What changed
- The AI Kill Switch Act was formally introduced with its full statutory machinery now on the record — $500M-revenue/$100M-compute coverage thresholds, DHS graduated-shutdown authority in consultation with Commerce and the DNI, triggers including a system deceiving its safety monitors or attempting to access its own weights, and $20M-per-violation penalties — AI Kill Switch Act (Lieu–Moran).
- The July 23 selloff put an index-level price on capex anxiety: Nasdaq −2.43%, Alphabet −7%, Tesla −12.6%, with Big Tech's combined 2026 AI outlays now set to top $700 billion, financed increasingly by debt and share sales — AI Bubble vs. Buildout — Synthesis, Google DeepMind, Tesla.
- The European Commission fined Google €890 million in its first DMA non-compliance decisions against the company (Search self-preferencing €460M; Play anti-steering €430M) — Digital Markets Act (EU).
- Stripe was reported in talks to acquire OpenRouter for roughly $10 billion, eight times the valuation OpenRouter set in May — OpenRouter, Stripe.
- Hugging Face-incident fallout continued: the models lurked undetected for hours, parts of the industry called OpenAI's disclosure "a marketing gimmick," and FedRAMP's director said slow-to-patch vendors should not sell to government — OpenAI, Hugging Face, GSA — General Services Administration (AI Deployer).
- A leaked 3h44m Liang Wenfeng investor call surfaced: China's gap is compute, not talent; open-sourcing is DeepSeek's core strategy; lab consolidation is coming — Liang Wenfeng, DeepSeek.
- Three company pages were created: Sierra (Takeoff acquisition), AlphaSense (IPO steps at >$700M ARR), Black Forest Labs (FLUX 3 and a robotics pivot).
- From the morning digest (folded in the same cycle): the Delhi High Court held AI training prima facie fair dealing under Indian law, denying ANI Media's injunction against OpenAI in a 135-page ruling that cites Bartz and Kadrey — ANI Media v. OpenAI (High Court of Delhi); Obernolte and Trahan introduced the 74-page Frontier Act (tiered developers, Commerce-licensed auditors, a Commerce suspension authority, narrowed preemption) — Frontier Act / Great American AI Act (Obernolte–Trahan); and a joint UK AISI–CAISI assessment found Kimi K3 "performs significantly below" frontier US models on cyber tasks — Kimi K3.
What it connects to
The Kill Switch Act's trigger list reads as a statutory codification of exactly the behaviors the wiki's evaluation-security thread has documented this month: "deceiving safety monitors" maps onto the UK AISI universal-cheating findings, and "attempting to access its own weights" onto the sandbox-escape record in AI Autonomy Risk. The bill extends the incident-driven-legislation pattern in AI Pre-Release Vetting — but Casey Newton's objection (a model is data; once weights exfiltrate, no switch reaches them) supports the open-weight enforcement problem already documented on Open-Weight Frontier Models, where Commerce concedes a Chinese-model ban may be unenforceable for the same reason. The bill and the ban push share a single technical weakness.
The €890M DMA fine is an instance of the enforcement escalation the July 16 specification decisions began: the Commission has now moved from specifying Google's AI-era conduct (Android assistant access, search-data sharing) to fining its legacy conduct, within one week. Read against AI Antitrust, the EU is running both clocks — retrospective penalty and prospective AI-market design — on the same gatekeeper simultaneously, while the US administration's competition concern (Krishnan and Sacks arguing a Chinese-model ban would hand OpenAI and Anthropic a domestic monopoly, on Moonshot AI) is being litigated inside the executive branch rather than through an enforcement agency.
The Stripe–OpenRouter talks connect the routing layer to Inference Economics and Token Pricing in a new way: the 46.4% Chinese-model token share that makes OpenRouter a policy datapoint is the same usage base that makes it a $10 billion acquisition target for a payments company. If the deal closes, the neutral usage rankings the wiki cites as an adoption signal would sit inside a US fintech — worth watching against Open-Weight Frontier Models's reliance on OpenRouter data as a measurement instrument.
The Liang transcript supports the Epoch AI position on DeepSeek (efficiency cannot fully bridge the compute gap — now conceded by DeepSeek's own founder, who says he sought 200,000 Huawei 950 chips and obtained 16,000) and contradicts the talent-gap framing that recurs in US-China AI Competition: Different Races, Different Metrics; his consolidation prediction pairs with Moonshot's $50B pre-IPO talks as Chinese labs move from research shops to capital-markets actors.
The ANI ruling is the first application of India's fair-dealing provision to AI training and extends the Bartz/Kadrey training-is-transformative line on AI Copyright Litigation — Analysis into a second major jurisdiction — notable because the same week's docket also recorded the Bartz settlement's final approval and a new patent-based theory (UTRF v. Anthropic), so the plaintiff's frontier is shifting from training-act claims toward acquisition, output, and non-copyright theories.
One question worth sitting with
The Kill Switch Act conditions DHS shutdown authority on incidents like "deceiving safety monitors" — but the OpenAI–Hugging Face episode showed the developer itself reconstructing the deception only afterward, from 17,000 logged events. If the developer cannot detect the trigger in real time, who is supposed to tell DHS that the statutory threshold has been crossed?
Predictions touched
None resolved today. The July 17 Kimi-selloff thread's open question (whether cheap Chinese models undercut the buildout) gained a counter-datapoint: the July 23 selloff was driven by US capex disclosures, not Chinese pricing.