AI this week: Anthropic vs. the Pentagon, OpenAI kills Sora, and the White House unveils its AI playbook

This week saw courtroom drama between Anthropic and the Pentagon, OpenAI killing off Sora and raising record capital, the White House laying out its AI regulatory vision, and NVIDIA doubling down on enterprise AI agents at GTC. Here are the 10 most important AI stories from the past seven days.

1. Federal judge calls Pentagon’s Anthropic blacklisting “troubling”

Source: NPR

U.S. District Judge Rita F. Lin heard arguments on March 25 in Anthropic’s lawsuit challenging the Pentagon’s decision to designate the company a “supply chain risk” after it refused to allow unrestricted military use of Claude. The judge said the government’s actions appeared aimed at crippling the company rather than addressing genuine security concerns, and indicated she would rule on the preliminary injunction before the end of the week. Over 30 employees from OpenAI, Google DeepMind, and Microsoft filed briefs in Anthropic’s support.

2. OpenAI shuts down Sora video app, cancels Disney deal

Source: TechCrunch

OpenAI announced on March 24 that it is discontinuing its Sora video generation app, including the iOS app, API, and Sora.com. The shutdown also torpedoes a landmark $1 billion deal with Disney that never closed. The decision reflects OpenAI’s pivot toward enterprise products and infrastructure as compute resources remain scarce across the industry. Sora’s research team will refocus on world simulation for robotics applications.

3. OpenAI’s funding round hits $120 billion as Altman shifts focus to infrastructure

Source: CNBC

OpenAI CFO Sarah Friar confirmed that the company is raising an additional $10 billion, bringing its record-setting funding round to roughly $120 billion and valuing the company at about $850 billion. Sam Altman told staff he is stepping back from direct oversight of safety and security teams to focus on capital, supply chains, and “building data centers at unprecedented scale.” The company also revealed it has completed initial development of its next major model, codenamed “Spud.”

4. White House releases national AI policy framework, targets state regulation

Source: CNN

On March 20, the White House released a legislative blueprint urging Congress to adopt federal AI regulation that would preempt state-level AI laws deemed “unduly burdensome.” The framework’s seven pillars cover child protection, intellectual property, free speech, innovation, workforce development, and federal preemption. It also proposes conditioning $42 billion in broadband infrastructure funding on states repealing certain AI regulations — a controversial mechanism that has drawn pushback from state attorneys general.

5. NVIDIA launches Agent Toolkit and NemoClaw at GTC 2026

Source: VentureBeat

NVIDIA used GTC 2026 to unveil its Agent Toolkit, an open platform for building and deploying autonomous enterprise AI agents. The toolkit includes Nemotron open models for agentic reasoning, AI-Q for enterprise knowledge integration, OpenShell for policy-based security, and cuOpt for optimization. Seventeen major enterprises including Adobe, Salesforce, SAP, and Cisco are early adopters. The hybrid architecture can cut query costs by more than 50% by routing between frontier and open models.

6. NVIDIA Nemotron 3 Super leads open-weight coding benchmarks

Source: Decrypt

Also at GTC, NVIDIA released Nemotron 3 Super, a 120-billion-parameter hybrid Mixture-of-Experts model with only 12 billion active parameters per forward pass. It scores 60.47% on SWE-Bench Verified (versus GPT-OSS’s 41.90%) and 91.75% on RULER at 1M tokens, making it the top open-weight model for real-world coding tasks. The release is part of NVIDIA’s $26 billion investment in open-model AI infrastructure.

7. Google ships major Stitch update, Figma shares drop

Source: Google Blog

Google released a significant redesign of Stitch, its free AI design tool, on March 19. The update adds an infinite canvas, a voice interaction mode that can interview designers about their goals and offer real-time critiques, and integrations with coding tools like Claude Code and Cursor. With 350 free design generations per month, Stitch is positioning itself as a serious alternative to traditional design tools — Figma shares reportedly dropped 8–10% in the two days following the announcement.

8. Alibaba’s Qwen 3.5 Small punches above its weight

Source: BuildFastWithAI

Alibaba released Qwen 3.5 Small on March 1, a family of four dense models (0.8B to 9B parameters) that are all natively multimodal — supporting text, images, and video — and licensed under Apache 2.0. The 9B model scores 81.7 on GPQA Diamond versus GPT-OSS-120B’s 71.5, outperforming a model 13 times its size on graduate-level reasoning. The models support a 262,000-token context window and multi-token prediction.

9. Transformer architecture research treats depth as a retrieval problem

Source: Turing Post

On March 16, two major Chinese AI labs — Kimi Team (Moonshot AI) and ByteDance Seed — independently published papers proposing that Transformer depth should be treated as an addressable, searchable dimension rather than a passive stack. Both papers diagnose the same problem (useful information from earlier layers gets diluted in deep networks) and propose letting attention mechanisms query across layers, not just across tokens. This represents a potentially significant shift in how Transformer architectures handle depth.

10. DOE announces $294 million in AI funding under Genesis Mission

Source: Inside Government Contracts

On March 17, the Department of Energy issued a Request for Application under the Genesis Mission, a White House-led AI initiative first announced in November 2025, with $293.76 million in anticipated total funding. The program aims to accelerate AI research and deployment across the national laboratory system, representing one of the largest federal commitments to AI infrastructure outside the defense sector.