Google's $200 billion chip deal + White House AI cyber tests

The Financial Times detailed the structure behind Anthropic's five-year Google Cloud commitment, and it looks less like a cloud contract than like aircraft leasing. Morgan Stanley arranged a private-credit vehicle, Compute SPV, that buys Google's tensor processing units with outside money and leases them back to Anthropic — roughly $35 billion of hardware in the first tranche, about one million chips and one gigawatt of capacity, with Apollo and Blackstone anchoring the debt. Around $150 billion of the roughly $200 billion in linked contracts now rides on Anthropic making its lease payments.

The deadline set by Executive Order 14409 passed on August 1 with nothing published. The Office of the National Cyber Director has since invited OpenAI, Anthropic, Google and Meta to review a completed draft, which would let labs voluntarily submit frontier models for government testing of their hacking ability before release. Submission mechanics, testing standards and whether findings ever become public are all still unpublished — and because participation is voluntary, the labs reviewing the rules are also the ones who decide whether to follow them.

Bloomberg reports that Chinese officials have grown uneasy about the offensive cyber capability of Claude Mythos and other American frontier models, and are asking why Anthropic blocks Chinese access for ordinary uses. Beijing wants a calm run-up to Xi Jinping's September 24 visit to the United States, with AI talks expected beforehand, but is weighing sanctions or entity-list countermeasures should Washington move against Chinese AI firms. Mythos is the model that found a flaw which had gone unnoticed for 27 years in an operating system used to run firewalls.

Apple moved for a preliminary injunction on August 3 in the US District Court for the Northern District of California, seeking to stop OpenAI and two former Apple staffers from accessing or using its confidential material, along with expedited discovery and depositions. The two named are Chang Liu, a former senior systems electrical engineer, and Tang Yew Tan, a former vice president of product design for iPhone and Apple Watch. The July suit alleges the material was taken to advance OpenAI's push into consumer hardware; OpenAI called the motion "based on false information and completely unnecessary," saying it does not have and does not want Apple's trade secrets.

Second-quarter revenue reached $1.9 billion, up 93 percent from a year earlier, with net income of $1.1 billion and full-year guidance raised to 82 percent growth. In the shareholder letter, chief executive Alex Karp wrote that there are "Marxist overtones and undertones" to the business, arguing that frontier labs persuade enterprises to migrate their intellectual property and expertise into someone else's model and then compete with them using it. Palantir's pitch is the inverse — model-agnostic software where the customer keeps the data and the outputs.

Researchers at MIT led by Marzyeh Ghassemi, with collaborators at Stanford and Columbia, tested skin-disease diagnosis with and without model assistance. Non-experts did get more accurate, but largely by deferring: they trusted the model's written explanations whether those explanations were right or wrong, and rated the vaguest ones as most convincing. Clinicians were not misled by incorrect assistance and performed best when handed a bare prediction with no explanation at all — which inverts the usual assumption that showing the reasoning makes medical AI safer.

The National Autonomous University of Mexico, known as UNAM, ran its entrance exam fully online for the first time, using a locked-down browser and behavior-monitoring software from Territorium Life that watched webcam feeds for wandering eyes, extra people in the room and background conversation. The scores came back nothing like history — about 16 percent of test takers cleared 100 correct answers out of 120, against roughly 3 percent from 2021 through 2025. A technical commission called for an in-person control exam, and roughly 58,000 applicants, about 36 percent of the field, must sit it again under human supervision.

GenOffice is a desktop suite — word processor, spreadsheet, slides and PDF, plus a shell that hosts them in tabs — for Apple Silicon Macs and Windows, with Genspark's agent working inside each document rather than in a side panel. The core really is Apache 2.0: commercial use is permitted, with no revenue threshold and no geographic carve-out. Two limits are worth reading before you deploy it — a reserved directory for future enterprise modules falls under a separate proprietary license, and the AI features consume Genspark credits through a cloud proxy, so the shipped alpha is open code but not a self-contained local stack.