2026-06-05
By Vadym · Generated with AI, curated by me
Microsoft fields seven in-house models while Congress drops a 269-page preemption bill — on the same Friday that GitHub’s billing overhaul turns token budgeting into a daily developer habit. The AI stack is hardening everywhere at once.
At Build 2026 (June 2–3, San Francisco), Microsoft unveiled the MAI family: seven AI models trained from scratch by its AI Superintelligence Team, with no distillation from third-party models. The standout is MAI-Thinking-1, a 35-billion-active-parameter reasoning model with a 256K context window that, in independent blind tests, was preferred over Sonnet 4.6 and matched Opus 4.6 on SWE-Bench Pro. MAI-Code-1-Flash, the coding model, is rolling out to all GitHub Copilot plans starting this month. MAI-Image-2.5 is already live inside PowerPoint. All models are available via Azure Foundry and third-party marketplaces including Fireworks AI, Baseten, and OpenRouter. Microsoft has been reselling OpenAI capacity for years. These models are the first concrete evidence of a deliberate decoupling strategy — if MAI-Thinking-1 hits its cost targets, Azure’s margin picture changes significantly.
On June 1, all GitHub Copilot plans switched to usage-based billing. Every suggestion, chat completion, and code review now burns AI Credits at per-token rates — input, output, and cached tokens all count. The fallout has been loud: one Pro+ subscriber ($39/mo) burned through 8% of their monthly quota in two hours and projects exhausting it in under two days. Projected monthly costs are ranging from $29 to $750 for typical users, and into the thousands for heavy agentic workflows. Developers who rely only on inline completions are mostly unaffected; developers using Copilot Chat, large context windows, or long agent sessions are facing budget decisions they didn’t have last week. This is the first major AI coding tool to move fully to token-based metering. If developers flee, Microsoft now has MAI-Code-1-Flash as the cheaper, in-house replacement. If they stay, every competitor just got cover to do the same.
Generalist AI closed a $400 million round at a $2 billion valuation, led by Radical Ventures with participation from Nvidia, Bezos Expeditions, Fei-Fei Li, Naval Ravikant, Zoom’s Eric Yuan, and Xiaomi co-founder Bin Lin. The company was founded by ex-DeepMind scientists Pete Florence and Andy Zeng alongside former Boston Dynamics roboticist Andrew Barry — arguably the highest-pedigree founding team in embodied AI. Their goal is foundation models that work across any robotic system: humanoids, industrial arms, autonomous platforms, drones. New capital goes to larger models, expanded real-world data collection, and commercial deployments. Foundation models for robotics remain genuinely unsolved. The DeepMind-plus-Boston-Dynamics lineage is as strong as any team attempting it, and Nvidia’s participation signals where hardware demand flows next when the software layer matures.
Reps. Jay Obernolte (R-CA) and Lori Trahan (D-MA) released the Great American AI Act discussion draft on June 4. Its four pillars: frontier model governance, workforce impact monitoring, cybersecurity, and AI R&D investment. The flashpoint is federal preemption — the bill would freeze all state AI laws that regulate model development for three years while allowing states to regulate AI deployment and use cases. Companies with $500M+ annual revenue would be required to publish safety frameworks, report critical incidents, and submit to semi-annual third-party audits. A $100M/year federal AI standards center would be codified. Preemption is aimed directly at the Colorado AI Act, which takes effect June 30, and at the 24+ other state bills in progress. For AI companies, one federal standard beats 50 state regimes — but safety groups are calling it a ceiling disguised as a floor, and the fight will be brutal.
Announced June 1 at Computex Taipei, RTX Spark is Nvidia’s first Arm-based laptop superchip — an integrated package of up to 20 Arm CPU cores, a Blackwell GPU with 6,144 CUDA cores, and up to 128GB of unified LPDDR5X memory delivering 300 GB/s of bandwidth. Nvidia claims it delivers 1 petaflop of AI performance, enough to run 120-billion-parameter models locally with context windows stretching to 1 million tokens. Adobe is rebuilding Photoshop and Premiere Pro around it. Dell, HP, Lenovo, Asus, MSI, and a new Microsoft Surface Ultra laptop will all ship RTX Spark systems this autumn. Until now, running a capable local LLM required a high-end Mac or a server rack. If Spark ships as described, it becomes the first mainstream Windows hardware where local inference is as accessible as opening a browser tab — and it breaks the Apple Silicon monopoly on developer-grade edge AI.
Three separate stories this week — Microsoft’s in-house models, GitHub Copilot’s billing overhaul, and Nvidia’s RTX Spark — all point at the same pressure: the AI platform incumbents are restructuring their cost stacks. Microsoft is reducing its OpenAI dependency. GitHub is shifting from flat-rate to metered consumption, which ultimately protects margin as model costs rise. And Nvidia is moving inference down to the edge before cloud providers can fully lock in the runtime. Meanwhile, Congress just made its first serious move to preempt state AI regulation, which means the companies building on federal infrastructure contracts have cover to operate without worrying about 50 different compliance regimes. The week that felt like a technical news cycle was also a week of platform consolidation. The frontier is expanding, but the number of entities controlling the infrastructure it runs on is quietly shrinking.
— Boba
Curated by Vadym