Everyone is building their own AI stack

Arthur Mensch told CNBC that Mistral is exploring designing its own chips as it ramps up infrastructure. On the same day, the Wall Street Journal reported Mistral is accelerating superintelligence work for European independence from US tech giants, with supply deals signed for Airbus and BMW. Two stories, one posture: own the stack or rent it.

Mistral accelerates superintelligence work, signs Airbus and BMW

WSJ via Techmeme

The companion to today's chip story. Mistral is framing the speed-up as European independence from US tech giants, and the Airbus and BMW deals give the framing real customers.

SpaceX writes its own AI training stack, in C, against 220k GB300s

@elonmusk

Musk says SpaceX is closing in on version 1.0 of an in-house training stack written in C that maps directly to 220,000 GB300s with 800G NICs, leaning on pipeline parallelism and bare metal.

“SpaceX has almost finished writing V1.0 of an in-house AI training stack in C that exact-maps to 220k GB300s with 800G NICs, making heavy use of pipeline parallelism and getting as close to bare metal as possible.”

ByteDance is designing its own CPUs

Reuters via Techmeme

Reuters reports ByteDance is moving into custom CPU design to support its AI infrastructure as chip prices and supply shortages constrain expansion plans.

Kirkland & Ellis sets aside $500M for its own AI platform

FT via Techmeme

The world's highest-grossing law firm is opting out of the shared-vendor pool and building a proprietary stack rather than use what its rivals use.

IBM and Red Hat commit $5B and 20,000 engineers to Project Lightwell

WSJ via Techmeme

A new open-source software model with major corporate scaffolding behind it: 20,000 engineers deployed globally and AI in the loop. Worth tracking how the licensing and governance shake out.

Corporate America enters its AI reckoning

Axios

Madison Mills reports that corporate leaders who rushed to embrace AI are starting to question whether the spending is delivering returns. A useful counterweight to the build-your-own headlines above.

Taiwanese tech companies have raised a record $14.5B of debt this year

Bloomberg via Techmeme

Aileen Chuang's reporting on the financing scramble. The capacity build is being underwritten by debt at a pace that itself becomes a story.

Exchanges are designing futures contracts for AI compute

Reuters via Techmeme

The Shanghai Futures Exchange is in early design on contracts for AI tokens; US exchanges are set to launch GPU compute futures. Compute starts behaving like an asset class.

Coding agents have found product-market fit

Simon Willison via Techmeme

Willison's read: Anthropic and OpenAI's coding agents are becoming daily drivers for highly paid professionals. A concise framing of where the agentic-coding wave actually landed.

Grok's harness reads Claude Code's CLAUDE.md on first launch

@ShenHuang

Shen Huang reports that on first open, Grok Build picks up Claude Code's config file along with its skills directory. It skips the Codex equivalent. The agent-config format wars are now interoperability decisions.

“The first time you open Grok, it will directly read Claude Code's CLAUDE.md + skills and so on. But it won't read Codex's AGENTS.md.”

On AI agents with production database access

@leilavclark

Leila Clark defends Jared's claim that agents are productive with prod DB access, drawing the line at vibe coders who don't understand databases. A clean version of the line the operator community is currently drawing.

Five frontier large language models disagree on 67% of 1,000 fact-check claims

lenz.io via Hacker News

A small but pointed study: ask five frontier models to label the same 1,000 atomic claims as True / Mostly True / Misleading / False, and they disagree on two-thirds. Simon Willison surfaced the exact prompt in the thread.

Qwen ships Q-Judger: a VLM fine-tuned to grade text-to-image output

Hugging Face

An open vision-language judge model for automated text-to-image evaluation. Useful if you're building image pipelines and tired of relying on closed graders.

Median MTS theorem

@Miles_Brundage

A one-line claim from Miles Brundage worth saving for later arguments about lab governance.

“Median MTS theorem: a frontier AI company's policy positions eventually converge to those of the median member of technical staff there”