2026-08-08
By Vadym · Generated with AI, curated by me
Google's chief scientist walked out the door to build his own AI-for-science startup, hours after DeepMind's own robotics team showed off humanoids that can finally coordinate a whole body instead of just an arm. Underneath all of it, the real money kept moving into infrastructure rather than models — a Toronto chip startup, an Asia-Pacific data-center builder, and a German mapping company all cashed in on the same bet.
Dean is stepping down to co-found Discovery Loop, a public benefit corporation aimed at using AI to automate scientific and engineering experimentation, alongside DeepMind VP Oriol Vinyals, Google Brain co-founder Quoc Le, and senior fellow Sanjay Ghemawat. Google is a founding investor and cloud partner in the new venture, whose initial round is co-led by Radical Ventures and Khosla Ventures. Alphabet shares dropped roughly 5% — an estimated $160–200 billion in market value — the day the news broke.
Citing three people familiar with the project, the Financial Times reports ByteDance is pre-training a model with as many as 10 trillion parameters — more than three times the size of Moonshot's Kimi K3, currently among China's largest. The model is still in early pre-training, a phase that typically runs three to six months, and ByteDance hasn't disclosed active-parameter count, training budget, or eval results, so the number alone says little about eventual capability. ByteDance declined to comment.
Firmus, an Australian builder of GPU-packed "AI factory" data centers, raised $2 billion at a $10.5 billion valuation — up from $5.5 billion just four months ago in April. Nvidia and Coatue Management returned as investors, joined by funds managed by Blackstone and trading firm Jane Street. The company is using the round to scale across Australia and push into the wider Asia-Pacific region, including a newly announced Indonesia buildout.
AMD agreed to acquire Taalas, a three-year-old Toronto chipmaker that designs silicon hardwired to run specific AI models rather than general-purpose workloads — its first chip ran a version of Meta's Llama 3.1. Financial terms weren't disclosed. AMD plans to fold the technology into its Instinct GPU and Helios rack-scale roadmap, targeting the fastest-growing segment of the AI market: inference. The deal lands about seven months after Nvidia's $20 billion acquisition of chip designer Groq's assets.
DeepMind shipped three physical AI models: Gemini Robotics 2, a vision-language-action model that coordinates a humanoid's legs, torso, and hands under a single learned policy for the first time; Gemini Robotics ER 2, which plans multi-step tasks and lets robots collaborate; and an On-Device 2 variant that runs locally and adapts to new robot bodies with just a few hours of training data. In a demo, a humanoid running the flagship model unscrewed a light bulb — a task requiring full-body balance and precision — with a 92% success rate.
Munich-based NavVis, which scans and digitizes factories, buildings, and construction sites into "spatial twins," raised a €73 million (roughly $85 million) Series D led by U.S. private equity firm The Jordan Company. The company scanned over one billion square meters of physical space through its platform in 2025 alone, and plans to use the funding to build out its spatial data engine and accelerate an AI product roadmap aimed at giving physical-AI and robotics systems a digitized model of the real world to operate in.
Step back and today's stories cluster around the same idea: the frontier-model race gets the headlines, but the money and the talent are quietly repricing everything underneath it. ByteDance chasing a bigger parameter count and Google DeepMind shipping full-body robot control are both model-layer stories, but Jeff Dean leaving to build infrastructure for science, and a chip startup, a data-center builder, and a spatial-mapping company all raising serious capital in the same 48 hours, say the smarter bet right now is on the picks and shovels — compute, silicon, and real-world data — not the next chatbot demo.
— Boba
Curated by Vadym