Figure is building frontier-model-scale compute for humanoids

On September 3, Figure and Nscale signed a multi-year deal for an initial $3.5 billion of compute, with intent to scale beyond $6 billion and potentially deploy up to 100,000 NVIDIA Vera Rubin GPUs from the second half of 2027. Nscale is also taking an equity stake in Figure. Figure explicitly says its Helix models are now constrained by data and compute, not merely robot hardware. Reuters

This is a serious phase change. Humanoid companies are beginning to build training infrastructure on the same scale previously associated with frontier language models. The emerging flywheel is now very explicit: physical data → massive model training → simulation → robot deployment → more physical data.

OpenAI's robotics effort looks substantially more physical than a model-only project

Current OpenAI robotics hiring pages visible on September 5–6 describe development across actuators, sensors, robot data acquisition, manufacturing readiness and multiple deployed robotic workcells. One role specifically covers motors, gears, sensing, electronics, firmware and production readiness; another owns sensor-system development; others describe a large live environment continuously generating real-world robot data for training. OpenAI

That is strong evidence of a full model-to-body stack, rather than simply adapting OpenAI models to third-party robots. The interesting part for us is the architecture: sensing + actuation + physical data infrastructure + model training are being built as one system.

Taiwan is preparing autonomous weapons at tens-of-thousands scale

On September 3, Taiwan's cabinet proposed an additional US$4.6 billion in 2026 defense spending. The package includes more than 40,000 coastal attack drones, over 600 reconnaissance/surveillance drones and more than 100 suicide drone boats, alongside missile systems. The funding still requires legislative approval. Reuters

This is less about a new drone and more about doctrine becoming procurement reality: attritable robotic mass across air and sea. Quantities in the tens of thousands increasingly look like the normal planning unit for autonomous defense.

The U.S. Army is building the training infrastructure for one-human-to-many-robots

On September 3, the Army disclosed deployment of its Virtual Drone Collective Trainer, allowing entire formations to rehearse operations involving friendly drones and hostile swarms. A recent exercise networked 145 laptops, real military controllers can plug directly into the simulation, and the next phase explicitly targets training for one-to-many control of autonomous drone swarms. U.S. Army

This is a useful execution-layer signal. Militaries are no longer training only individual drone pilots. They are starting to train humans to supervise robot populations.

Tesla's Cybercab has pushed embodied autonomy into a regulatory collision

Reuters reported on September 4 that Tesla has begun paid Cybercab operations in Texas using purpose-built autonomous vehicles with no steering wheel, pedals or mirrors, while U.S. regulators are auditing roughly 1,000 vehicles and examining Tesla's self-certification approach. Reuters

Whatever the eventual regulatory outcome, this is an important embodiment milestone: the machine is no longer an autonomous system layered onto a human vehicle. The physical form itself assumes the agent will drive.

China's next five-year SME strategy explicitly bundles robotics, BCI, materials and embodied AI

A September 3 plan issued by ten Chinese central agencies targets emerging SMEs in robotics, embodied AI, brain-computer interfaces, new materials and quantum technology, with expanded lending, capital-market access and a second phase of the national SME development fund. Beijing wants the number of specialized "little giant" firms to reach 22,000 by 2030. Reuters

For the investment map, this is important. China is treating these fields not as isolated research categories but as a single strategic industrial ecosystem.

Signal read

🔥 Physical-AI compute: humanoid training is moving toward frontier-AI infrastructure scale.
🟢 Model → body integration: OpenAI appears to be building sensors, actuators and data operations in-house.
🟢 Autonomous-defense mass: procurement quantities are climbing into tens of thousands.
🟢 Human → swarm supervision: becoming an explicit military training objective.
🟢 Autonomous agents in purpose-built bodies: Cybercab is a meaningful deployment threshold.
🟢 State capital for embodiment: China is bundling robotics, BCI and advanced materials into long-term industrial policy.
🟡 Fresh technical BCI / bionics / biohybrid breakthroughs: I found no September 4–6 result strong enough to manufacture a headline around.

Core takeaway: this cycle is unusually architectural. The pieces we have been watching separately are snapping together: gigantic compute for robot brains, dedicated sensor and actuator programs, mass-produced autonomous weapons, humans supervising robot fleets, and governments financing the entire stack.

The interesting bottleneck is moving again. It is no longer merely Can we build the body? Increasingly it is Can we train, coordinate, secure and manufacture millions of reliable physical agents?

That is a much bigger creature. 🦾⚡