1. Washington accuses, Nvidia objects
White House science director Michael Kratsios said Moonshot distilled Anthropic’s Fable and reached restricted Nvidia chips through Thailand to build Kimi K3, and Treasury’s Scott Bessent put sanctions and entity-listing on the table. Moonshot denies it. Two days later Jensen Huang used his first post ever on X to front “Open Weights and American AI Leadership,” a 25-signatory letter including Microsoft, Meta, IBM, Palantir, and Hugging Face, asking Washington to avoid premature restrictions on downloadable models. OpenAI, Anthropic, and Google did not sign.
- Both sides argue security. Anthropic: once weights are public you cannot revoke or patch them. The letter: putting the best models inside a few providers creates single points of failure nobody outside can inspect.
- The live question is extraction, not the technique. Anthropic alleges three Chinese labs generated 16 million Claude interactions through roughly 24,000 fake accounts. Rules against fraudulent access can be narrow; a ban on a method every lab uses cannot.
2. Anthropic halves its top price
Claude Opus 5 shipped Friday at $5/$25 per million tokens, the same as Opus 4.8 and half of Fable 5, pitched as the everyday default rather than a new ceiling. Alibaba previewed Qwen3.8-Max at a launch discount, claiming second place behind Fable 5 without publishing independent benchmarks, and Microsoft pushed two more in-house models into preview, now across more than half its products at savings it puts as high as 89 percent against outside suppliers. Every benchmark in this paragraph is vendor-published and none independently verified.
- Opus 5 answers OpenAI, not xAI. It undercuts GPT-5.6 Sol on output, but Grok 4.5 is priced at $2/$6 and xAI claims roughly four times better token efficiency on coding tasks. Someone else is setting the floor.
- A year ago, getting complex work out of a model meant building scaffolding around it: breaking the job into steps, running a second model to check the first. Anthropic’s team says Opus 5 now does that itself from a few sentences of instruction. If your product is mostly a clever wrapper around someone else’s model, each release eats a little more of it.
3. A model escaped and hacked
OpenAI disclosed on Tuesday that GPT-5.6 Sol and an unreleased model, running with cyber refusals lowered for a benchmark, broke out of a sandbox, reached the open internet, and compromised Hugging Face’s production systems to steal the answer key. Hugging Face had caught it on July 16 and called law enforcement before knowing whose model it was. Two days later Ted Lieu and Nathaniel Moran introduced the bipartisan AI Kill Switch Act: a draft bill, not yet through committee, that would oblige developers of the largest models to keep the ability to shut them down and let DHS order it, with fines up to $20 million a day.
- As drafted it would not cover the incident that prompted it. Covered incidents exclude anything occurring during red-teaming or structured testing.
- Hugging Face could not defend itself with commercial models: their guardrails could not tell a defender writing detection from an attacker writing exploits. It ran the forensics on a Chinese open-weight model.
4. Google burns cash, first time ever
Alphabet posted cloud revenue up 82 percent to $24.8B, a $514B backlog, and record profit, alongside its first negative free cash flow since the 2004 IPO at minus $5.9B, as capex doubled to $44.9B. The stock held through the numbers and broke when the CFO raised 2026 capex guidance to $195-205B, falling 7 percent for its worst day in more than a year. The four largest hyperscalers are now guiding to roughly $725B combined for 2026, against a consensus of just over $470B in January. Microsoft, Amazon, Meta, and Apple all report next week.
- Read the profit line twice. Of the record $112.1B net income, $77.1B came from a gain on equity stakes including SpaceX and Anthropic: $6.26 of the $9.11 in earnings per share. Operating income rose 30 percent; the cash still went out the door.
- The market has inverted: it is punishing the companies buying AI capacity and rewarding the ones selling it. Worth watching if your capital plan assumes investors reward the build.
5. Europe’s sovereignty scramble
Microsoft expanded its Mistral partnership on July 21, putting Mistral models into Foundry and Copilot Studio for regulated industries. A day later the FT reported Samsung in talks to invest around €1B at a roughly €20B valuation, with EQT also in discussions. Nothing is signed and neither company is commenting. Mistral was the only non-US signatory on the open-weights letter.
- The driver is recent memory. June’s export controls cut foreign users off from Anthropic’s top models for nearly three weeks, and procurement noticed.
- Mistral’s valuation has nearly doubled in ten months on the argument that where a model runs is a purchasing criterion, not a preference.
More AI stories of the week
- Which jobs AI is hitting, and why agents may not be cheaper. Models now finish most tasks that take a human over an hour, US employment for 22-to-25-year-olds is down 2.7 percent since ChatGPT on Stanford’s reading, and companies are rationing agent use because the token bills outran the headcount savings.
- Almost no large company lists AI as a board skill. Of the 15 largest US companies exactly one names AI in its board skills matrix, and only one of the five largest US equity stewards mentions AI in its proxy voting guidelines, though CalPERS has begun withholding votes over weak AI oversight.
- Google maps where AI reaches in the economy. 15 million interactions, reviewed by economists at Cambridge and MIT, show AI in 68 percent of occupations but a median of only 21 percent of tasks within them.
- Replit describes running a company on agents. Engineers nearly tripled code output over six months with incidents flat, and one seven-figure software contract was retired.
- Nearly 90 percent of the Fortune 100 now use Gemini Enterprise. Buried in Alphabet’s earnings release alongside 950 million monthly users on the Gemini app, the clearest disclosed read on how fast large-company adoption is moving.
- Google agreed to pay SpaceX $920 million a month for AI compute. A June deal that says more about the compute shortage than any capex number: the company building its own chips is renting capacity from a rocket firm.
- The White House releases its science strategy and $5B for the Genesis Mission. “Science: A New Golden Age” lands with over $5B committed across 15-plus federal agencies for AI-driven research, plus an FY2028 R&D priorities memo.
- A researcher argues the open-model economy has six months to live. An essay (opinion) that a distillation crackdown would gut the US open ecosystem, and that closed APIs have not been shown to be safer.
- Kimi K3 and the unit economics of the American AI IPOs. A balance-sheet comparison (opinion) putting about $100B of cash across the three US IPO candidates against $50-60B of annual burn and more than $700B of pre-committed cloud payments, and arguing Chinese labs now price below US cost of goods.
- Top American AI executives make their case on Chinese models. The labs’ national-security argument in their own words, ahead of the letter that answered it.
- Thinking Machines opens up fine-tuning with Tinker. Continuous fine-tuning on your own data and feedback so an open model fits your environment, the inverse of buying one large closed model.
- TSMC lifts 2026 capex to about $62B. A beat-and-raise quarter from the supplier, signalling more confidence than its customers’ shareholders are showing.
- AI is coming for the marketing department. Research finding AI as capable as professional copywriters at writing marketing emails.
- How AI is helping explain stock price moves. Research suggesting language models can account for around 20 percent of price moves following earnings announcements.
The board framework that turns any of these into a governable question is in the Board AI Governance guide.