Microsoft Build was the week's centrepiece, and the message was unmistakable: Redmond is building its own model stack. The headline release was MAI-Thinking-1, Microsoft's first reasoning model — 35 billion active parameters in a trillion-parameter mixture-of-experts architecture, trained entirely on commercially licensed data with no distillation from OpenAI. AI chief Mustafa Suleyman claimed that after tuning the model for McKinsey, it outperformed GPT-5.5 on quality at roughly ten times better cost efficiency. Alongside it came MAI-Code-1 (already live in GitHub Copilot), MAI-Image-2.5, MAI-Transcribe-1.5 covering 43 languages, and MAI-Voice-2. Microsoft also unveiled Project Polaris, the model that will replace GPT-4 Turbo as Copilot's default engine starting August — a mixture-of-experts architecture with specialised sub-modules for different programming languages. For anyone building on Microsoft's stack, the takeaway is clear: the future of Copilot runs on Microsoft models, not OpenAI's. The broader move — Windows as an agent runtime, Azure Agent Mesh for orchestration, Project Solara for adaptive desktops — positions Windows itself as the operating system for AI agents, not just the platform you use them on.

Anthropic filed a confidential S-1 with the SEC on Sunday, formally beginning the IPO process just four days after closing its $65 billion Series H at a $965 billion valuation. The listing is expected around October, ahead of OpenAI's reported September target. "This gives us the option to go public after the SEC completes its review," the company said. The filing puts the two largest private AI companies on a collision course toward public markets within weeks of each other — a dynamic that will force investors to make direct comparisons for the first time. Anthropic's run-rate revenue is above $47 billion and it just posted its first operating profit. OpenAI has 900 million weekly users and is approaching $2 billion in monthly revenue. The question isn't whether there's demand — it's which story Wall Street prices higher.

Nvidia used Computex in Taipei to enter the PC market outright. Jensen Huang unveiled the RTX Spark superchip — an Arm-based platform co-developed with Microsoft and MediaTek that pairs a 20-core Grace CPU with a Blackwell RTX GPU carrying 6,144 CUDA cores and up to 128 GB of unified memory, delivering one petaflop of AI performance. More than 30 laptops and 10 desktops from Dell, HP, Asus, Lenovo, and MSI will ship this autumn. Huang also announced the Vera CPU, designed specifically for AI agent workloads and claiming 1.8 times faster task completion than x86 chips, and laid out a multi-generation Spark roadmap running through Rubin and Rosa Feynman architectures. For businesses weighing where to run local AI — inference, fine-tuning, agent hosting — the RTX Spark makes the case that serious AI compute belongs on the desktop, not just in the cloud. It also signals that the laptop refresh cycle ahead will be driven by AI workloads in the same way mobile was driven by the camera.

On the policy front, Congress produced the most comprehensive federal AI bill to date. Representatives Jay Obernolte (R-CA) and Lori Trahan (D-MA) released a 269-page bipartisan discussion draft called the Great American Artificial Intelligence Act, which would create a federal AI governance framework, establish the Center for AI Standards and Innovation with $100 million per year in funding, require frontier developers to publish safety plans before model releases, protect AI whistleblowers, and — most controversially — preempt state AI laws for three years. The preemption clause is the fulcrum: it would freeze the patchwork of 1,200-plus state bills that Fortune documented last month, giving federal standards time to take hold but stripping states of their ability to experiment. Public Citizen called it a giveaway; Bloomberg Law called it necessary. Either way, it's the first serious attempt to answer the structural question that's been building all year — who actually regulates AI in the United States? SpaceX, meanwhile, began its IPO roadshow on Wednesday, pricing at $135 per share for a $1.77 trillion valuation and a planned June 12 listing on Nasdaq — which would make it the largest IPO in history.

And a few stories that say something about where value is moving. Suno, the AI music platform, raised $400 million at a $5.4 billion valuation — still facing copyright suits from Universal and Sony, but now with a Warner Music licensing deal and over 2 million subscribers. It's one of the clearest examples of AI creating an entirely new category of consumer product, even as the legal framework around it remains unresolved. OpenAI announced it will grant EU institutions access to GPT-5.5-Cyber, expanding its defensive cybersecurity tools into European governments and businesses — a distribution play that has no equivalent from any competitor yet. And Colorado, whose original AI Act was set to take effect June 30, signed a replacement bill that strips out most of the original compliance obligations and pushes the effective date to January 2027 — a real-time illustration of how the regulatory ground keeps shifting under companies trying to build on it.

Signals

Week in Review: June 5