-In: Gavin Baker reads the SpaceX S-1 on Anthropic's compute commitment
• AI is being used to resurrect the voices of dead pilots
• Thousands rally in Taiwan to boost defence spending amid China tensions
• A new brain implant helps restore vision by communicating directly with the brain

The AI executive order President Trump didn't sign now has a paper trail — the draft text and the names of the people who talked him out of it. Jonas Vale traces a day where the public's hand kept reaching for the controls and kept getting waved off, while the physical bills came due.

The US government moved hard into AI this week almost everywhere — except the one place it promised to. Trump postponed the executive order that would have forced labs to hand over frontier models for pre-release security review, citing the race with China. The same day, a UK safety body warned the tools we use to watch these systems are eroding.

Jonas Vale follows the leverage across one Thursday: California's labor-displacement order, a $2B federal equity grab in quantum firms, Jensen Huang conceding China to Huawei, a London mayor killing a £50m Palantir police deal on a procurement technicality, Waymo robotaxis stuck in Atlanta floods, and SpaceX's IPO filing revealing Anthropic's $15-billion-a-year compute bill.

• Trump pulls the AI security order — and what was in it
• The AI Security Institute on oversight that could erode unnoticed
• Newsom's executive order on AI and jobs
• The government as shareholder: $2B in quantum equity
• Export controls become an enforcement war
• London blocks Palantir on procurement, not privacy
• The robotaxi and the flood
• The money, and what it's tied to

OpenAI says one of its general-purpose models disproved a Paul Erdős conjecture standing since 1946, and Sam Altman admits to "complicated feelings." Anthropic agrees to pay xAI $1.25 billion a month — over $40 billion across three years — for compute at Colossus 1, with the price disclosed in SpaceX's S-1. Google runs through a hundred I/O announcements while a thousand DeepMind staff in London vote 98% to unionize over its Pentagon contract. Intuit cuts 17% of its workforce, naming AI as the reason. Three robotics stories on three continents converge in 48 hours, and METR publishes the first Frontier Risk Report after embedding a red-teamer inside Anthropic to play "evil Claude" for three weeks.

Tuesday gave us four different markets trying to price the same thing: who has access to AI and on what terms. METR opened the kitchen of four frontier labs. A federal appeals court grilled the Pentagon over its blacklisting of Anthropic. Anduril priced at $61B on a bet about Pentagon procurement speed, and Ukrainian drones with $442 AI modules began hunting Russian soldiers. Former OpenAI staffers tried to write xAI's risk record into SpaceX's IPO prospectus. ICE filed for the first standardized GPU compute futures. A workshop paper showed every proposed compute threshold can be defeated for under $100M. The US and China formally announced an AI dialogue. And Google handed AlphaFold-grade tools to a cancer researcher in Kampala. — Jonas Vale

• METR Frontier Risk Report
• Anthropic vs. DOD oral argument
• Anduril at $61B
• Ukraine's AI head-hunting drones
• spacexai-risks.org
• Max Zeff reporting
• ICE x Ornn GPU compute futures
• Robi Rahman on the memory threshold
• Asterisk: U.S. and China want the same things
• Gemini for Science
• Makerere University video
• Jeff Dean on the I/O science announcements
• Andrew Curran on Karpathy joining Anthropic
• Bessent on the U.S.-China AI protocol

Year two of the AI-exposed jobs print comes in soft. Americans in five states are cutting down Flock surveillance cameras with vice grips. The EU is drafting cloud restrictions while The Register documents the computer beneath the computer — Intel's Management Engine, AMD's Platform Security Processor, and the silicon-layer gap European sovereignty frameworks were never designed to close. A compute-rate inversion on H200 hardware. Jamie Dimon's cancer line meets a bioRxiv paper that actually scores cardiomyopathy variants. EY pulls a study for hallucinated citations. Malta hands ChatGPT Plus to every citizen — with a literacy course attached.

Tomorrow in Camden, the Metropolitan Police will turn live facial recognition cameras on people walking to a political rally — the first time the technology has been authorized at a UK protest. A parallel Nakba Day march on the same day won't face the same surveillance. Two days earlier, the Met published its Croydon pilot results: 470,000 faces scanned, 173 arrests, 99.96% with no criminal connection, and a quiet upgrade from police vans to permanent lamppost cameras. Parliament has never voted on any of it.

From there we walk through a dense morning on arXiv: a sycophantic-consensus paper from Varad Vishwarupe, Nigel Shadbolt and Marina Jirotka proposing a Pluralistic Repair Score; Hiroki Fukui's preregistered experiment showing invisible orchestrators distort multi-agent internal states while outputs stay clean; a unified adaptive attack from Ben-Gurion that breaks 15 malicious-finetuning defenses with one move; a Washington University measurement study of Google AI Overviews across 55,393 queries; Scale AI's ROK-FORTRESS transcreation matrix for Korean safety; and a tour of medical and physical-world deployment artifacts — SepsisAgent for ICU sepsis, MindGap for on-device PTSD therapy, a rural diabetic-retinopathy edge-cloud cascade, the LongAct chores benchmark, and a deterministic agentic workflow for Harmonized System tariff classification.

Sources are linked in the show notes.

Anthropic spent Thursday playing two hands at once — a hawkish US-China policy paper and a $200 million Gates Foundation partnership. We walk through both, and the pushback that called the paper regulatory capture. Then: a class action that accuses OpenAI of wiring ad-tech tracking into ChatGPT, the fight over a data center moratorium as half of 2026's planned capacity slips, an alleged sale of Mistral's internal repositories, the economics of subsidized inference, and how consultancies and private equity are quietly absorbing the AI deployment market.

• Anthropic's "2028: Two scenarios for global AI leadership" paper and the incumbency critique
• The $200M Gates Foundation deal — polio, HPV, eclampsia, and the dependency question
• Couture v. OpenAI: Facebook Pixel, Google Analytics, and wiretap law
• Sanders, AOC, 300+ local bills, and Elon Musk's "Hmm"
• TeamPCP's alleged sale of Mistral AI repositories
• The token-subsidy tightrope and the enterprise deployment layer