2026-06-01

1. Microsoft Build 2026 — Project Polaris: GitHub Copilot Without OpenAI ⚡

• Microsoft gave an early preview of Project Polaris, its in-house coding AI model. It will replace GPT-4 Turbo, GitHub Copilot's default model (GA: 2026-08). (source)
• Architecture: CoT + Tree-of-Thought reasoning to support complex multi-file refactoring
• Code Content Guarantee: trained exclusively on licensed data → IP-infringement indemnification (an industry first)
• Announced alongside: Windows Agent Framework open source, Azure Agent Mesh, Copilot Workspace GA, Foundry Local GA (on-device)
• GitHub Copilot AI-credit billing switches over starting today (6/1) — a billing redesign for the Polaris era
• Why it matters: "Claude Code has overtaken GitHub Copilot's market share" was reported as the direct trigger. Microsoft is breaking free from its dependence on OpenAI licensing, developing a frontier coding AI on its own and deploying it instantly through GitHub (tens of millions of developers) — the coding AI market is now reshaping into a three-way race among Anthropic (Claude Code), OpenAI (Codex), and Microsoft (Polaris).
• → Project Polaris (🆕 new), Microsoft, Agents (LLM Agents)

2. Anthropic Files Confidential IPO S-1 ⚡

• Anthropic submitted a confidential draft registration statement (S-1) for an IPO to the U.S. SEC. Ahead of OpenAI. (source)
• Recent valuation of $965B (2026-05-28 Series H), annualized revenue of $47B
• A listing could place it among the top tier of the S&P 500 (share count and offering price undecided)
• Why it matters: A public listing by an AI safety company is a fundamental shift in governance structure. How quarterly earnings disclosure plus pressure to maximize shareholder profit can coexist with a "safety-first" mission becomes a structural question for the AI industry. An IPO race with OpenAI and SpaceX — a signal marking the end of the AI startup era.
• → Anthropic

3. NVIDIA Cosmos 3: A Fully Open Physical AI Foundation Model ⚡

• NVIDIA unveiled Cosmos 3, the world's first fully open omnimodal physical AI foundation model. MIT license. (source)
• Cosmos 3 Super (64B): 32B reasoner + 32B generator (Mixture-of-Transformers). Unifies world generation + physical reasoning + action prediction into a single model
• Trained on 20 trillion tokens (~1 billion images, 400 million videos). #1 open model on R-Bench.
• Cosmos Coalition launched: global partners including Agile Robots, Black Forest Labs, Runway, Skild AI
• Use cases: synthetic data for robot training, autonomous-driving simulation, physical AI post-training
• Why it matters: The world foundation model — the equivalent of the LLM for the "physical AI era" — has appeared for the first time as fully open source (MIT). Together with DeepMind Gemini Robotics ER 1.6, the race to accelerate physical AI is now in full swing — meaning robotics/AV startups have a base model ready for immediate real-world use.
• → Cosmos 3 Super (🆕 new), NVIDIA, Embodied Agents