AI Sector Daily Digest — June 3, 2026

Today's five: Trump signs a voluntary AI security review order; Microsoft launches seven MAI models at Build 2026; Nvidia releases Nemotron 3 Ultra as the most capable open US model; 130+ mathematicians publish the Leiden Declaration warning AI threatens their discipline; and Alibaba ships Qwen3.7-Plus with full multimodal input.

President Trump signed an executive order on June 2 asking the country's leading AI developers — including OpenAI and Anthropic — to voluntarily submit their most capable models to government cybersecurity reviewers up to 30 days before public release. 1 The order, titled "Promoting Advanced Artificial Intelligence Innovation and Security," directs NIST and relevant national security agencies to develop testing procedures for frontier models. Participation is voluntary, not mandatory, a design the White House chose deliberately to avoid deterring investment. 2 This is a slimmed-down version of an order the administration shelved in May after internal disagreements; the original draft included binding disclosure requirements that AI labs opposed. 3

At its annual developer conference in Seattle, Microsoft announced a new MAI model family comprising seven models it developed internally — the clearest signal yet that the company is building its own AI stack alongside its OpenAI partnership. 4 The headline model is MAI-Thinking-1, a reasoning model; MAI-Code-1 is tuned for GitHub and VS Code and is already available inside Copilot; MAI-Transcribe-1.5 outperforms both Gemini and OpenAI's transcription models on internal benchmarks; and MAI-Image-2.5 handles image generation. 5 Microsoft also unveiled Project Solara, a new agent-first OS platform for devices that run AI agents without a traditional app-based interface. The company confirmed it is incorporating Anthropic's Claude models into Microsoft 365 alongside OpenAI, a change from its historically exclusive arrangement. 6

Nvidia announced Nemotron 3 Ultra at Computex 2026 in Taipei — a 550B-parameter mixture-of-experts model with 55B active parameters at inference time. 7 It scores 48 on the Artificial Analysis Intelligence Index, making it the highest-rated open-weights model built by a US company. The benchmark leader remains Moonshot AI's Kimi K2.6, a Chinese open model that scores 54 on the same index. 8 Nemotron 3 Ultra is designed for long-running enterprise agents — coding, research, and multi-step workflows — and is built on Nvidia's NeMo framework. The model becomes available this month via the Nvidia API catalog and is also integrated into Microsoft's Windows AI stack. 9

A group of 16 mathematicians from 15 universities, led by Jim Portegies of Eindhoven University of Technology, published the Leiden Declaration on AI and Mathematics on June 2. 10 More than 130 researchers have signed it. The 11-page declaration raises concerns about AI-generated proofs that are not independently verifiable, the risk of tech industry influence over research funding and publication, and the accuracy and reliability of AI systems in formal reasoning. The timing is pointed: a week earlier, an internal OpenAI model produced a solution to the Erdős unit-distance conjecture, an 80-year-old open problem, prompting questions about closed access to AI-derived mathematics. 11 The declaration calls on journals to require disclosure of AI use, funders to protect independent mathematical research, and institutions to maintain human-led peer review. 12

Alibaba released Qwen3.7-Plus on June 2, a multimodal model that adds image and video understanding to the Qwen3.7 family. 13 Priced at $0.40 per million input tokens and $1.60 per million output tokens, it supports deep reasoning and tool invocation via Alibaba's Bailian platform. The model ranked 16th on Vision Arena, placing Alibaba fifth among model providers globally in vision benchmarks. One notable constraint: it is proprietary and cannot generate images or video, only interpret them. 14 It is not available as open weights, a departure from earlier Qwen releases.