Four throughlines run through today’s edition. The first is money moving faster than governance can track it: Nvidia cut its Ohio backstop to under $120 billion, Stripe is paying over $7 billion for OpenRouter, SpaceX closed on Cursor, and OpenAI turned a $100 warrant into a $2.3 billion Cerebras stake.
The Ownership Shuffle: Money Consolidates Around Compute and Distribution
Four separate deals reshaped who controls AI’s infrastructure today, from chip financing to routing to the coding tools built on top of it.
- Nvidia Halves Ohio Data Centre Backstop to Under $120 Billion. Nvidia cut its financing guarantee for the SB Energy campus by more than half and narrowed it to the project’s first phase, Reuters reported.
- Stripe Reportedly Agrees to Buy AI Router OpenRouter for $7B+. Bloomberg reports Stripe will pay over $7 billion for OpenRouter, more than five times the router’s valuation from a funding round just three months earlier.
- Cursor Confirms SpaceX Now Owns It, Closing the Option From May. Cursor’s own blog confirmed SpaceX completed the acquisition, the outcome of a $60 billion call option Bloomberg first reported in May, without disclosing final terms.
- OpenAI’s $100 Cerebras Warrant Hides a $2.3 Billion Stake. OpenAI paid $100 to exercise vested Cerebras warrants now worth an implied $2.3 billion, a filing that landed two days before its Cerebras powered Ultrafast preview.
- Wispr raises $280M at $2B valuation, looks past dictation. The voice AI startup tripled its valuation in under a year and is betting an interface lab and meeting notetaker can outrun free rivals.
The Open Weights Clock: China Sets the Cadence Everyone Else Follows
Open models keep setting the pace on benchmarks, and the download data shows which ones teams actually run once the hype settles.
- GLM-5.3 leads a cybersecurity benchmark, and Z.ai still opens the weights. Z.ai’s new open weights model topped a vulnerability benchmark and doubled its own exploit score, with weights due to publish in two weeks.
- Nathan Lambert: China’s AI edge is speed, not stolen weights. Interconnects argues Chinese labs match Western frontier models mainly by shipping in days what OpenAI and Anthropic sit on for months.
- Hugging Face Data Shows Hype and Usage Split on Open Models. Hugging Face’s summer ecosystem report finds likes track excitement while downloads track deployment, and Qwen has quietly become the default base model.
- Qwen 3.8 27B Is Great, But Its Default Reasoning Wastes Time. Simon Willison’s hands on tests found the open 27B model excellent, but its default xhigh reasoning setting burns minutes on trivial prompts.
Agents as Infrastructure: Memory, Roles, and the Architecture Underneath
Today’s agent and model research is about plumbing more than raw capability: how systems remember, organize, and hold state.
- Files, structured stores, and trained memory tested head to head. A controlled benchmark found structured memory stores beat curated markdown files by 29 points, while trained experience only helped weaker models.
- Google adds file-based Custom Agents to Antigravity, IDE next. Antigravity 2.0 and its CLI now let developers define agent roles as markdown files, a fix framed around token cost rather than smarter agents.
- Microsoft preprint stops transformers from discarding their own state. A new architecture feeds a model’s own top layer hidden state back into itself each step, replacing today’s practice of throwing it away.
- A Language Model That Only Knows What a Fifth Grader Knows. LittleLearner trains models solely on elementary school material to test whether new skills are learned or just surfaced by scale and prompting.
- Google Reportedly Taps AMD to Design a CPU-Heavy TPU. A SemiAnalysis note reportedly says Google is working with AMD on a TPU variant with on package CPU cores built for reinforcement learning.
Trust Catches Up: Watermarks, Campaigns, and the Backlash Ahead
The industry’s credibility problem showed up today in a product detail, a policy pattern, and a direct admission from a lab CEO.
- Claude’s Watermark Draws Three Separate Lines of Attack. A blogger, a paraphrase tool builder, and a legal outlet each fault Anthropic’s Claude watermark for a different reason, from usability to enforceability.
- AI now outranks other issues in nearly 40% of US campaign sites. A review of over 1,200 candidate websites found AI mentioned in nearly 40 percent of House, Senate, and governor races, driven largely by data center fights.
- Amodei: AI’s backlash is a trust crisis, not a messaging one. Answering an investor’s critique on X, Anthropic’s CEO defended his regulatory theory and admitted AI firms have not delivered on their promises.
- Zvi Mowshowitz says Greenblatt’s 2031 timeline holds, his optimism doesn’t. Mowshowitz backs Ryan Greenblatt’s AI R&D automation forecast but rejects his narrow training plan and his faith that verification alone contains the risk.
The Human Side: What Gets Easier, What Still Doesn’t
Two stories from the research and workplace side of AI show where automation genuinely helps and where judgment still has to fill the gap.
- AI cut a PL paper’s hardest step, and submissions nearly doubled. Ilya Sergey says machine checked proofs that once ate months now take weeks, but his own data shows reviewing didn’t get any cheaper.
- Andrew Ng’s AI Engineering Map Puts Judgment Inside the Job. Andrew Ng’s four skill framework quietly argues that product sense, not just coding agent fluency, now belongs to the engineer.
Quick Hits
The rest of what moved today, in one line each.
- New coding agent turns word problems into checked math proofs. MathCode converts plain language math questions into Lean 4 theorems and attempts formal proofs, according to its project page, a narrow but concrete step toward machine checked reasoning.
- OpenAI taps Wiz’s Dali Rajic as CRO amid exec exodus. Axios reports Greg Brockman is driving OpenAI’s leadership rebuild, including new safety exits, as the company adds a chief revenue officer ahead of an expected IPO.
- AI Video Market Rebounded, But Not on Sora’s Terms. Revenue and valuation gains at rival video startups show the category recovered around products Sora never delivered, per The Decoder, undercutting OpenAI’s early lead in AI video.