$510M in smuggled AI chips

The supply chain just got indicted. Federal prosecutors charged Supermicro co-founder Wally Liaw and two others with smuggling $510 million in servers containing export-controlled Nvidia AI chips to China through shell companies in Southeast Asia. Supermicro stock dropped over 25% in pre-market trading. Separately, the DOJ filed its response in the Anthropic blacklisting case, arguing that the company’s AI ethics policies constitute “conduct, not speech” and defending the Pentagon’s supply chain risk designation. Nearly 150 retired judges filed an amicus brief supporting Anthropic, and a hearing before Judge Rita F. Lin is set for March 24. Anthropic has multiple billions in 2026 revenue at risk. One indictment targets chips flowing out of the country; the other targets a company that refused to remove human oversight from weapons systems. Export controls and procurement blacklists are now the same policy toolkit applied in opposite directions.

While Washington litigates who can build AI and where, the labs are quietly buying the developer infrastructure that determines how it gets built. OpenAI announced it will acquire Astral—the company behind uv and Ruff, tools installed by several million Python developers—giving Codex a package manager that runs 10–100x faster than pip. Codex already has over 2 million weekly active users with 3x user growth in 2026; Anthropic made a parallel move in December 2025 when it acquired Bun, the JavaScript runtime. The pattern is explicit: Python tooling goes to OpenAI, JavaScript tooling goes to Anthropic, and every developer workflow gets one step closer to a single vendor. Britannica and Merriam-Webster are suing OpenAI over nearly 100,000 articles, but the novel claim isn’t copyright—it’s a Lanham Act trademark argument that ChatGPT hallucinations falsely attributed to Britannica constitute trademark infringement. If that theory holds, every AI company that generates citations becomes liable not just for copying content but for inventing it under someone else’s name. The labs are acquiring infrastructure faster than the legal system can define what they owe for the data already inside the models.

And the models themselves are getting harder to ignore at any price point. MiniMax shipped M2.7 this week—a model that autonomously handled 30–50% of its own reinforcement learning research workflow during development. It scores 56.22% on SWE-Pro and 57.0% on Terminal Bench 2, with a 205K context window, priced at $0.30/$1.20 per million input/output tokens. For context, that output price is roughly 20x cheaper than Claude Opus 4.6 and 12x cheaper than GPT-5.2. A model that participated in its own training pipeline, priced for volume, with competitive coding benchmarks. The gap between frontier capability and frontier pricing continues to narrow—and the labs charging 20x more need an answer beyond benchmark tables.

Latest from PulseMark

Gemini Embedding 2 Multimodal RAG: Build Cross-Modal Search 5 modalities in one API call, 768-dimension vectors that save 75% storage, 70-second audio chunking windows, and the preprocessing pipeline you can finally delete. Read the tutorial →