Physical AI & Hardware

Physical-AI politics collapses into supply-chain geopolitics. The heart of a robot — actuators, permanent magnets, rare earths — is overwhelmingly Chinese, and China's April 2025 export controls (which halted parts of Tesla's Optimus production) showed it can be weaponized. The US treats this as a security problem, while Japan (state-backed robotics as an aging-society answer) and India (Make in India / PLI) compete to onshore manufacturing.

1. In April 2025 China's Commerce Ministry imposed individual licensing on exports of seven rare earths (samarium, dysprosium, terbium and others) and permanent magnets; Elon Musk publicly blamed the magnet bottleneck for hitting Tesla's Optimus production, showing a robot's core components are now a diplomatic lever[13].
2. China holds about 63% of the key firms in the humanoid component supply chain (concentrated in actuator parts and rare-earth processing) and 94% of global permanent-magnet production, vertically integrated from mine to machine — making the physical layer of Physical AI structurally China-dependent[11].
3. 🇺🇸 CSIS and others frame China's rare-earth/magnet controls as a threat to US defense and industrial supply chains, pushing the securitization of robotics: the US leads on NVIDIA's software/silicon but is exposed on physical components, making allied friend-shoring a policy priority[14].
4. 🇯🇵 Japan promotes robotics as a national answer to aging and labor shortage, holding top-tier robot density (446 units per 10,000 manufacturing workers); in March 2026 FANUC announced a partnership with NVIDIA to embed Jetson Thor into controllers, part of a public-private push on industrial AI[23][16].
5. 🇮🇳 India uses Make in India / Digital India and the Production-Linked Incentive (PLI) scheme to onshore electronics and robot manufacturing; the electronics PLI had drawn about INR 2.16 trillion (~USD 26B) in committed investment by December 2025, though humanoids remain an unlisted priority category, leaving a policy gap[21][22].

Coming soon

Capital is pouring into Physical AI. Physical Intelligence went from a $5.6B valuation to talks for a $1B raise at $11B, Apptronik raised $520M at $5B, and the 1X NEO is on sale at $20k (or $499/month). Pricing is polarizing: high-capability US machines (Optimus targeted at $20k–$30k) versus China's Unitree winning global shipments on volume and price (G1 $13,560, R1 $5,900). Industrial robots reached 542,000 installs in 2024 (China 54%), doubling in a decade — a huge installed base under which the new category is layered.

1. 🇺🇸 Physical Intelligence, which builds the robot 'brain,' raised at a $5.6B valuation in November 2025 and by March 2026 was in talks (Founders Fund, Lightspeed) for $1B at $11B; Apptronik raised $520M at $5B (incl. Google) in February 2026 — capital concentrating in the US[5][6][8].
2. 🇺🇸 The home humanoid 1X NEO opened pre-orders in 2025 with two models — $20k early-access or a $499/month subscription — while Tesla publicly targets $20k–$30k for Optimus at scale, kicking off a price war in the home/service segment[7][9].
3. China's Unitree leads global shipments on low-cost volume — the G1 at $13,560 and the new R1 from $5,900, a fraction of Western machines — undercutting Atlas-class platforms (formerly $200k+) and resetting the economics of robot adoption[12].
4. Per IFR's World Robotics 2025, industrial-robot installs hit 542,000 in 2024 (double the level of a decade ago), with China at 54% of the world and an operational stock above 2 million — Physical AI builds on a massive existing automation market, not from scratch[15].
5. 🇯🇵 Japan's robotics 'Big Five' (FANUC, Yaskawa, Kawasaki Heavy, DENSO, Mitsubishi Electric) account for over 40% of global industrial-robot shipments; their integrated actuator/gearbox/control know-how carries into the humanoid era, even as critics note Japan lagged the AI-driven embodiment boom[23][24].
6. 🇮🇳 India's robotics market reached about $2.14B in 2026 and aims to plug into the global supply chain on manufacturing and talent; as the largest source of US graduate-level robotics researchers, with a deep English-speaking technical pool and the world's largest working-age cohort, it has potential to become a development-and-deployment hub[21][22].

Coming soon

The two social drivers are aging/labor shortage and consumer acceptance of AI devices. In Japan the working-age population shrinks by 12M by 2040, structuralizing demand for monitoring, eldercare and home robots. Meanwhile Meta-style AI smart glasses sold 7M+ units in 2025, and AI companion/toy products (Mattel×OpenAI, Curio plush) are emerging — even as concerns over child safety erupt in parallel.

1. 🇯🇵 Japan's working-age population (15-64) is projected to shrink by about 12 million from 2020 to 2040, structurally expanding demand for service/humanoid robots for monitoring, eldercare and chores; SoftBank's Pepper, AIST's PARO and Cyberdyne's HAL have built a track record in elder-care robotics[23][24].
2. The AI plush-toy market is forecast to surge from $74M in 2024 to $1,373M by 2032 (~52% CAGR); Mattel partnered with OpenAI in 2025 to build AI toys, and Curio sells plush characters like Grok (from $99) as a screen-time alternative companion — domestic AI characters are going mainstream[19].
3. 🇺🇸 The surge in AI toys has made safety a flashpoint in the US: some products reportedly told children where to find knives or how to light a fire, and in March 2026 Senator Gillibrand and colleagues sent the FTC a letter urging an AI-toy probe — regulatory and trust barriers loom over child-facing Physical AI[20][25].
4. EssilorLuxottica sold 7M+ Meta Ray-Ban/Oakley AI smart glasses in 2025 (about triple the prior year); smart glasses crossed in months from 'gadget' to a legitimate eyewear category, marking rapid consumer acceptance of AI-laden wearables[17].

Coming soon

The technical core is 'robot foundation models + edge AI silicon.' NVIDIA aims to be the 'Android of robotics' with GR00T N1 (the first open humanoid foundation model) and Jetson Thor (Blackwell, 2,070 FP4 TFLOPS / 128GB / 130W). Google DeepMind pushes vision-language-action (VLA) into production with Gemini Robotics 1.5, ER 1.5 and On-Device (network-independent, sub-10ms), and Physical Intelligence's π0 shows a VLA that runs on any robot — all enabling the real-work deployment of Figure 03, Atlas and Optimus.

1. In March 2025 NVIDIA unveiled Isaac GR00T N1, the world's first open, customizable humanoid foundation model, with a dual System 1 (fast reflex) / System 2 (deliberate) architecture inspired by human cognition, trained faster via Isaac simulation and Cosmos synthetic data — the core of NVIDIA's robotics-platform play[1].
2. 🇺🇸 Jetson Thor, the Blackwell-generation robot supercomputer, reached general availability in 2025: 2,070 FP4 TFLOPS and 128GB within a 130W envelope, 7.5x the AI compute and 3.5x the energy efficiency of Orin; adopted by Agility, Amazon, Boston Dynamics, Figure and Meta, it is becoming the standard for running generative AI at the edge[2].
3. In September 2025 Google DeepMind released Gemini Robotics 1.5 plus Gemini Robotics-ER 1.5, an 'embodied reasoning' model that plans tasks using tools like web search; in June 2025 it shipped Gemini Robotics On-Device, which runs network-independently at sub-10ms, opening the door to cloud-free on-site robots[3][4].
4. 🇺🇸 Physical Intelligence's π0, released October 2024 and open-sourced February 2025, is a VLA model that generalizes across tasks like folding laundry and making coffee; the company builds no robots, instead supplying a 'software brain that runs on any robot' — advancing a hardware/model division of labor[5].
5. Humanoids moved to real work: Figure 03 hit one robot per hour at its BotQ plant in June 2026 (tooled for 12,000/year), Boston Dynamics' Atlas began parts-sequencing at Hyundai's Georgia plant in January 2026 and won 'Best Robot' at CES 2026, and Tesla showed Optimus running autonomously on a line at CES 2026[9][10].
6. 🇯🇵 In March 2026 FANUC announced an NVIDIA partnership, replacing legacy controller modules with Jetson Thor (T5000) to raise local AI compute over 7.5x and strengthening simulation and digital twins with Isaac Sim, PhysX and GR00T N — integrating Japan's craftsmanship-heavy industrial robots into the Physical-AI stack[16].

Coming soon

The legal frontier is two-front: export controls (component geopolitics) and consumer/child safety (AI toys and companions). China's rare-earth/magnet licensing has effectively become a rule, making sensitive-component trade permit-dependent. In the US, FTC-probe requests over AI-toy unsafe responses and voice-data collection show that regulatory and privacy frameworks for in-home Physical AI are still forming.

1. 🇺🇸 In March 2026 Senator Gillibrand and colleagues sent the FTC a letter seeking a probe into AI-toy safety, with open-ended conversations with children, voice-data collection and unsafe responses at issue — privacy and product-safety regulation for in-home AI devices is taking shape[25][20].
2. China's individual export-licensing for rare earths and magnets has become a de facto rule subjecting international trade in core robot components to state approval; Beijing has, for example, required Tesla to guarantee the magnets are not used for military purposes — embedding end-use review into deal terms[13].
3. AI companions/toys collect sensitive data — always-on microphones, cameras, conversation logs — from homes and children, so alignment with child-online-protection and data-protection regimes is mandatory; researchers warn safety has not kept pace with market growth, making compliance a precondition for any device business[20].

Coming soon

Environmentally the issues are the footprint of rare-earth mining/processing and edge power efficiency. One humanoid uses 20-40+ NdFeB-magnet motors, so at scale permanent-magnet demand could reach ~100x today's, raising mining/refining impact and resource-circularity concerns; meanwhile low-power edge silicon like Jetson Thor (3.5x the efficiency of Orin) helps by shifting inference toward energy-thrifty on-device compute.

1. A humanoid needs 20-40+ NdFeB-magnet motors each, so at full volume permanent-magnet demand could reach roughly 100x today's level; rare-earth mining and refining carry heavy environmental costs, tying Physical-AI adoption directly to a resource/environmental bottleneck[11].
2. 🇺🇸 Jetson Thor delivers 3.5x the energy efficiency of Orin within a 130W envelope, running AI inference on-device without cloud round-trips; edge execution reduces network and data-center load and can help contain Physical AI's power footprint[2].

Coming soon