AI News October 7, 2026: Microsoft Bets the PC on Local AI Today, Mistral Drops a 1-Trillion-Parameter Open-Weight Giant, Google Buys 3.6GW of Power

Two currents are pulling AI in opposite directions this week: models are getting radically more open, and the infrastructure behind them is getting radically more locked-down. Today Microsoft bets the entire PC platform on AI that runs on your desk, Mistral and Reflection ship the largest open-weight models the West has ever seen, Google buys gigawatts of nuclear-adjacent power to feed its data centers, and Meta’s agent security story gets an uncomfortable behind-the-scenes look. Here are the five stories that matter.

1. Microsoft’s Big Bet: Windows and Surface Go All-In on Local AI Today

At 10 a.m. Pacific in San Francisco, Microsoft hosts its Windows and Surface event with a pointed tagline — “something new is coming from Windows” — and a stated focus on how local AI will shape the next chapter of the PC. CEO Satya Nadella, Windows and Devices chief Pavan Davuluri, and NVIDIA CEO Jensen Huang are all expected on stage, with the NVIDIA RTX Spark platform front and center.

Windows Central reports the event should surface the Surface Laptop Ultra and the Surface RTX Spark Dev Box — devices built to run serious AI models locally — alongside Project Zenith, Microsoft’s codename for a new class of developer-first local AI machines. After two years of cloud-first Copilot, this is Microsoft telling developers the next platform shift runs offline, on-device, and on their silicon.

2. Mistral Large 4 “Le Chonk” — A 1-Trillion-Parameter Open-Weight Flagship From Europe

Mistral launched the public preview of Mistral Large 4 on October 6, and the numbers are arresting: a 1-trillion-parameter natively multimodal mixture-of-experts model with 49 billion active parameters, trained from scratch over two months on roughly 3,800 NVIDIA Grace Blackwell GPUs in Mistral’s own European data centers — a sovereignty flex aimed squarely at regulators and enterprises.

Performance-wise, Mistral claims it beats every open-weight model from the US or Europe, is state-of-the-art among open models on cybersecurity, finance, and legal workloads, and even surpasses frontier closed models on visual grounding. It scores 38 on the Artificial Analysis Intelligence Index — the top Western open-weight result, though still behind leading Chinese open models. API pricing is aggressive: $1.36/M input, $4.18/M output. Weights drop at the end of October, after red-teaming with cybersecurity partners and state authorities.

3. Reflection AI’s Beam: Frontier Coding Skill at a Fraction of the Compute

Reflection AI unveiled Beam, its first open-weight model: a 501B-parameter sparse MoE with 23B active, pretrained on 23.8 trillion tokens and sharpened by a reinforcement-learning run that generated over 100 million rollouts on 10,500 NVIDIA GB300 GPUs in four weeks.

The benchmark sheet is the story: 77.2 on SWE-Bench Pro v2-Hard, 80.1 on Terminal-Bench v2.1, 97.8 on AIME 2026, and 90.5 on GPQA Diamond — matching Z.ai’s GLM-5.2-class performance while using 3–4x less inference compute, and approaching far larger systems like Qwen 3.8-Max and Kimi K3 on efficiency rather than raw scale. Weights, the technical report, and fine-tuning artifacts ship later this month under Apache 2.0 — arguably the most permissive license ever attached to this capability class.

4. Google Signs 3.6-Gigawatt Power Deal, a Quarter of It New Nuclear

Google contracted 3,590 MW with Constellation Energy, with roughly 25% coming from new nuclear capacity across 11 reactor upgrades in Illinois, Pennsylvania, and New Jersey, per Reuters. The structure: a 20-year PPA for 890 MW of nuclear plus a separate long-term contract for 2,700 MW from other sources, triggering more than $4.3 billion in Constellation investment, with first deliveries in 2028 under PJM’s “bring your own power” framework for AI data centers. The AI land grab is now, literally, a power grab.

5. Meta Patched Muse VM Escapes Days Before Launch

404 Media reports Meta engineers worked nights and weekends from August 27 to the September 8 launch patching KVM escape vulnerabilities in Muse — flaws that could have let a Muse user break out of the agent’s per-user virtual machine and reach internal Meta databases. VM escapes sit in Meta’s highest bug-bounty tier at up to $300,000. Researchers quoted in the piece say the fixes narrow but may not fully close the exposure, calling production “one KVM escape away.” As agents get their own computers, agent isolation becomes the new perimeter.

Worth tracking too: Anthropic expanded its Claude Startups program with a $45K partner stack; a16z’s latest consumer AI report says ChatGPT has ~3x the US paid subscribers of Claude or Gemini (but 7.5% of Claude subscribers pay $100+/month); and Mizuho is deploying a Claude-based “Cowork” agent to 30,000 staff.

The through-line this Wednesday: openness is scaling up faster than anyone expected — trillion-parameter weights and Apache 2.0 frontier coders in the same week — while the energy, silicon, and security foundations underneath get heavier, hotter, and harder-edged. The open frontier is no longer the sideshow; it’s the main stage.