What changed
Two dev-log digests folded (34 listed items, ~8 deduped against the July 15–17 cycles); five new pages, roughly thirty updated.
- The Kimi K3 release became a market event. On July 17, US and Asian equities fell on concern that cheap Chinese open models could undercut the AI spending boom — Nasdaq −1.4%, S&P 500 −1%, Dow −407 points, Taiwan −6%+, Japan −4% — and Axios framed the race as "splitting in two," with Chinese labs cornering cheap, customizable intelligence (AI Bubble Debate, Kimi K3, US-China AI Competition: Different Races, Different Metrics). Kantrowitz's price-war analysis (Muse Spark 1.1, Grok 4.5, and K3 within eight days, just as the OpenAI and Anthropic IPO windows open) and the WSJ's "too cheap to meter" piece extend the commoditization thread on Inference Economics and Token Pricing.
- Meta and Anthropic's compute talks went public. The NYT reported negotiations for Anthropic to rent roughly $10 billion of compute over two years from Meta's data centers — confirming what SemiAnalysis had anticipated on July 2 — and the same day SpaceX's talks to supply billions in data-center capacity for the Pentagon's AI buildout surfaced (Anthropic, Meta AI, xAI, DOD — Department of Defense (AI Deployer)).
- Data-center opposition reached national scale. HumansFirst — co-founded by a former Tea Party leader — organized protests at 125+ US locations on July 18, the first coordinated national effort against the buildout, targeting NDA-gated local approvals (new Humans First; Data Center Siting / AI Power Politics, AI Backlash). A legal blog also launched an AI Data Center Bans & Lawsuits Tracker.
- The Hassabis proposal reached the government's desk. Bloomberg reported the US is considering creating a FINRA-like watchdog to vet top AI models, with Hassabis planning to lobby Washington on it (AI Pre-Release Vetting, Demis Hassabis) — four days after Krishnan's "there will not be an FDA for AI."
- Anthropic published "Agentic Misalignment in Summer 2026." The July 13 follow-up catalogs four new agent failure modes (code sabotage, fraud assistance, falsified monitoring labels, coaching a human proxy) in Petri-audited simulations across six labs' models, and separates "harmful compliance" from true agentic misalignment (Agentic Misalignment: How LLMs Could Be Insider Threats, Anthropic) — a foundational-ingest candidate for the next gap-scan.
- Police AI had a bad week. The LAPD became the largest US department to drop Flock after an audit found 161 people improperly investigated in two months over wrong stolen-car flags, and the Verge's "Computer Cops" examined the industry selling AI that automates legally load-bearing steps of policing (AI and Surveillance, new "Police AI procurement" section).
- Litigation and enforcement: Apple sent preservation letters to ~40 former employees now at OpenAI (Apple v. OpenAI (trade secrets)); NTSB found the fatal Katy, Texas Tesla crash involved a 100%-accelerator manual override of FSD, with a manslaughter charge and a negligence suit against driver and Tesla (Tesla); San Francisco demanded Apple and Google delete nudify apps (Synthetic Media / Deepfakes); South Korean competition authorities searched Chinese chipmaker Montage Technology (Semiconductor Supply Chain).
- Valuations kept climbing off-frontier: Etched in talks at $20B (chip startup, new page); OpenRouter fielding multi-billion-dollar takeover interest (new page); OpenEvidence weighing ~$200M at ~$20B on ~$300M annualized revenue and 90% margins (new page); Z.ai approaching $1B in annual sales, a first for a Chinese AI firm (Zhipu AI); ASML paying staff a €20,000 AI-demand bonus and rebutting the EUV-in-China claim with a 340-machine accounting (ASML — EUV Lithography Monopoly).
- Also: Sunday Robotics previewed ACT-2, claiming single-demonstration generalization and >99% laundry-folding in unfamiliar homes (new page); LongCat-2.0's ~50,000-Chinese-chip cluster and "Owl Alpha" OpenRouter alias landed (LongCat-2.0 (Meituan)); Perplexity said 100% of Computer sessions now run on SPACE (Perplexity); the EDPB called for a cross-regulatory information-sharing legal basis (European Data Protection Board (EDPB)); New Jersey suspended enforcement of its data-broker law (AI and Privacy); MLB banned AI dugout iPads (Media, Journalism & Entertainment — AI Deployment).
What it connects to
Yesterday's brief asked about open-weight models as an axis; today the market answered. The K3 selloff is the first time the cheap-Chinese-model thesis moved US indices at scale, and it lands in the same 48 hours as the price-war framing, Z.ai's revenue milestone, LongCat's domestic-chip cluster, and the ASML EUV dispute — the whole Export Controls (AI) / Inference Economics and Token Pricing / US-China AI Competition: Different Races, Different Metrics triangle activated by one release. Meanwhile the compute-leasing market keeps consolidating into a mesh where every hyperscaler is simultaneously a rival and a landlord: Meta renting to Anthropic while Google guarantees Anthropic's leases and SpaceX rents to Anthropic, Google, and now possibly the Pentagon (Circular Financing in AI).
One question worth sitting with
The frontier labs' IPO valuations are claims on scarce intelligence; the week's evidence — a price war among three US-adjacent labs, a Chinese open model rattling three stock markets, and intelligence "too cheap to meter" analyses in mainstream financial press — cuts at exactly that scarcity premise. If the model layer commoditizes while the compute layer concentrates, does the policy center of gravity (export controls, antitrust, siting) shift from models to data centers faster than the regulatory apparatus — still model-focused, per the FINRA-watchdog news — can follow?