The AI Brief

Blackstone is in early talks with institutional investors about a second mega private credit package of at least $36 billion to finance Anthropic's use of Google's custom TPU chips, Bloomberg reported August 5. If finalized near that figure, the deal would exceed the $35 billion arrangement Apollo Global Management and Blackstone closed roughly two months ago, itself described by Bloomberg as among the largest private credit transactions in history. The size, structure, and whether Blackstone ultimately leads the deal remain under discussion and may change, per people familiar with the matter who asked not to be identified.

The earlier $35 billion deal was structured through an SPV that purchased Google TPUs and leased them back to Anthropic, keeping the debt off Anthropic's own balance sheet. That debt was divided into approximately $6 billion of A1 notes, $25 billion of A2 notes, and $4.5 billion of B notes; Broadcom provided residual-value support on the senior tranches, meaning Broadcom covers any shortfall for A1 and A2 investors if a chip resale fails to cover what is owed. Apollo, Blackstone, and Broadcom assembled the original deal through a partnership called the AI XPV Platform. The new financing proposal arrives as Anthropic has filed confidentially for a US IPO, targeting a listing as early as October, with Morgan Stanley, Goldman Sachs, and JPMorgan working on the process. Representatives for Blackstone, Apollo, Anthropic, and Google declined to comment.

The proposed second package would arrive in stages tied to chip deliveries. The first $35 billion deal included roughly 16 separate draws over just over a year, with portions of the debt available to institutional buyers via the 144A market by early 2027. The new round implies Anthropic's TPU commitments are growing faster than the first facility can cover, ahead of its pre-IPO investor roadshow.

Disclosure: Claude, which generates this brief, is built by Anthropic.

Koray Kavukcuoglu, who took over as Google DeepMind CEO after Demis Hassabis stepped back to become Alphabet chief scientist and DeepMind chair, disclosed the 950 million monthly active user figure in his first all-hands message to staff, which was published August 8 on the Google DeepMind blog. The message described the Gemini app as "rocketing to over 950M monthly users" and characterized DeepMind as "the AI engine room of Google," with Gemini delivering AI Mode, AI Overviews, and the consumer assistant globally.

The figure lands against a backdrop of continued Gemini 3.5 Pro absence. The model, announced at Google I/O in May and originally promised "next month", has now missed every public and leaked target since June. As of August 8, no model card, no API model ID, and no pricing have been published. Prediction market Kalshi has more than $320,000 wagered on the launch window; traders assign a 46% probability to a release before August 16, 63% before August 21, and 76% before the end of August. Google's Made by Google hardware event in New York City on August 12, scheduled to introduce the Pixel 11 series, is now the next plausible launch window, though Google has not confirmed whether Gemini 3.5 Pro will appear alongside the hardware. A model matching its description appeared in LM Arena under the name "gemini-3.5-flash" in early August, suggesting deployment infrastructure is ready.

Anthropic published the self-hosted environments announcement on August 6, 2026. The feature, now in public beta for Claude Team and Enterprise plans, lets a platform or developer-experience team configure named environments in the claude.ai admin console. Sessions started from the web, mobile, desktop, the terminal, or a scheduled routine then execute on the organization's own infrastructure rather than Anthropic's servers. The feature supports both fixed runners (always-on machines) and on-demand runners (containers spun up per session). It ships off by default.

Anthropic draws a sharp line between self-hosted environments and Remote Control, an existing feature that lets developers continue a session running on their own laptop from a phone. Remote Control sessions are tied to the individual user and end when the host machine stops; self-hosted environments run on shared infrastructure a platform team operates and are accessible to any user in the organization. The August 6 release also coincided with Claude Code v2.1.224, which added cross-session messaging via a new SendMessage API, removed the subagent spawn cap, and patched a permission-preview flaw where tool inputs containing bidirectional-override characters could manipulate terminal display.

The launch follows a concentrated enterprise push: on August 5, Anthropic added inline DLP (data loss prevention) hooks to Claude Enterprise, letting compliance teams inspect and block prompts and tool calls before they reach the model, and separately confirmed it is building an in-house silicon team to design custom AI chips for Claude alongside its AMD, AWS, Google, and Nvidia supply.

Disclosure: Claude, which generates this brief, is built by Anthropic.

OpenAI filed a Motion to Dismiss Apple's trade-secrets lawsuit on August 6, 2026, following a blog post published the prior evening titled "Apple is getting this wrong." Apple sued OpenAI in July, alleging that Chief Hardware Officer Tang Tan and engineer Chang Liu systematically transferred confidential product designs and manufacturing secrets when they joined OpenAI from Apple. Apple subsequently sought a preliminary injunction to bar both employees and OpenAI from accessing the alleged confidential information during litigation.

OpenAI's public response centers on two procedural arguments and a factual dispute. On procedure: OpenAI asserts that Apple's outside lawyers "emailed the wrong person after confusing two Asian last names" when attempting to contact OpenAI in February, an error Apple acknowledged only after OpenAI raised it. On the merits: OpenAI published iMessage transcripts between Liu and former Apple colleagues that it argues show Liu helping those colleagues locate files in iCloud Drive after his departure, not exfiltrating Apple confidential material. Apple's motion, however, separately documents 37 documents Liu downloaded from a third-party cloud provider (not iCloud) on five occasions through April 27, 2026, three months after his last day, a pattern the published transcripts do not address. Independent analysis by Daring Fireball noted the transcripts explain iCloud activity Apple says is not in dispute, while leaving the Box-platform downloads at the center of Apple's case unexplained.

The litigation runs parallel to OpenAI's AI hardware development, the company's first foray into consumer devices. Tang Tan, formerly Apple's VP of Product Design for iPhone and Apple Watch, was hired to lead the effort. Apple's preliminary injunction, if granted, could prevent him from working on the project while the case proceeds.

First covered in Vol. I, No. 74. Google holds its Made by Google event in New York City on August 12 to launch the Pixel 11 smartphone series, the first hard date on Google's calendar since Gemini 3.5 Pro missed its August 7 window, which prediction markets had assigned 73% probability. Kalshi traders have now placed more than $320,000 on the launch timeline. As of August 8, they assign a 46% chance of launch before August 16, rising to 63% before August 21 and 76% before August 31. Google has confirmed no release date, pricing, or features.

Community leakers report that a model matching Gemini 3.5 Pro's description appeared in LM Arena under the identifier "gemini-3.5-flash" in early August, and that enterprise partners have been quietly testing early builds. Leaked performance notes describe improvements in frontend generation, SVG generation, and multi-file coding, with an updated Deep Think mode that breaks down complex queries before answering. Google has not confirmed any of this. The DeepMind blog's first Kavukcuoglu message, published this morning, makes no mention of Gemini 3.5 Pro's launch timeline.

The broader context: Google DeepMind's new CEO has inherited a Gemini pipeline where the flagship model has missed five consecutive launch targets and Gemini 3.5 Flash, while capable, trails on hard reasoning benchmarks and the long-context retrieval tasks Pro is specifically designed for. SemiAnalysis published a detailed analysis August 7 arguing "Gemini is Cooked but GCP is Cooking," noting that Google's inability to retain AI talent, Jeff Dean, Quoc Le, Oriol Vinyals, and Sanjay Ghemawat all left in August, compounds the model-quality gap.