Physical AI

Mega-rounds still concentrate capital in a few apex physical AI labs

Capital is extremely lumpy: the week of 2026-09-28 logged $6.4B across just 4 deals, and seed-stage deals account for $6.3B of 90-day volume. Project Prometheus, founded by Jeff Bezos, is reportedly close to a $10B raise at a $38B valuation, and Nvidia's Jensen Huang publicly praised it as a compute buyer. A $500M growth round on 2026-10-15 shows late-stage money following the same pattern. Smaller recent rounds, like a $12.5M a16z Speedrun/NFX seed and an $18.7M Series A, are tiny by comparison, so the market is barbelled between frontier-lab mega-rounds and small seed bets.

Physical Intelligence's π0.5 is the default baseline across papers, serving as the System 1 controller and being evaluated through world-model proxies with 0.984 ranking correlation. Dyna, Generalist, and AgiBot are framed as the research-oriented peers. Dyna describes a 'product research' approach in which commercial loops probe the model's intelligence ceiling.

Why it matters · If a few VLA stacks become the standard, hardware makers and application builders will depend on them, much as they now depend on LLM providers.

World models can grade competitors' policies without physical access, and General Intuition trains action-labeled world models on gaming video. Separate research shows a general-purpose model, GPT-6-Astra, beating specialized VLAs via in-context learning. Interpretability work is also mapping how VLAs route information from vision to action.

Why it matters · Cheaper simulation-based evaluation and generalist models could erode the moat of specialized VLA vendors and cut data costs.

Unitree G1, AgiBot G2/World, Fourier GR-1 and ActionNet, and Galaxea R1 are repeatedly used as benchmark hardware and training data. Dyna's CEO called Unitree's sim-to-real gap a strong breakthrough. Meanwhile, research argues Figure, Boston Dynamics, and Apptronik over-focus on end-effector dexterity rather than whole-body contact.

Why it matters · Standardized, low-cost hardware and open datasets lower barriers for model developers and shift value toward software and data.

Ulysses runs a full Nvidia 5090 onboard its Mako underwater vehicle, roughly 100x the compute of other AUVs, and was backed at pre-seed by Lowercarbon Capital. Sacca frames its vertical integration as a SpaceX/Tesla-style approach, and the Saronic comparison shows defense autonomy competing on integration.

Why it matters · Cheap, compute-rich autonomous platforms threaten incumbent defense and infrastructure hardware vendors.