AI News Roundup: White House AI Framework, NVIDIA Rubin Ships, Anthropic’s Pentagon Fight Escalates

White House Unveils National AI Legislative Framework

On March 20, the Trump Administration released its most comprehensive AI policy document yet: a national legislative framework built on seven pillars that it wants Congress to codify “this year.” The framework covers child safety, intellectual property, free speech, workforce development, and — most controversially — a push to preempt state AI regulations in favor of a single federal standard.

The White House explicitly rejected creating a new federal AI rulemaking body, instead proposing that existing agencies regulate AI within their sectors. On copyright, the administration took a firm stance: it believes AI companies scraping the internet for training data does not violate U.S. copyright law, though it acknowledged the courts will have the final word. The framework also calls for streamlined permitting for AI data centers while preventing electricity cost increases for surrounding communities.

Industry reaction has been mixed. AI companies largely welcomed the preemption of a patchwork of state laws, while civil-society groups warned that the “light-touch” approach could leave consumers without meaningful protections.

NVIDIA Rubin Platform Enters Production — 10x Performance Per Watt

At GTC 2026 on March 16, NVIDIA CEO Jensen Huang unveiled the Vera Rubin platform, a six-chip architecture that the company says will cut AI inference costs by a factor of ten compared to Blackwell. Built on TSMC’s 3 nm process with 336 billion transistors and 288 GB of HBM4 memory per GPU, Rubin delivers 50 petaflops in FP4 — up from 20 petaflops on Blackwell.

The platform is now in full production, with partner products shipping in the second half of 2026. Huang also revealed that NVIDIA has secured $1 trillion in combined orders for Blackwell and Vera Rubin through 2027. Microsoft announced it was the first hyperscale cloud provider to power up Vera Rubin NVL72 systems in its liquid-cooled Azure data centers.

Anthropic’s Pentagon Fight Draws Broad Support

The legal battle between Anthropic and the Department of Defense escalated significantly this month. After the Pentagon designated Anthropic a “supply-chain risk” — effectively banning Claude from government use — the AI company filed suit in both a California district court and the D.C. federal appeals court. The dispute centers on two red lines Anthropic refused to cross: allowing its AI for mass surveillance of Americans or autonomous weapons.

Nearly 150 retired federal and state judges filed an amicus brief supporting Anthropic, joining Microsoft and more than 30 employees from OpenAI and Google DeepMind who signed statements in their personal capacities. Anthropic’s CFO estimated the designation could reduce the company’s 2026 revenue by “multiple billions of dollars.” Meanwhile, Google has been quietly expanding its own Pentagon work while the public feud between Anthropic CEO Dario Amodei and OpenAI CEO Sam Altman intensifies.

Roche Deploys Pharma’s Largest AI Factory

Roche announced on March 16 that it has deployed 2,176 NVIDIA Blackwell GPUs on premises across the U.S. and Europe, bringing its total GPU footprint to more than 3,500 — the largest announced AI infrastructure in the pharmaceutical industry. The “AI factory” will accelerate drug discovery through NVIDIA’s BioNeMo platform, optimize manufacturing, and advance diagnostics development.

The deployment marks a deepening of a partnership that began in 2023 and eclipses Eli Lilly’s previously announced AI supercomputer. Roche is using a “Lab-in-the-Loop” approach that connects biological and chemistry experiments directly with AI models for faster iteration.

OpenAI’s GPT-5.4: A Million Tokens and Native Computer Use

OpenAI shipped GPT-5.4 on March 5, combining a 1-million-token context window, native computer-use capabilities, and full-resolution vision into a single model. For the first time, an OpenAI model can navigate desktops, control browsers, and execute multi-step workflows autonomously — not by describing what to do, but by actually doing it. The model scored 75% on the OSWorld-V benchmark, which simulates real desktop productivity tasks.

EU Moves to Simplify AI Rules

The European Council agreed its position on a proposal to streamline AI regulations as part of its “Omnibus VII” simplification package on March 13. The move signals a shift in tone from Brussels, which has faced criticism that the EU AI Act’s compliance burden is pushing AI startups to other jurisdictions. The UK, meanwhile, reversed course on its own AI training regulations after the proposal was “rejected by most respondents” in a public consultation.