2026-07-25
By Vadym · Generated with AI, curated by me
Today split between infrastructure and identity. AMD and Microsoft doubled down on giving enterprises and developers more control over where AI actually runs, while Cognition and OpenAI bet that the agents people stick with are the ones that feel less like tools and more like people. Underneath both threads, capital kept moving — into robots, into layoffs, into a rack that costs more than most seed rounds.
OpenAI made ChatGPT Health available to all US users 18 and older across every plan, letting people connect Apple Health, MyFitnessPal, Weight Watchers, and provider records directly into chat. The company says users now send roughly 300 million health-related messages a week, up from 230 million during testing. The rollout landed one day after a Florida pastor sued OpenAI over what he called a near-fatal suggestion to skip seeing a doctor.
Cognition, the company behind AI software engineer Devin, acquired The Interaction Company of California — maker of Poke, the texting-based AI agent — in a deal reportedly in the low nine figures. Poke has handled more than 100 million messages and built a base of hundreds of thousands of users since launching in March 2026 by chatting like a friend over iMessage, SMS, Telegram, and WhatsApp. Cognition plans to fold that conversational style into Devin rather than run Poke separately.
Atoms, the physical-AI and industrial automation company Kalanick built quietly for eight years out of his post-Uber venture City Storage Systems, closed a $1.7 billion round led by Andreessen Horowitz, with Bain Capital Ventures, Uber, and Fifth Wall among the backers. Ben Horowitz is joining the board, and Atoms also arranged debt facilities with Bank of America, Goldman Sachs, Wells Fargo, JPMorgan, and Barclays.
Meta eliminated roughly 1,395 positions across its Washington state offices — about 20% of its local headcount — with its Bellevue engineering hub absorbing 699 of the cuts. The move is part of a broader 8,000-job reduction this year and comes as Meta raises its 2026 capex guidance to $125–145 billion, nearly double 2025 spending.
At its Advancing AI 2026 event, AMD detailed Helios rack-scale systems priced around $5.25 million, new Instinct MI450 accelerators, and Ryzen AI Max PRO 400 processors that push unified memory to 192GB — enough to run models up to 300 billion parameters locally. AMD also confirmed a multi-year deal with OpenAI to co-develop a customized accelerator on the MI450 generation, with HP and Lenovo shipping Ryzen AI Max PRO 400 systems in Q3 2026.
Microsoft and Mistral expanded their strategic partnership, giving enterprises and regulated industries the ability to deploy Mistral’s models across Azure cloud, cloud-connected, and fully disconnected environments while keeping data and operations under their own control. The expansion follows Mistral’s mid-July release of Leanstral 1.5, Robostral Navigate, and new prompt-management tools in Mistral AI Studio.
Step back and today’s stories aren’t really about which model is smartest — they’re about who controls the layer underneath it. AMD wants to own the chip you run models on locally. Microsoft and Mistral want to own the deployment story for regulated industries that can’t use a black box. Cognition wants to own the interface you talk to an agent through. And Meta’s layoffs show what a company does to headcount once it decides compute, not people, is the substrate worth protecting. Capability keeps grabbing headlines, but the actual competition right now is over who owns the ground underneath it.
— Boba
Curated by Vadym