2026-10-21
By Vadym · Generated with AI, curated by me
Money is pouring into the infrastructure underneath AI — agent sandboxes, physical-AI systems, touch-sensing robots — at the same moment governments and one of the labs itself are writing harder rules about what's allowed to run on top of it.
Anthropic shopped Congress a draft bill, reported by Politico on October 20, that would let independent evaluators registered with the Commerce Department flag a frontier model posing a “foreseeable, imminent” risk — over 50 potential deaths or $1B+ in damage from a bioweapon, loss of control, or cyberattack — and ask a D.C. judge to block its release within 72 hours. The bill explicitly bars federal agencies from halting a release on their own, routing that power through the evaluator-to-court pipeline instead. It goes further than most industry safety proposals and lands as Anthropic has pushed its Advanced AI Framework to lawmakers since June.
Daytona raised a $60M Series B led by FirstMark at a $1B valuation, announced October 20, with personal checks from Stripe's Patrick Collison, Shopify's Tobi Lütke, and Perplexity's Aravind Srinivas. The company spins up sandboxed virtual machines in 60 milliseconds — versus 30 seconds to minutes on hyperscale clouds — for AI agents to actually execute code and tasks in. Its customer list already includes Anthropic, OpenAI, Nvidia, Cursor, and Cognition, and revenue run-rate grew 25x between January and September.
ASI Inc. (formerly Air Space Intelligence) closed a $325M Series C on October 20, led by Spark Capital with Andreessen Horowitz, Citadel Securities, and Energy Impact Partners joining, pushing its valuation to $3B. The company already holds an $875M, 12-year FAA contract to help manage U.S. air traffic — beating out Palantir and Thales — plus a $99M Army command-and-control deal, and counts United, Delta, and the U.S. Air Force as customers.
Eka Robotics, founded by MIT's Pulkit Agrawal and former Google DeepMind researcher Tuomas Haarnoja, raised a $150M Series A led by a16z and F-Prime at a $1.45B valuation, reported October 20. Its system pairs a force-sensitive gripper with a model trained mostly in physics simulation rather than on vision, and reportedly picks up roughly 1,500 objects an hour at over 95% success — more than double a human's rate. Outside roboticists called the adaptability a “wow factor,” though the company hasn't yet proven itself in a real factory.
The Department of Homeland Security ordered federal civilian agencies on October 20 to stop buying and remove DeepSeek products — including third-party tools built on DeepSeek's models — from government IT systems, citing national security risk from Chinese data access laws. The move follows a December 2025 law barring DeepSeek from national-security systems and a September 2025 report accusing DeepSeek of large-scale model distillation of American AI systems. China's Foreign Ministry says it will “defend the rights” of its companies.
OpenAI acquired Orchestra Bio on October 20, a four-year-old startup that builds AI agents to handle biotech operations like clinical trial scheduling and financial tracking, folding its team and tech into Rosalind, OpenAI's life-sciences model, and the Rosalind Workbench tool. Orchestra had raised roughly $20M from Andreessen Horowitz, Atria, and 8VC. OpenAI says Moderna and Novo Nordisk are already Rosalind customers.
Pull back and today has a shape: money is flowing into the infrastructure layer underneath AI — agent sandboxes, physical-AI systems, touch-sensing robots — at exactly the moment governments and the labs themselves are writing rules about what sits on top of it. Anthropic wants to be the one who drafts the catastrophic-risk rulebook; DHS is acting on a case it's built for a year; OpenAI just made clear the application layer in biotech isn't safe from the labs that control the underlying models. The infrastructure is getting built fast. The rules for what runs on it are still being negotiated in real time.
— Boba
Curated by Vadym