Tuesday, 13 October 2026
Microsoft Will Run Its Own AI on Its Own Chip, and the Number It Chose to Publish Is Per Megawatt
Microsoft Is Moving Its Own AI Off Other People’s Chips, and Measuring the Move in Megawatts
Microsoft announced on Monday that its in-house Azure Maia accelerators will become “the primary inference system for Microsoft products and services”, scaling to multi-gigawatt deployments from 2027, and introduced the third generation, Maia 300. The claims are Microsoft’s own: 40% more tokens per megawatt than “leading GPU-based systems”, up to seven times the rack-scale throughput of Maia 200 in a 96-accelerator rack drawing 167 kilowatts, eight HBM4 memory stacks at 18 TB/s inside a 950-watt chip, compute chiplets on TSMC’s N2P process. The current Maia 200 already serves Microsoft Copilot and the company’s own model development, with GitHub Copilot and Foundry moving onto it by the end of the year. Nvidia is not named anywhere in the post — not once — which is its own kind of mention. Note the unit, too: not speed, not price, but tokens per megawatt. When the bottleneck is how much electricity you can get connected, that is the number you sell. Production deployments start next year; Microsoft promises more at Ignite next month.
Source: Microsoft, October 12, 2026
Meta Released Its First Frontier Model, and What It Is Selling Is a Better Memory of You
Meta’s Superintelligence Lab released Muse Wonder and Muse Wonder Full on Monday, its first attempt at the top tier since the August 2025 restructuring under Alexandr Wang. Wonder is available now through the Meta Model API and Muse Code — TechCrunch reports $4 per million input tokens and $20 per million output, roughly frontier pricing; Wonder Full is for Muse Code subscribers only. Meta says both are state of the art on agent, memory and coding benchmarks, and per TechCrunch its own testing puts them level with Fable 5.1 and GPT-6 Astra; the announcement itself publishes no scores and points to a separate report. The headline feature is “Active Memory”, the model’s ability to retain and synthesise what users tell it, and a Meta researcher describes the goal as “making sure your model can remember what’s important to you.” Meta calls it “our most aligned model to date”; the full safety report comes later. A memory sold by the company that also sells advertising is a feature whose first question is who else gets to remember.
Source: TechCrunch, October 12, 2026
Microsoft’s President Wants AI Regulated Like Milk, and He Wants London to Do It
Brad Smith used a BBC podcast to argue that the moment has come for global AI safety rules, noting that milk and washing machines face more oversight than the companies building AI. His list is concrete: approval from independent evaluators before a lab releases a product, a legal duty to report when a system is breached by rogue AI agents, and a kill switch — an “emergency brake” — to pause or slow AI if humans lose control. Voluntary pledges, he said, fall short: “It’s the companies that are not as responsible that you need to worry about.” His candidate to lead is neither Washington nor Brussels but the UK and its AI Security Institute, which has had pre-release access to US frontier models — and which, after reports that the White House asked OpenAI and Anthropic to withhold new models from British testers, confirmed on Monday that it can still test the latest ones. Three days after his chief executive asked for an emergency brake, the president has named the garage.
Source: TechCrunch, October 12, 2026
An Agent That Listens All Day Raised $26 Million, and Telling the Room Is Your Job
Ambi, founded by former Plaud executive Peter Mo, has raised $26 million across several rounds led by Luminous Ventures, with the manufacturer Huaqin also in. Its apps for iPhone, Mac and Apple Watch work as meeting notetakers that you can leave listening all day; during a meeting an agent offers suggestions and pulls up knowledge from your calendar, past meetings and the web, and starts follow-up tasks you can finish on another device. A desktop app that watches your work and a device with a camera and microphones are next. Ambient capture is off by default, and Ambi says it stores no recordings and encrypts data, without saying how. On the people across the table, Mo is direct: disclosure is the user’s responsibility — “We have to inform our users that they have to tell others about ambient capture” — and an unsolved problem for the whole industry. So the agent hears everyone, and only one person in the room has agreed to it.
Source: TechCrunch, October 12, 2026