Briefing · Thursday, June 25, 2026

Good morning. It's Wednesday, June 25, and we're covering OpenAI taping out its first custom silicon, Anthropic accusing Alibaba of extracting Claude's capabilities, Google adding computer use to Gemini Flash, and Qualcomm buying Modular for $4 billion.
The OpenAI chip story hit 705 points on HN. The Anthropic-Alibaba thread drew 749 comments and counting. Here is the signal, sourced.
In today's brief:
THE BIG ONE
OpenAI announced (705 points on HN) a custom inference chip built with Broadcom and manufactured by TSMC. The company claims the chip went from concept to tapeout in nine months, with OpenAI models "accelerating parts of the design and optimization process."
The HN discussion remained skeptical. A chip CEO in the thread noted nine months RTL-freeze to tapeout is "fairly typical, even somewhat unimpressive" for a complex 3nm design. The community parsed what "AI-assisted design" actually means - engineers confirmed they use LLMs for testbench generation and documentation, but called these incremental improvements rather than revolutionary breakthroughs. Broadcom handled physical design, IP blocks, and supply chain, meaning OpenAI's contribution focused on logic specifications.
Why it matters: OpenAI joins Google TPU and Nvidia's Vera Rubin in the custom silicon race. The timeline claims warrant scrutiny, but the strategic direction is clear - inference costs drive vertical integration.
SECURITY
Anthropic alleged (433 points, 749 comments on HN) that Alibaba "illicitly extracted Claude AI model capabilities" through model distillation - querying Claude's API at scale and using the responses to train competing models. The extraction reportedly targeted valuable post-training behaviors like reasoning chains and tool usage patterns.
The HN thread turned sharply critical of Anthropic. Commenters highlighted that Anthropic trained Claude on "over seven million books from pirate sites like LibGen" according to settlement details. The hypocrisy argument gained traction: "You're trying to kidnap what I've rightfully stolen." Others framed the complaint as lobbying for export controls rather than genuine IP protection, noting that distillation is essentially indistinguishable from API evaluation.
Simon Willison covered adjacent territory on June 24, quoting Tom MacWright on how LLM-generated portfolios reveal "nothing about this person, other than that they use particular tools."
Why it matters: If distillation becomes a standard practice, API providers face a choice between restricting access and accepting that outputs train competitors. The legal and ethical lines remain contested.
PLATFORMS
Google shipped computer use capabilities in Gemini 3.5 Flash (219 points on HN), enabling the model to interact with computer interfaces via screenshots and controls. The feature allows automated task completion across web applications and desktop environments.
Benchmark data shows Gemini 3.5 Flash performing comparably to Claude Sonnet 4.6 but trailing Claude Opus 4.8 and GPT-5.5 on OSWorld evaluations. The discussion ran negative - users reported the model abandoning tasks after errors and overly strict safety guardrails blocking routine queries. One user praised it as "3x cheaper than 5.5" for agentic workflows. Missing MCP support in the Gemini app drew multiple complaints.
Why it matters: Computer use is becoming table stakes for frontier models. Google's entry at Flash pricing creates pressure on the premium tier, but execution quality determines adoption.
ACQUISITIONS
Qualcomm announced (198 points on HN) an all-stock acquisition of Modular, the AI compute startup led by LLVM creator Chris Lattner. The deal values Modular at approximately $4 billion.
The discussion focused on the Mojo programming language's future. Qualcomm committed to open-sourcing Mojo's compiler this year, though commenters questioned whether a chip company will prioritize software development. The acquisition fits Qualcomm's broader play to build AI stacks independent of NVIDIA's CUDA ecosystem, joining their Ventana and Tenstorrent investments.
Why it matters: If Qualcomm executes, Mojo becomes the first serious CUDA alternative with major corporate backing. If they don't, another promising compiler project gets acqui-hired into obscurity.
TOOLS WORTH A LOOK
RubyLLM (391 points) - Unified Ruby framework for OpenAI, Anthropic, Gemini, DeepSeek, and 10+ other providers. Chat, vision, audio, embeddings, and tool calling through one interface. (OSS)
Krea 2 (383 points) - Open-weights 12B text-to-image model ranking top 10 on Artificial Analysis. Diffusion transformer with Qwen 3 VL encoder. Permissive license. (OSS)
Nub (241 points) - Rust-based Node.js toolkit that layers productivity features without replacing the runtime. 24x faster script running than npm, 2.5x faster installs than pnpm. (OSS)
GLM-5.2 (242 points) - Z.ai's open-weight model matches Opus 4.8 on agent benchmarks at max thinking effort. MIT licensed. (OSS)
WHAT ELSE IS HAPPENING
FROM THE SITE
Simon Willison published browser-compat-db on June 24 - Mozilla's browser compatibility data converted to SQLite using Claude Code, hosted on GitHub CDN with open CORS headers for Datasette Lite exploration.
Every link above goes to a primary source or our sourced coverage. Tomorrow's brief lands when the news does - subscribe to get it by email.
The daily brief, delivered. Free, unsubscribe anytime.