The Physical AI Inflection: Mini Thesis (Eximius Ventures)

Summary

• Robotics and Physical AI startups raised a record $27.6B across 1,009 deals in 2025, more than double 2024.p.6
• Humanoid bill-of-materials (BOM) costs have roughly halved since 2024, with the Unitree G1 now listing at about $13.5K to $16K.p.6
• Teleoperation hardware costs have fallen from roughly $20K to $300 in two years, while usable robot data costs declined from about $340 to $118 per hour.p.6
• In defence robotics, deal value grew over 260%+ with a median Series A valuation of $105M compared to $50M for non-defence.p.10
• Operating costs for Figure robots at BMW are around $25/hr today, with projected robot operating costs dropping below $5/hr by 2030.p.12
• The market for robot fleet-management software is estimated at $2.93B in 2026 and projected to reach $12.67B by 2035.p.15

Key insights

1. Physical AI relies on Vision-Language-Action (VLA) models that unify perception, language understanding, and motor control to achieve generalisable robot behaviour.p.5
2. NVIDIA Cosmos reached roughly 2M downloads, demonstrating the growing industry focus on physics-aware simulation environments to close the sim-to-real gap.p.6
3. Current robot data represents only 0.1 to 1% of LLM pretraining scale, making data access and synchronised physical collection a critical bottleneck.p.7
4. NVIDIA demonstrated synthetic data scaling by generating 780K synthetic trajectories (equivalent to 6,500 hours of human demonstration data) in just 11 hours.p.8
5. Proprietary physical-world fine-tuning data acts as the main competitive moat as foundational models transition from open pretraining to specific deployments.p.9
6. Defence is the earliest viable market for Physical AI adoption due to higher strategic value, controlled deployment environments, and military share reaching 68.1% in autonomous robots.p.10
7. Industrial and manufacturing settings form the second wave of adoption, supported by 542K industrial robots installed globally in 2024 and placement accuracy exceeding 99%+.p.11
8. Service sectors like cleaning and hospitality are expanding adoption, with 42K+ hospitality robots sold in 2024 (up +34% YoY) and a $1.2B projected hospitality service robot market in 2026.p.12
9. A structural geographic divergence exists, where the US leads in frontier foundation models and capital, while China controls hardware, refining, and mass-manufacturing supply chains.p.14
10. India's Physical AI startups raised approximately $155M across 31 deals in Jan-Jul 2026, positioning the country around cost-efficient data and MSME software integration.p.16