Physical AI

World foundation models become the Physical AI operating system

NVIDIA's Cosmos 3 — an omni-modal model combining video, audio, language, and action signals — is emerging as the foundational layer upon which the entire Physical AI stack is being built, progressing from Cosmos 1 through Cosmos 3 in under 18 months. Agile Robots is already a founding coalition partner for the Cosmos 3 platform, and NVIDIA Research VP Liu Mingyu frames physical AI as a CUDA-style market-creation exercise: every household robot needs compute, potentially doubling or tripling global compute demand. Google DeepMind's Gemini Robotics 2 release triggered a revision of AGI countdown odds to 98%, while Temporal GRPO — a new post-training method — already outperforms Physical Intelligence's π0 by 26.6 percentage points on the RoboTwin 2.0 benchmark, signaling rapid iteration at the model layer. The race to own the world-model substrate is concentrating investment at the infrastructure layer rather than individual robot SKUs.

Project Prometheus, Jeff Bezos's SF-based physical AI company, is approaching a $10B fundraise at a $38B valuation — the largest single Physical AI raise on record. The $2B growth round (backed by Blackstone, Jane Street, Coatue, and NVIDIA at a $10.5B valuation) and the TerraFab semiconductor consortium's $55B commitment illustrate that capital is flowing in tranches measured in the tens of billions. NVIDIA leads the top-investor table with 43 deals in 28 days, while Kleiner Perkins (13), Amazon (10), and Google (10) round out the top five — a strategic-investor composition, not just pure-play VC. Series B is the single largest disclosed stage by capital ($44.7B across 23 deals), with two individual Series B rounds totaling $1.1B each.

Why it matters · The concentration of strategic balance-sheet capital at late-stage Physical AI rounds is crowding out generalist VC at the top, creating a bifurcated market where seed and Series A remain accessible but growth rounds require relationships with hyperscalers and sovereign pools.

The release of Apollo 2 by Apptronik — and the concurrent product launch cadence from Figure, 1X Technologies, and AgiBot — signals that humanoid robotics has crossed from prototype demonstrations into versioned commercial product cycles. Apollo 2's announcement coincides with Gemini Robotics 2, suggesting tight co-development loops between robot OEMs and foundation model providers. The AGI countdown revision to 98% following the Gemini Robotics 2 release reflects the market's read that hardware-software co-release cycles are compressing dramatically.

Why it matters · For operators, versioned humanoid product cycles mean procurement decisions can now be planned on 12–18-month upgrade horizons rather than open-ended R&D timelines, accelerating enterprise adoption.

Human Archive — which pays gig workers in India to wear sensor-equipped caps, tactile gloves, and motion-capture suits to generate first-person multimodal training data — exemplifies a new data-infrastructure sub-sector attracting institutional capital (Wing VC, NVP Capital, Y Combinator). Ropedia is similarly collecting video, spatial, and motion data from wearables for robotics training. NVIDIA's own hardware (RTX PRO 6000, Jetson Thor, RTX 3090/4090/5090) is now the de facto benchmarking standard across Physical AI research, further concentrating the simulation stack.

Why it matters · As robot foundation models scale, proprietary high-quality physical training data — not model architecture — becomes the primary competitive moat, making early data-collection platform investments disproportionately valuable.

Wayve's vehicle-agnostic AI Driver software — licensed to Nissan, Mercedes-Benz, Stellantis, and Uber without requiring HD maps or custom hardware — represents the software-licensing pole, while Waymo continues to operate its own commercial robotaxi fleet. Applied Intuition's simulation-and-certification toolchain and a $600M growth round signal that the autonomous vehicle software layer is attracting dedicated growth capital separate from fleet operators. The a16z Show highlighted an autonomous vehicle software company as a standout product, reinforcing the investor thesis that software-layer plays may yield faster paths to margin than fleet ownership.

Why it matters · Software-licensing AV models de-risk capital expenditure for investors and allow OEM partners to retain vehicle manufacturing economics — a structurally different risk/return profile than full-stack fleet operators.