AI Briefing

Today's theme is sovereignty and stack control: who owns the model, the silicon, and the data. Anthropic dragged the White House into a fight with Alibaba, accusing China's giant of quietly siphoning Claude to school its own Qwen models. Meanwhile OpenAI started building its own chips, ByteDance smashed the AI-video duration ceiling, and Mistral made a pitch that your documents should never leave the building. The most fun thread to pull: someone allegedly copied Anthropic's homework 28.8 million times.

01IndustryTom's Hardware ↗

Anthropic Tells the White House Alibaba Used 25,000 Fake Accounts to 'Distill' Claude Into Qwen

In a letter to White House officials and several U.S. senators, Anthropic alleged that Alibaba orchestrated roughly 25,000 fake accounts to query Claude about 28.8 million times between April and June 2026, then used the outputs to 'distill' Claude's capabilities into its own Qwen models. It is the first time a leading U.S. lab has publicly accused a major Chinese rival of large-scale model theft through ordinary API access. The claim escalates U.S.-China AI tensions and exposes how hard it is for any lab to stop a competitor from cloning behavior one prompt at a time.

02ProductOpenAI ↗

OpenAI and Broadcom Unveil 'Jalapeño,' OpenAI's First Custom AI Chip

OpenAI and Broadcom revealed Jalapeño, OpenAI's first 'Intelligence Processor' — a purpose-built ASIC designed specifically for LLM inference rather than repurposed training hardware. Co-developed from initial design to manufacturing tape-out in just nine months (with OpenAI's own models speeding up parts of the design), it is the first step in a multi-generation compute platform slated for initial deployment by end of 2026 with partners including Microsoft. The move is OpenAI's clearest signal yet that it wants to 'build the full stack' and reduce its dependence on Nvidia.

03ProductTech Times ↗

ByteDance's Seedance 2.5 Claims the First Native 30-Second AI Video Clip

At its Volcano Engine FORCE conference, ByteDance announced Seedance 2.5, claiming it can generate a single continuous 30-second clip natively — without stitching shorter segments together — while accepting up to 50 multimodal reference inputs for character and brand consistency. If the claims hold up under independent testing at the early-July public launch, it roughly doubles the practical native-clip ceiling of major commercial tools and reshapes the economics of short-form ad and social video. The catch: its predecessor drew cease-and-desist letters from every major Hollywood studio, and those copyright disputes remain unresolved.

04ProductTech Times ↗

Mistral Ships OCR 4, a Document AI That Runs Entirely on Your Own Servers

Paris-based Mistral released OCR 4, a document-intelligence model that returns paragraph-level bounding boxes, typed-block classification (titles, tables, equations, signatures) and per-word confidence scores across 170 languages — and can run in a single Docker container fully on-premise. The pitch is data sovereignty: unlike US-headquartered cloud OCR services exposed to the US CLOUD Act, Mistral's self-hosted option keeps sensitive documents inside a customer's own infrastructure. That positioning lands just as the EU AI Act's fine-enforcement provisions take effect in August 2026.