Physical Intelligence vs hardware giants Figure and Tesla

Physical Intelligence operates out of San Francisco and secured $1 billion in total funding following a $600 million Series B led by Alphabet Inc. This follows a $400 million Series A led by Jeff Bezos and OpenAI. This Vision-Language-Action architecture uses flow matching, a variant of diffusion models, to output motor commands at 50Hz. While LLMs predict text, pi0 predicts physical trajectories by using semantic knowledge from a 3 billion parameter vision-language backbone based on Google’s PaliGemma. The flow matching method utilizes deterministic Ordinary Differential Equation solvers to achieve faster inference compared to traditional diffusion techniques. This allows the pi0 model to output continuous, high-frequency motor commands that enable human-like dexterity when the robot peels a grape or assembles a box. The company uses a "Recap" method to train models through reinforcement learning with experience and corrections to reach 99.9% reliability. Physical Intelligence uses cross-embodiment learning by training on data from eight distinct robot types. This approach allows the model to handle tasks like laundry folding, bussing tables, and assembling cardboard boxes. Physical Intelligence bypasses the data bottleneck by using this diverse dataset. The company develops software that functions for any robot with a motor. This software focus competes with the vertical integration of Tesla and Figure.

Figure 02 vs Tesla Optimus

Figure 02 maintains a $40 billion valuation because it earns revenue from external customers like BMW. Figure’s Helix VLA model operates at 200Hz, which allows the robot to coordinate tasks across both hands without scripted choreography. Tesla targets 5,000 to 10,000 units in 2026 with a cost target between $20,000 and $30,000, a price that places it significantly below the $50,000 cost of Figure 02 for enterprise buyers. Tesla’s production in Fremont and Texas includes 500 Gen 2 units for internal use. Figure operates 150 units in commercial settings, including a deployment at BMW Spartanburg. Figure’s $50,000 price tag results in a 12 month payback period, whereas Tesla’s $25,000 cost targets a payback period of under 6 months. Tesla prices its robots below the $52,000 to $68,000 annual loaded warehouse labor cost in the US. Tesla uses its $10 billion FSD compute and Dojo infrastructure to train Optimus policy models using 1.5 billion FSD miles per quarter. Figure relies on a 2024 funding round that included Microsoft, NVIDIA, OpenAI, Jeff Bezos, and Intel Capital. Figure 02 maintains a $40 billion valuation because it earns revenue from external customers like BMW, while Tesla focuses on internal production within its own factories to iterate on the robot. You should ignore public demos and watch the cumulative shipment numbers instead.

Warehouse manipulation and market scale

Rebl Industries achieves 99.7% picking accuracy in Nordic warehouses using Zebra and Photoneo perception systems. These robots complete 800 to 1,000 picks per hour. Apptronik ships 50 Apollo units to Mercedes-Benz and GXO Logistics following a $350 million Series A. 1X Technologies has Neo Gamma deployed in Norwegian retail pilots. Figure has expanded to a second undisclosed Fortune 500 customer and a third pilot in logistics. China shipped 10,000 humanoid units across all manufacturers in 2025, while US companies shipped an estimated 600 units. The humanoid market reaches $38 billion according to Goldman Sachs estimates. Total real revenue across all 50+ humanoid companies in 2026 remains under $500 million. A German survey shows 82 percent of economic stakeholders consider humanoid robots drivers of future innovation. The industry expects a massive push to collect embodied data from the real world to bridge the sim-to-real gap. Unitree G1 ships at $16,000, while UBTech Walker S2 costs roughly $30,000. Figure wins on revenue per unit, but Tesla wins on total volume. Which company will control the hardware ecosystem?