The hearing started. Judge Rita F. Lin began hearing arguments today in San Francisco in Anthropic’s lawsuit challenging the Pentagon’s “supply chain risk” designation—the blacklisting that blocks Anthropic from military contractor work and puts hundreds of millions to billions in 2026 revenue at risk. Anthropic’s position: it refused to remove safety guardrails for surveillance and fully autonomous weapons systems. The Pentagon’s position: those refusals constitute an unacceptable national security risk. Court filings revealed the Pentagon told Anthropic the two sides were “very close” on the disputed issues just days before publicly cutting ties. Microsoft and 22 former high-ranking U.S. military officials filed briefs supporting Anthropic. The case turns on whether a company’s safety policies can be treated as a supply chain deficiency—and that answer will shape how every AI lab negotiates government contracts from here.
While the courtroom argues over who controls AI in the military, the developer market is grappling with who controls AI in the code editor. Cursor confirmed that Composer 2, its new coding model launched March 19, is built on Moonshot AI’s Kimi K2.5—a Chinese foundation model—after initially declining to disclose the underlying architecture. Cursor has over 1 million daily active users and 50,000+ business customers. Composer 2 pricing sits at $0.50/$2.50 per million input/output tokens on standard and $1.50/$7.50 on fast, and it scored 61.7% on Terminal-Bench 2.0—outperforming Claude Opus 4.6 at 58%. The largest U.S. AI coding tool quietly ran on Chinese-built infrastructure, and the disclosure only came after external pressure. Supply chain transparency isn’t just a Pentagon concern anymore—it’s a developer tooling question that affects every company writing code through an AI intermediary.
The infrastructure underneath all of it is scaling on two fronts. Gimlet Labs raised an $80 million Series A led by Menlo Ventures—$92 million total—for what it calls the first multi-silicon inference cloud, running workloads across NVIDIA, AMD, Intel, ARM, Cerebras, and d-Matrix chips simultaneously. The company claims 3–10x faster inference at equivalent cost, already has eight-figure revenues, and counts a top-3 frontier lab and a top-3 hyperscaler among its customers. On the power side, NVIDIA and Emerald AI partnered with AES, Constellation, Invenergy, NextEra, Vistra, and Nscale to build power-flexible AI factories that adjust consumption based on grid demand—a design that could unlock up to 100 GW of flexible capacity if adopted nationwide, with Nscale targeting 2 GW of data center capacity by 2027 scaling to 8 GW by 2030 and a 96 MW pilot at NVIDIA’s Aurora AI Factory in Virginia going live mid-2026. One startup is making chip choice irrelevant for inference; another consortium is making data centers part of the grid solution instead of the grid problem. The bottleneck is shifting from compute to coordination.
That’s Tuesday. Four stories, zero fluff.
— The PulseMark Team
Get the Daily Pulse
Sharp analysis on what's actually moving in AI. No hype, no filler, no weekly digest.