Robotics & Physical AI

Neil Susan argues that investing in physical AI — automation, autonomy, and advanced manufacturing — can lift US economic growth back to over 5% annually, creating a structural tailwind for related equities and private ventures.

Physical AI requires managing massive sensor data at the edge (video, 3D, logs) from deployed robots in poor connectivity environments. Fox Glove's developer infrastructure solves this, attracting customers across autonomous vehicles (Waabi, Wave), warehouse robotics (Dexterity), and defense (Anduril, Shield AI). Traditional SaaS margins make it familiar to generalist VCs, driving valuation premiums.

A fruit-fly connectome mapped by Google is now simulated in software and connected to a Hugging Face SO-100 arm via OpenAI's Astra/Codex, enabling painting, Doom, and parallel parking — signaling a 'deep research moment' for robotics where cheap hardware + frontier models unlock DIY physical AI.

Lior Susan argues 85% of global GDP resides in physical industries (manufacturing, semiconductors, defense) now undergoing AI-driven automation and autonomy convergence; Eclipse's portfolio across compute, energy, robotics and defense is scaling rapidly, with IPOs like Cerebras and SpaceX proving the thesis and attracting LP capital for the reindustrialization opportunity.

Frontier models now transfer to robotic control with minimal adaptation; cost-vs-completion curves predict saturation of general-purpose manipulation within months, not years.

Expanding commuting radius via hybrid work and self-driving cars increases developable land geometrically (square of radius), potentially reviving the American ownership society model that historically underpinned political stability and middle-class wealth creation.

Cheap open-source robot arms (Hugging Face SO-100) combined with vision-language models like Astra enable rapid iteration on physical tasks — painting, mail sorting, weed pulling — lowering the barrier for a DIY robotics explosion akin to the 3D printing craze.

The 85% of global GDP in physical industries (manufacturing, semiconductors, defense) is being transformed by AI-driven automation and autonomy, creating a generational opportunity for startups and investors in robotics, specialized compute, and energy infrastructure.

RoboCurve benchmark shows GPT-6 controlling robot arms at 100% task completion using only raw video frames and joint actuation; cost-capability frontier saturating rapidly — humanoid general-purpose manipulation expected within months.

Progress in perception, reasoning, and planning for physical world interaction—exemplified by autonomous vehicles—signals robotics as the next frontier where AI meets the physical economy.

Robotics data management (logging, edge collection, cloud sync, debugging) is a horizontal SaaS-like business serving the entire physical AI ecosystem from warehouse robots to defense, attracting traditional VC interest despite hardware expertise gaps.

Ty shows that current vision-language models can plan and execute physical painting tasks on a low-cost open-source robot arm, suggesting a near-term 'deep research moment' for robotics where cheap hardware + agentic software enables widespread DIY automation of physical tasks like mail sorting and weeding.

Robot deployment is already scaling in China, indicating physical AI automation will extend beyond digital to real-world labor, with broad implications for manufacturing, services, and economic productivity.

Hugging Face's SO-100 open-source 3D-printable robot arm (~$200 actuators) combined with frontier models (Astra) enables low-cost physical AI experimentation; YC humanoid under $2K and Nat Friedman's leaf robot suggest near-term consumer robotics boom for tasks like mail sorting, weeding, and cooking, with models already capable but speed/latency the remaining bottleneck.

Lior Susan argues the convergence of AI, automation, and autonomy across physical industries represents a generational opportunity to reindustrialize the US, with Eclipse's portfolio spanning semiconductors, robotics, and energy infrastructure.

Physical AI progress mirrors digital AI: perception (world modeling), reasoning (multi-step planning), and action (motion control) are all improving — autonomous vehicles and robotics are leading indicators of AGI's physical manifestation.

Hugging Face's SO-100 arm (~$200) and sub-$2K humanoids from YC companies, combined with models like Astra that can plan and execute physical tasks via Codex, lower the stakes for experimentation; near-term use cases include mail sorting, weed pulling, and cooking — an IRL spam filter — with models already capable of object recognition and manipulation.

Physical AI spanning manufacturing, semiconductors, defense, and robotics is where automation meets autonomy, potentially enabling US GDP growth above 5%. Companies in semiconductors, energy, and robotics are becoming large businesses for the first time, making this the best moment in history to build in physical AI.

Frontier models (Astra, GPT-6) plugged into robotic arms with vision achieve saturated manipulation benchmarks; cost-capability curves predict general-purpose humanoid robot tasks solved by GPT-6.5/6.1 within months, collapsing the embodiment gap.

Many software-focused VCs are exploring robotics investments after Nvidia's Jensen Huang highlighted physical AI, but most admit they don't invest in hardware, creating a gap between interest and capability that favors specialized firms like Eclipse and infrastructure plays like Foxglove.

Managing dense sensor data at the edge is a universal bottleneck for robotics companies; Fox Glove's developer infrastructure solves logging, upload, and debugging across autonomous vehicles, warehouse robots, and defense systems, attracting both specialist (Eclipse) and generalist (Bessemer) VC interest.

Ukrainian drone production cycles compress battlefield feedback into hardware redesigns within weeks, proving that software-defined manufacturing and autonomous systems are the new paradigm for defense and industrial production.

Whole-brain emulation of a fruit fly (via Google connectome) now runs in simulation and drives a $200 Hugging Face SO-100 arm, signaling a 'deep research moment' for robotics where cheap hardware + foundation models enable rapid DIY experimentation.

Physical AI — combining automation, autonomy, and AI in the physical world (manufacturing, semiconductors, defense, robotics) — represents 85% of global GDP and offers a historic opportunity to reindustrialize the US; Eclipse's portfolio spans compute, energy, defense, and manufacturing companies becoming large businesses.

Matteo argues that AI advancements enable humanoid robots to perform complex tasks, creating a trillion-dollar market that traditional automakers must enter or lose to competitors, citing China's rapid progress in humanoid Olympics and robots outperforming humans in athletics.

Fox Glove's developer infrastructure for robotics data attracts traditional SaaS investors because it offers understandable margins while providing exposure to the entire physical AI ecosystem from warehouse robots to defense.

Cheap commodity hardware + frontier vision-language models enable rapid robotics prototyping; the 'deep research moment for robotics' is arriving as agents close the IRL-to-digital loop (mail sorting, weed pulling) without requiring $50k humanoids.

85% of world GDP is physical industries (manufacturing, semiconductors, defense); AI-driven automation and autonomy meeting in these sectors creates unprecedented opportunity for US reindustrialization, with compute, energy, defense, and robotics companies becoming large businesses.

RoboCurve benchmarked GPT-6 controlling a robotic arm with visual input: near-100% completion on pick-and-place. Cost-vs-completion frontier collapsing predictably. Wezner predicts GPT-6.5/6.1 will solve general-purpose humanoid robot tasks within months. Video generation and robotic embodiment converging — video gen becomes just another output modality in GPT-8/9.

Chinese EV makers like Xiaomi and BYD are already expanding into components and robotics; Franceschetti predicts all Chinese automakers will become humanoid robot companies within years, making humanoids the next automotive-scale market, as evidenced by China's 'Olympics of robots' showing rapid progress.

Progress in robotics and AVs demonstrates AI mastering physical world modeling (perception), multi-step reasoning, and motion planning — the three pillars of physical intelligence — unlocking labor automation in manufacturing, logistics, and transportation.

Investor enthusiasm for physical AI — AI moving into the physical world to perform real-world tasks in homes, factories, and warehouses — is driving astronomical valuations for robotics startups, particularly humanoid and foundation model companies, pricing some out of early-stage lists.

The developer infrastructure layer for physical AI is structuring as a traditional SaaS business with familiar margins and pricing, making the broader robotics boom accessible to generalist investors who understand software economics but not hardware complexity.

Investor enthusiasm for physical AI — AI moving into the physical world via robots — is driving astronomical valuations for robotics startups, with some humanoid and foundation model companies priced out of The Information's TI50 criteria.

Robots are already scaling in China, and as physical AI converges with foundation models, automation will move from digital to physical worlds across labor, manufacturing, and domestic tasks — creating a parallel investment cycle to data-center buildout.

Maritime autonomy has almost no open-source perception datasets because commercial autonomous boats lack regulatory approval (manning requirements persist) and acoustic/sub-surface data gets classified quickly. Combined with saltwater, biofouling, bird nesting, and 20G+ accelerations at the air-water interface, this forces full-stack proprietary development — a high barrier but also a moat for early movers.

Gary Tan reports hard tech companies now comprise 25% of YC batch (40x increase), as AI tools like GPT-6 Astra enable small teams to solve complex engineering problems previously requiring massive capital and specialized hires.

Antonio Lee argues ultra-low-cost humanoid robots (80-90% of home tasks, 1.5kg lift) serve as consumer wedge to accelerate AI development for robotics, with coding agents like GPT-5 Astra enabling one-shot task deployment.

Unlike Figure or 1X, Tesla can deploy thousands of Optimus robots in its own manufacturing facilities and warehouses to iterate on hardware and software before external commercialization, a vertical integration advantage that accelerates learning cycles.

Fleets of 250-lb autonomous drones carrying foam can deliver suppression power equivalent to a wildland engine within 10 minutes, showcasing robotics and AI applied to high-stakes physical-world problems.

Robot taxi services already widely available in San Francisco and LA could scale massively, tapping into the enormous transportation industry and creating huge opportunities for companies with autonomous vehicle and AI capabilities.

Kempczinski explicitly states he does not foresee a future without people in restaurants, instead framing AI as a co-pilot to make jobs easier and remove stress, suggesting labor-intensive service sectors will pursue augmentation over replacement.

Stewart Landersberg contends that autonomous drone fleets responding in 5-10 minutes can stop catastrophic wildfires, without which the American West faces an existential cascade of lost home insurance, mortgages, and community viability—making fire-tech a necessary infrastructure investment.

The same edge computing chips, AI models, data infrastructure, and electric powertrain technology developed for autonomous vehicles transfer directly to humanoid robots and flying cars, creating a compounding R&D advantage for full-stack players like XPeng.

To capture abundance from automation, Immad advocates a massive infrastructure program building 100M robots owned by the people (via sovereign wealth funds), upgrading US infrastructure while distributing ownership. Elon Musk stated at G20 that one robot does the work of five humans, confirming the labor displacement trajectory.

Robotics and logistical automation adoption is just starting; Symbotic's backroom automation for Walmart demonstrates the trajectory with over 100% earnings growth.

Physical AI adoption in logistics/back-office automation just beginning; Symbotic's Walmart partnership demonstrates scalable robotics deployment with triple-digit earnings growth — early innings of industrial AI transformation.

Tony Xu describes a 'war on atoms' where companies like DoorDash organize offline information and logistics, distinct from the digital 'war on bits' for consumer attention, and predicts increasing intersection between the two arenas over time.

Commercializing robo-taxi at scale (Tesla 2025) generates hundreds of billions in cash flow to fund the humanoid robot problem, which requires multi-hundred-billion-dollar AI compute investment. Despite architectural improvements, humanoid robots at meaningful scale are a 2029-2030 event. The robo-taxi transition transforms Tesla from car sales to platform revenue on miles driven.

Antioch's blend of classical simulation and learned gap models closes the sim-to-real gap, enabling data-efficient robotics training; partnership with Amazon Ring shows applicability beyond humanoids to smart devices, drones, vacuums — any hardware/ML system where real-world testing is expensive.

Klein references Chinese warehouses with 95% robotic automation as evidence that physical AI is already reshaping industrial workforces, implying similar transformation coming to factory maintenance and logistics globally.

Elevators are being equipped for autonomous robot and humanoid operation to deliver services to aging populations, with early lab proofs demonstrating the building infrastructure layer for physical AI.

The Pentagon's Drone Dominance Program aims to procure 3 million drones, implying 120K-240K headset displays at a 1:20 ratio; Kopin supplies displays to 5 of the top 10 competitors and is installing OLED deposition capacity to meet this demand, providing a near-term volume driver independent of the longer-term AR headset program.

Removing crews from vessels allows radical design simplification, enabling mass production, cost reduction, and new market creation — a structural shift comparable to autonomy in other domains.

Tesla's Optimus humanoids (22 DOF hands) will precede humans to Mars for base construction; Chinese humanoid videos show rapid progress; Bank of America estimates 1.2M humanoids by 2030 and 10M by 2035, potentially justifying Tesla $3T valuation and SpaceX Mars economics.

Robotics faces exponentially larger action spaces than self-driving, scarce teleop data, and severe cross-embodiment gaps (policies don't transfer across hardware); solving this requires world models pre-trained on internet video plus small action-conditioned fine-tuning, but tactile/epidermal sensing parity with humans is still missing.

Terra Firma's founder argues retrofitting existing Caterpillar and Deere fleets with an AI intelligence layer solves construction's fragmentation and labor shortage, enabling small teams to orchestrate large robot fleets for infrastructure on Earth and Mars.

Physical AI constraints (battery, edge compute) favor fine-tuned, task-specific models over general foundation models, leading to a SaaS-like explosion of vertical robotics companies rather than a few dominant labs.

Cosmos world foundation models solve physical AI's data scarcity by generating robot-perspective synthetic data from compute — a shared multimodal representation across embodiments enables transfer learning because all robots operate under the same physics, making robotics capabilities scale with GPU cycles rather than fleet collection.

Cosmos Dreams demonstrates action-conditioned world models running real-time sensor simulation on a single RTX Pro 6000 (down from 64 GB300 GPUs), letting autonomous vehicle developers do passive verification and rare-scenario training entirely in simulation before fleet deployment.

Physical AI requires real-world deployment data to handle edge cases, creating a chicken-and-egg problem that favors companies with deployed fleets; a single 'robotics brain' is decades away due to hardware integration and safety requirements.

Training world models on diverse video/game/simulation data then fine-tuning for specific embodiments (cars, humanoids, drones) dramatically reduces real-world data requirements versus vision-language-action models, making general-purpose robotics economically viable.

Both Tesla (Optimus, Cybercab, robotaxi) and Mobileye (robotaxi commercial deployment by year-end, humanoid robotics) are positioned as rare companies spanning autonomous vehicles and humanoid robotics; the market is pricing them for perfection but the physical AI thesis remains a show-me story with near-term capex dilution.

Warehouses are designed for forklifts, not humans, making them natively robot-friendly; Dexory's no-infrastructure-needed deployment (2-5 days) enables rapid scaling in existing 250K+ global warehouses without requiring customers to rebuild facilities.

Dexory's 1.3B pallet scans and 70K days of real warehouse operations create a unique 'machine-ready' dataset that enables other robots (forklifts, pickers, humanoids) to operate without synthetic simulation, giving a structural advantage in physical AI deployment.

The explosion of cheap drones from Ukraine-Russia war production makes directed energy (Epirus high-power microwave, lasers) the only affordable countermeasure — traditional interceptors are too expensive per shot, so scaling directed energy to protect bases, ships, planes, and major events is a funded priority.

While US labs (Google, OpenAI, Anthropic) abandon consumer video generation for robotic world modeling, Elon's Grok Imagine aims to generate full-length movies by year-end as a stepping stone to 'Digital Optimus' — a pixels-to-actions computer-use agent that navigates screens like humans navigate the physical world, leveraging video understanding for autonomous knowledge work.

Current surgical robots use proprietary, non-interoperable connections limiting them to Level 1 teleoperation. Nvidia's data-centric approach — Ethernet backbone, Holoscan Sensor Bridge converting any sensor signal to RDMA-enabled Ethernet, and IGX edge compute — reduces latency from 50ms to 1.2ms, enables seamless sim-to-real transfer, and creates a pathway to Level 4-5 autonomy by making sensor/actuator upgrades as simple as adding nodes to a data center.

Building human-like hands, feet, and eyes is extraordinarily difficult and will take a very long time; even if solved, consumer applications like folding clothes don't justify trillion-dollar investment, making Optimus a persistent cost center with no revenue visibility.

Building human-like hands, feet, and eyes is extraordinarily difficult without evolutionary models; Optimus will take a very long time and consumer applications like folding clothes don't justify trillion-dollar investment, though military/space uses make sense.

Jensen states Nvidia's robotics/autonomous vehicle business is already ~$10B and will become the next $100B business within 10 years. The 'ChatGPT moment for robots' occurred years ago with sim-to-real reinforcement learning. Self-driving cars provide the scale to kickstart the flywheel, and Nvidia is open-sourcing its full stack (Alpaca Myo) for agriculture, delivery, warehouse AMRs.

Travis Kalanick's Atoms ($1.7B raise, a16z-led) targets full supply chain automation via robotics + real estate; OpenAI develops consumer hardware (Codex keyboard, pendant); Tesla FSD regulatory delays in Europe contrast with Chinese ADAS approvals — physical world automation requires massive capex but offers defensible moats beyond software.

Humanoid robots (Optimus, Figure) will be smuggled onto early uncrewed lunar landers within 4-6 years to perform dangerous infrastructure build-out, logistics, and resource extraction, making human EVAs the exception rather than the rule.

The data market for robotics (RL environments, simulation, real-world physical data) is currently small relative to GenAI and autonomous vehicles but will grow substantially as labs solve simulation-to-real transfer; Mercor is positioning early to capture this demand which requires high-fidelity environments mimicking deployment settings.

The convergence of humanoid hardware approaching human form factor and AI pre-training on internet video data is accelerating rapidly. The CEO of 1X believes we are less than a decade away from hard takeoff where robots build robots, data centers, and chip fabs, creating true labor abundance. His personal bet is 3 years, with 10 years as the outer bound.

At 20 hours/day operation and a 5-year lifespan, a humanoid robot delivers ~40,000 hours of work, bringing effective labor cost to ~$1/hour versus $20-40/hour for human factory workers. This 90% cost compression gives robotics companies enormous pricing headroom while still undercutting human labor, with pricing remaining inelastic for a long time because it is anchored to human wage rates.

Robotics will have its breakthrough moment when users can directly interact with robots performing useful tasks, similar to ChatGPT's accessibility. Without robotic automation, humans become mere actuators for cloud AI — a 'very bad' outcome. Progress is accelerating toward general-purpose robotic manipulation.

Inum builds scalable sensing infrastructure for undersea domain awareness, using AI-powered sonar buoys and C2 platforms to detect submarines and unmanned underwater vehicles in real time, filling a critical gap in maritime defense.

China's 130+ robotics companies and 30-40 near-term IPOs reflect manufacturing scale advantages; the winning paradigm may mix Chinese hardware bodies with Western foundation model brains, accelerating commercial humanoid deployment.

Critical drone components like gimbalized thermal cameras, high-energy-density batteries, and brushless motors lack domestic US suppliers, creating opportunities for new industrial startups.

Agents perform net-new work humans never did (reading all contracts for upsell signals, generating 5,000 personalized ads vs 5), expanding total economic output rather than replacing labor; robotics will extend this to physical world.

The FCC ban on foreign humanoid robots removes Chinese competitors like Unitree from the US market, but the bigger bottleneck for American humanoid companies is finding profitable commercial use cases. Capital is not the constraint — compelling teams have raised hundreds of millions to billions — but deploying that capital into genuine value creation in homes, workplaces, and factories remains the challenge.

Administration limited importation of non-US humanoid robots to force homegrown supply chain; cites drone import restrictions (Dec 2024) that triggered dramatic investment surge in US drone supply chain as model for robotics.

AI is moving from digital to physical worlds; robotics represents a massive, early-days opportunity where AI becomes the operating system for legacy industrial institutions.

Current voice systems cannot distinguish between multiple speakers or filter background noise, a fundamental barrier for humanoid robots and voice agents that must be solved before widespread deployment.

Traditional robotics is limited to single-function, expensive systems only large enterprises can afford. Foundation models for robotics provide the 'brains' for flexible, thinking, safe interaction — a pre-consensus area where Axiom is investing to democratize physical AI for 8 billion people, not just developers.

Atlas Motion leverages internal co-design software and standardized raw material inputs to collapse motor design iteration cost, enabling high-mix production for both enterprise primes and early-stage robotics companies; this mirrors a broader shift where AI-driven design tools eliminate traditional NRE barriers in hard tech manufacturing.

Humanoid robot deployments could scale from 60K units in 2026 to 13M annually by 2035 ($200B market, part of $1T physical AI), but progress is constrained by lack of physical training data — simulation hits a 'wall' on real-world messiness, requiring massive real-world data collection efforts.

Distributed autonomous vessels with onboard ML for navigation, sensor fusion, and counter-UAS can persist in contested 20-mile straits where manned ships face asymmetric threats from mines, drones, and fast attack craft.

Shield AI model — drones/cameras commoditized, value in AI layer; must invest 5-10 years before momentum, not during hype cycle.

Ryan argues that even if AIs remain poor at social/political tasks, superhuman capabilities in verifiable engineering domains (chip design, manufacturing, robotics) would enable an 'industrial explosion' radically transforming the physical economy.

Robotics lacks internet-scale training data. Companies like Sunday solve this by co-designing robot hands and teleop gloves for high-fidelity data, then using pre-training + post-training flywheel — mirroring the Waymo/Tesla playbook for autonomous vehicles.

The next major leg of AI compute demand will come from robotics, requiring massive video data for training; Rumble's 57M monthly active users and video platform provide a proprietary data moat that typical neoclouds lack.

Neros is adding AI-powered terminal guidance (Archer AI) so drones can complete attacks after comms are jammed near targets; this hybrid manual/autonomous approach addresses the gradient of comms degradation in contested environments.

Unlike LLMs, robotics requires constraining physical problems (factory vs construction site vs warehouse) before applying models; each vertical's nuanced physics and task variation favors specialized solution stacks over general-purpose foundation models, keeping the sector fragmented for years.

China's robotics demos prioritize kinetic feats (running, backflips) while US players (Figure, 1X, Optimus) focus on factory manipulation — unclear which approach wins commercial deployment, but pace gap is striking.

Chinese robotics companies showcase speed and agility (running, dancing) while US counterparts target fine motor skills for manufacturing, suggesting different near-term commercialization paths.

Treating reasoning as a latent variable and using variational inference to score reasoning traces for concision, non-triviality, and action predictiveness automatically discovers optimal reasoning types per embodiment (move+gripper for manipulation, structural affordances for locomotion, meta-action for driving), improving success rates and data efficiency across 1B-30B parameter models.

End-to-end robotics application companies (data centers, warehouses, manufacturing, food) that start with teleop on off-the-shelf hardware, iterate fast with real customers, and build operational moats beyond modeling will unlock massive markets with relatively small capital requirements.

Decomposing memory into short-term dense visual frames for dexterity and long-term compressed language representations for task progress allows policies to handle 10+ minute tasks, wait appropriately (grilled cheese), and recover from mistakes without retraining — a critical missing piece in current VLAs.