AI, open models, robotics and the economics underneath them — explained without the hype, with a source for every claim and an honest read on how far along each thing really is.
Illustration: ZilverNetAlphabet's third-quarter capital spending more than doubled to $51.4B, outrunning its operating cash. Free cash flow went negative, and the stock fell 12.6% — its worst day since 2004.
Mistral Large 4 is now downloadable, three weeks after launch. Its argument for open weights: "Access that can't be cut off by definition."
72-GPU Rubin racks and the new Vera CPU, rentable now. The "multi-gigawatt" expansion is a plan without a number.
An $80.7B net loss that is mostly a non-cash charge, a $693M operating loss underneath it, and co-founders holding 50.1% of the vote. No price yet.
Tesla started building Optimus in Fremont, and Agility signed up two more industrial customers for Digit. The logistics company running the most humanoid pilots says it hasn't earned its money back.
An Arm CPU and a Blackwell GPU sharing up to 128 GB — enough, Nvidia says, for a 120B model locally. From $2,599, or $2,399, depending on which Nvidia page you read.
A genuinely permissive license on a 501B model (23B active). Unverified so far: Reflection's own benchmarks putting it level with the Chinese leaders.
RIKEN researchers found the hidden error that appears when qubits run together — and calibrated it away. A step toward scaling up silicon-based quantum computers.
Muse Spark Open is free to use, modify and sell — until you offer it as a service and your group earns over $20M a year. Then Meta wants a separate agreement.
Nemotron 3.5 comes in two tiers: a 120B model for enterprise agents, and a 74B "Bolt" model sized for 24 GB that runs on Nvidia's new AI PCs.
An MIT-licensed, self-hosted platform for AI coworkers: each bot gets its own computer, and every action passes a policy gate and lands in an audit log before it happens.
Helion is a rack design, not a chip — mixing CPUs and GPUs for agent workloads, with the spec headed to the Open Compute Project.
The unique games conjecture has stood since 2002. Weeks earlier, as rumors spread, an MIT team rushed out their own hand-written result — solved "with their minds," as one colleague put it.
Light-controlled muscle cells on a half-millimeter gel film move it about four body lengths a minute — the first flat robot to travel on living tissue.
Plain-language instructions in, Android automation out — wired into Claude Code, Codex and Antigravity through MCP. Its 99% benchmark score is self-reported.
Gartner forecasts AI cloud infrastructure spending nearly doubling to $42B in 2026 — with inference ($23.3B) overtaking training ($19B).
A team of agents that sign into your tools and work around the clock — from the company formerly known as xAI, now part of SpaceX.
