The ChatGPT Moment for Physical Worlds

Physical AI refers to intelligence that allows physical systems to perceive, reason, and execute on hardware or other embodiments. While software has traditionally scaled at near zero marginal cost, physical AI requires a cognitive reasoning layer, an onboard inference layer, and a real-time control loop on top of hardware, all operating within strict physical and safety constraints. By separating the software brain from the local hardware, industries like manufacturing, logistics, and robotics can finally experience scale, enabling new models like robots as a service and hardware-enabled software.

I believe that we are nearing the “ChatGPT moment” for physical AI given a confluence of trends: improved backbone VLA/VFAs, emerging world models, high fidelity physics simulators, and the collapse of hardware costs. These coalescing tailwinds enable a new wave of intelligence that can generalize across environments and embodiments.

However, as of today, the field is still rooted in research rather than commercial scale. When ChatGPT’s 3.5 model released in late 2022, it excelled at basic text drafting and coding, but struggled with riddles, temporal awareness, and hallucinations. Today's physical AI suffers from a similar disconnect. We have yet to see smooth, real-time task completion in generalized, unpredictable environments. Developers still face a significant shortage of data to train their systems, which is being worked on today. Bridging the "sim-to-real" gap still requires many breakthroughs across world model research, dexterity/manipulation, reinforcement learning, and scaled data collection. Despite these challenges that the industry faces, it seems we have much more clarity on what it takes to make the vision of Physical AI a reality.

Why Now for Physical AI & the Next Frontier

An Influx but Concentrated Talent Pool

Robotics and Physical AI has enjoyed record funding, raising a record $27.6B in VC dollars in 2025, more than doubling the prior year. The category then posted its strongest quarter on record in Q1 this year (raising $16.3 billion across 492 deals), lifted by megadeals in Shield AI, Saronic and Neura Robotics. The biggest funding categories have been in defense, world models, and infrastructure/foundation models.

The influx of funding has coincided with researchers commercializing their work. Senior authors of foundational robot-learning literature have spun out into venture-backed companies. For example, authors of “Open X-Embodiment: Robotic Learning Datasets and RT-X Models” (October 2023) and “RT-2: Vision-Language-Action Models” (July 2023) have later founded Physical Intelligence, Skild, World Labs, Jacobi Robotics, Dyna, Covariant, and Cobot.

With the influx of venture funding and new companies, the number of academic papers mentioning robotics has increased. NeurIPS’ embodied AI and robotics track has moved from niche to a major theme. The industry has organized around the thesis that foundation-model dynamics will transfer from language to physical control. The world model frontier has collapsed talent pools that were typically siloed across NLP, computer vision, and robotics, attracting talent across disciplines.

Hardware Commoditization and the Next Frontier

Historically, hardware could be its own moat due to proprietary mechanical engineering. Teams a decade ago had to build their own hardware or buy from low-volume manufacturers, previously accessible to funded labs. The dissolution of this moat significantly accelerated in the last few years given cheaper actuators, off the shelf commodity arms, and fleets of accessible robots from Shenzhen or US based integrators.

It has become vastly easier to design a robot and get it manufactured quickly, with lead times to get a prototype custom-manufactured from months to just a few weeks. Components like a global-shutter camera or high-frequency IMU now cost single-digit dollars, and a depth-sensing camera has fallen from thousands to $150. The platforms themselves have also collapsed in price: more than a dozen manufacturers now sell 6 and 7 DoF arms under $10,000. Full humanoid platforms fell from roughly $115K in 2024 to about $35k in 2025, with entry-level platforms like Unitree's G1 retailing near $13,500. Hardware has gone from a bespoke, capital-intensive differentiator to a standardized, modular vessel. The body can be bought off the shelf while value migrates to the intelligence layer.

Hardware R&D hasn’t stopped, but instead shifted to gaps in the market today like dexterous hand manipulation and tactile sensing. Contact-rich tasks like reorienting an object in hand, plugging in a cable, or handling deformables is where hard mechanical problems still exist. Early stage teams are tackling dexterous hands with different philosophies in actuation, like smaller gearboxes versus tendon-driven motors. Ultimately, hardware and software co-design, like matching a hand’s morphology to human-video data, has become more important, rather than optimizing hardware in isolation.

Achieving Real-Time Latency on the Edge

The robotics revolution has been enabled by the arrival of 1) energy-efficient edge processors (like NVIDIA Jetson Thor) that pack massive compute into tiny power envelopes, 2) the shift towards optimized local models like VLAs and small language models that eliminate cloud latency, and 3) direct-to-GPU sensor fusion. Together, these advances allow modern robots to process high-bandwidth sensory data, make complex decisions, and execute movements on-device.

Achieving real-time physical AI requires bridging a speed mismatch between low-level hardware and high-level intelligence. While low-level motor controllers execute rapidly at 100 to 1,000 Hz, high-level cognitive reasoning in VLA models remains a bottleneck, averaging just 3 to 5 Hz on edge hardware. Achieving fluid, human-like motion demands pushing this reasoning loop to a video-rate target of ~30 Hz, requiring a 6x reduction in decision latency.

First-generation VLAs (such as RT-2 and OpenVLA) relied on autoregressive, token-by-token generation inherited from LLMs, a bottleneck that left inference severely limited by memory bandwidth rather than compute power. To overcome this, modern architectures like Physical Intelligence’s model decouple high-level planning from low-level execution. This introduces a dual-system hierarchy: a deliberate “System 2” vision-language backbone that handles slow, high-level environmental reasoning, paired with an ultra-fast “System 1” flow-matching policy optimized for continuous, real-time physical control.

Foundation Models: VLAs and World Models

Foundation models serve as the “brain” of Physical AI trained on multimodal datasets (telemetry, video, text) to generalize across tasks and embodiments. Pioneered in July 2023 with Google’s RT-2 (developed by researchers who later founded Physical Intelligence), VLAs have evolved through three phases:

• Phase 1 (Monolithic LLMs): Autoregressive, token-by-token generation inherited directly from LLMs.
• Phase 2 (Decoupled Architectures): A System 1 / System 2 split separating slow, deliberate reasoning (“thinker”) from fast motor control (“reflex”).
• Phase 3 (Edge-Optimized Policies): Continuous action policies driven by diffusion and flow matching optimized for edge compute.

Nevertheless, VLAs still face significant hurdles with long-horizon planning, error compounding, and heavy compute requirements on physical hardware. To bridge these gaps, the concept of Vision-Force-Action (VFA) models has recently been introduced. Traditional VLAs rely on vision and language, missing critical physical properties like contact force, friction, and geometry. This renders delicate, contact-heavy tasks (like threading a lightbulb) extremely difficult. By integrating tactile and force sensing directly into action frameworks, research like Tactile-VLA equips robots with physical intuition, boosting plug-insertion success rates from 25–40% (vision-only) to 90%. These newly dubbed Vision Force Action models take into account touch feedback and physical control as another leap forward

In parallel, World Models take a simulation-first approach, learning physics like gravity and momentum directly from data rather than explicit code. Operating as internal simulators, these models allow robots to forecast environment changes before acting. Current state-of-the-art work spans generative video (Decart), 3D spatial representations (World Labs), and latent-space planning (AMI Labs). However, two key challenges remain: internet video lacks interactive physical feedback, and world models do not natively output low-level motor commands. The logical endpoint is a Unified World Model: a single foundation model that seamlessly integrates photorealistic rendering, accurate physics simulation, and real-time action sequences.

Overcoming Data Collection Challenges

While LLMs leverage the vast expanse of the public internet, Physical AI operates on a tiny fraction of that data. Over 99% of internet content consists of text, static images, and edited video—leaving only a sliver of 3D/4D spatial interactions. To build generalist robot brains capable of handling unstructured real-world environments, developers structure their data into a Robotics Data Pyramid, layering abundant, low-cost internet data at the base and high-fidelity teleoperation at the peak.

Today's robotic data pipeline relies on three main pillars:

• Internet Data: Offers broad visual diversity, but heavily edited web videos disrupt continuous cause-and-effect (the Markov property). To learn physical laws like gravity and friction through failure, teams instead source raw, unedited RGB-D video streams.
• Egocentric Data: First-person video data, recorded via smart glasses and headmounts, provide a low-cost, scalable pipeline. To overcome the embodiment gap, researchers use skill-capture gloves to log precise joint angles, velocities, and grip forces in everyday human interaction.
• Teleoperation: Though expensive ($118–$136/hour), human teleoperation provides ground-truth data, capturing synchronized motor commands, torques, and contact signals that video alone cannot convey.

Because no single data source is sufficient on its own, leading teams deploy hybrid architectures. For instance, 1X Technologies combines ~900 hours of egocentric human video for mid-training with ~70 hours of real-robot teleoperation for task-specific fine-tuning, reflecting an industry standard of 80–90% egocentric data mixed with 10–20% teleoperation. Similarly, startups like Sunday use skill-capture gloves to log home manipulation tasks at up to 100x lower cost than running a full robot fleet.

NVIDIA’s EgoScale framework, released in February, serves as a prime real-world validation of this data hierarchy in practice. EgoScale leverages pre-training on 20,000 hours of egocentric human video with a small “bridge dataset” (50 hours of human video paired with just 4 hours of robot teleoperation) to translate human sight into robot action. This hybrid approach yields a massive 54% improvement in task success rates over non-pretrained baselines.

However, EgoScale also exposes the ultimate boundary of passive data scaling. While passive egocentric video efficiently bootstraps a robot’s core capabilities, pushing reliability from 95% to 99.9% purely through video would require billions of hours of raw footage. To cross this reliability threshold practically, physical AI must shift to the apex of the data pyramid: active, closed-loop Reinforcement Fine-Tuning (RFT) in simulation. Pipelines like CosFly-RFT replace passive observation with automated execution loops, allowing the model to test actions, observe errors, and automatically update its weights in real time without human intervention.

A Global Industrial Labor Shortage

The urge to deploy physical AI into the real world is driven by a need for labor. According to global talent surveys, 72% of employers worldwide report difficulty finding the talent they need due to aging populations, falling birth rates, and gaps in technical skills.

The numbers in the United States alone show why the status quo is broken. Manufacturing faces 600,000 unfilled jobs today and a projected 2.1 million worker shortage by 2030. Logistics, warehousing, and utilities are constantly battling 500,000 open roles. Healthcare is short 1.25 million workers, headed for 1.35 million mid-year. In Agriculture, only 182 domestic applications were submitted for nearly 415,000 advertised farm positions.

In the context of physical AI and robotics in this economic climate, there is an opportunity to support industries of high turnover (30-40% in warehousing and logistics) or severe burnout (healthcare). Additionally, sectors like advanced manufacturing and utilities see a technical mismatch of specialized skillsets. As much of workforce in this sector is 25% to 40% over 55 years old and on the brink of retirement, the industrial world is not waiting for the human labor pool to recover, but heavily investing in Physical AI as a viable path to stay operational.

Where to Invest: Market Opportunities

Layer 4 — Embodied Labor Applications

The grand, long term vision for embodied AI is the opportunity to directly capture global labor spend, representing what is arguably one of the largest total addressable markets in tech history. Across every capital and labor-intensive vertical — manufacturing, supply chain, healthcare, agriculture, construction, defense, retail, and consumer services — embodied AI promises to transition fixed human labor costs into scalable, high-margin software and hardware solutions.

While the market remains in its early innings of deploying reliable physical systems in unstructured real-world environments, the economic trajectory is undeniable. According to consensus forecasts from industrial and financial leaders like BCG and McKinsey, the global commercial and industrial robotics market is projected to expand from roughly $90B in 2024 to between $200B - $260+ billion by 2030, laying the foundation for a fundamental transformation in how physical work gets done.

Compelling industry opportunities and example companies:

• Supply Chain and Logistics: Modern platforms such as Locus Robotics develop autonomous mobile robots (AMRs) and software for warehouses, distribution centers, and e-commerce fulfillment operations. Their systems work collaboratively alongside human workers to handle picking, case packing, and inventory replenishment, improving worker productivity. Collaborative Robot's Proxie also offers a robot as a service deployment model, offering assistance in assembly, quality inspection and warehouse automation (sorting, picking, of goods). Pure software vendors like Gather AI provides computer vision and intelligence to track and manage physical inventory in real time, creating a digital twin of the warehouse.
• Manufacturing: Companies like Machina Labs and Amca are redefining modern manufacturing by transforming rigid production floors into software-defined, agile ecosystems. Machina Labs utilizes real-time, physics-informed AI to transform generic robotic arms to produce metal sheets, incorporating sensor feedback and intelligence, while Amca uses existing manufacturing machines (like CNC mills, hydraulics, and electronics assembly tools) with AI to automate the engineering, compliance, and supply chain. Together, these two companies demonstrate software innovation via adaptive manufacturing networks.
• Agriculture: Robotics can automate tasks such as harvesting and inventory management and crop with greater speed and precision. Moss Robotics, founded out of CMU's Robotics Institute, builds autonomous ground vehicles and retrofits for specialty crop farms to digitize their famrs amid chronic labor shortages. Orchard Robotics' tractor-mounted camera system helps fruit growers maximize crop yield by informing decisions on pruning, thinning, labor, and harvest.
• Construction: Bedrock Robotics, founded by ex-Waymo engineers, retrofits hardware and installs software to turn standard equipment into self-driving machines to excavate job sites without human operators. Companies like Renovate Robotics are working on installing roofing shingles faster in an industry with contractor shortages. These companies aim to multiply the output of existing construction crews, enhance job safety, and accelerate build timelines.
• Defense/Government: AI is transforming the defense sector by shifting the paradigm of slow innovation due to hardware design and supply chain constraints. Neoprime companies are building autonomous systems to execute dangerous missions in harsh environments. Saronic builds autonomous surface vessels capable of long-range unmanned operations in contested, GPS-denied environments, winning a $392M Navy production contract. Tera AI, founded by Google X's former ML lead, provides software-only visual navigation that lets any robot or vehicle navigate GPS-denied terrain using only cameras.
• Retail/Hospitality: In retail, robots can assist with tedious inventory management, restocking, and fulfillment, while hospitality operators face chronic labor shortages and razor thin margins. Companies like Miso Robotics automate hazardous commercial fryers while Chef Robotics assembles ingredients for restaurants, hotels, and fine dining. The impact of these companies is to eliminate inventory waste, ensure operational predictability, and help operators focus on high-touch customer service.
• Healthcare: Robotic applications in healthcare span a wide range: surgical applications, new mobility products, patient support/interaction, elderly care and companions, and patient monitoring. Companies like Intuitive Surgical are known for their da Vinci surgical system, which is a robotic system that allows surgeons to perform complex procedures with greater precision and control.
• Consumer & Home Automation: Robots can automate household tasks, freeing up time and providing a more convenient lifestyle. Recent examples include folding laundry (Physical Intelligence), cleaning floors (Matic), and even cooking. Sunday Robotics' Memo, a wheeled mobile manipulator that learns chores like loading dishwashers and folding laundry from human demonstrations via its Skill Capture Glove, begins home beta deliveries in late 2026. Driven by the global aging population, governments and insurance companies are expected to heavily subsidize home care robots for the elderly.

Layer 3 — Embodied AI Infrastructure

The market is aggressively shifting toward specialized, off-the-shelf physical AI infrastructure. Unlocking general-purpose robotics requires an entirely new technology stack spanning spatial foundation models, automated data pipelines, and real-time reinforcement learning engines. By providing this connective tissue, infrastructure startups are allowing robotics teams to deploy commercially, faster, without having to reinvent the wheel at every step.

The primary bottleneck and the source of the deepest competitive moat is building true world models that master geometry, dynamics, and physical law. Cutting-edge research is rapidly blurring the legacy boundaries between perception, simulation, and execution. Instead of treating these as isolated steps, modern spatial intelligence allows a robot to perceive an object, simulate its weight and friction on the fly, and execute dexterous manipulation inside a single, continuous feedback loop. Infrastructure platforms that solve this physics grounding will prove immensely valuable.

Most investable areas:

• World Models & VFAs: The “brain” layer is the highest-value prize in the stack. Vision-language-action and foundation models from Physical Intelligence, Skild AI, and Dyna Robotics give robots general-purpose reasoning that transfers across embodiments and tasks, while world models from AMI Labs let machines predict geometry, physics, and dynamics before acting. State of the art research today is training AI systems to understand latent physics rather than pixels, along with physical feedback. We are likely a fe years away from pairing a VLA’s high-level semantic reasoning with a World Model’s real-time physical intuition into a single closed-loop system. Analogous to LLM labs in the language era, the model layer is where most value will concentrate. The platforms that master unified spatial intelligence become the operating system every downstream robotics application is built on.
• Reinforcement Learning: RL is the connective tissue that makes the rest of the stack work, turning demonstrations and simulation into policies that survive contact with the real world. The category is still early and largely embedded inside the big robotics labs rather than standing alone, but emerging players like Trajectory AI and Mechanize point to a dedicated policy-training and rollout layer, much as RLHF tooling became its own market in language AI. The investable moment arrives as robotics teams standardize on off-the-shelf training infrastructure instead of pipelines in-house.
• The Scale AI for Robotics: True physical intelligence demands rich, multi-sensor interaction streams: 3D spatial depth, high-frequency tactile contact feedback, force/torque dynamics, and proprioception. Companies like Sensei Robotics, Kled AI, and Encord are expanding beyond basic video annotation to build sensorized demonstration suits, teleoperation networks, and multimodal pipelines that capture how robots feel and apply force, not just what they see. Whoever is able to aggregate cross-embodiment datasets at scale for these rich physical interactions will see rapid revenue growth, although long-term durability and moats remain questionable.

  Layer 2 — Advanced Manufacturing

Most investable areas:

• AI-Native Manufacturers: The most compelling manufacturing opportunity is vertically integrated manufacturers that own the factory and intelligence layer. For example, Hadrian’s AI-powered precision machining factories to directly supply parts to the aerospace and defense industry directly. Nox Metals is AI-powered metals supplier, building automated factories that turn raw metal into production-ready material faster, cheaper, and more reliably than legacy service centers. Owning production end-to-end creates a proprietary data flywheel no software vendor selling into legacy factories can replicate, and reshoring mandates and defense supply-chain urgency provide durable demand tailwinds.
• Generative Design & CAD: The design layer is where manufacturing begins, and it’s controlled by slow-moving, seat-based incumbents: Autodesk, Dassault, Siemens, Ansys, Bentley Systems. Startups like AdamCAD work on LLMs to generate editable parametric geometry from text and images, collapsing the barrier between intent and design. On the engineering side, nTop and PhysicsX replace brute-force physics simulation with AI surrogates that evaluate designs in seconds instead of hours. As these converge, the winner becomes the system of record for how physical products are conceived. Unseating deeply rooted incumbents requires exceptional GTM strategy and innovating on business models beyond SaaS.
• Industrial Factory Automation: The shop floor faces a structural labor gap, and the new generation of robotics companies attacks it with foundation models than fixed automation. Mind Robotics, spun out of Rivian, is building full-stack industrial AI trained inside a live vehicle plant; Skild AI is developing hardware-agnostic robot and brain; and humanoid makers like Figure are piloting in real manufacturing environments. Manufacturing is the natural beachhead for generalist embodied AI: environments are semi-structured, tasks are repetitive but with edge cases, and ROI is measured in uptime rather than demos. The teams embedded in live production earliest win a compounding advantage via proprietary training data.

Layer 1 — Digitizing Physical Assets

Most investable areas:

• Physics-Informed Models: We are standing at the threshold of an “LLM moment” for physical engineering. For decades, hardware design and digital twin creation have been throttled by slow, legacy finite element analysis (FEA) and computational fluid dynamics (CFD) solvers. These solvers were historically built for CPUs with traditional physical equations across meshes, making it labor-intensive. Swapping these legacy tools (Siemens, Ansys, Bentley Systems) for physics-informed neural surrogates represents a massive improvement in simulation speed. Rather than solving partial differential equations iteratively from scratch, large physics models learn from thousands of high-fidelity historical simulations and datasets. These models can approximate physical behavior and predict stress, strain, velocity, and thermal profiles, shortening slow engineering design cycles. PhysicsX is the category’s flagship: its large physics models are already in production at GM, collapsing two-week aerodynamics cycles into minutes.
• Closed Loop Digital Workers: A winning thesis lies in bridging predictive asset intelligence with closed-loop physical control. First-generation “predictive maintenance” was passive, doing little more than sending an alert when a part was about to break. The future belongs to full-stack orchestration software that ingests real-time telemetry, runs continuous predictive simulations, and autonomously acts on physical infrastructure. Emerald AI is an early example: its platform turns data centers into grid-responsive assets, autonomously flexing power consumption in real time so operators earn from energy markets. By combining energy economics, predictive simulation, and direct asset control, these platforms control physical operations and establish themselves as the default system of record.
• Energy & Data Centers: The AI data center buildout is the largest infrastructure project of this generation, and energy is the binding constraint. Gigawatt-scale campuses waiting years for grid interconnection while demand for compute compounds. This bottleneck creates urgent, investable opportunities across the stack: software that squeezes more out of existing facilities, like Aravolta’s data center infrastructure management platform and Phaidra’s autonomous cooling control; thermal innovation like Corintis’s microfluidic chip cooling, which removes heat 3x more effectively than today’s state of the art; and entirely new power infrastructure, from American Terawatt’s off-grid generation and private HVDC transmission for AI training clusters to Base Power’s distributed home batteries. The companies that unblock power delivery, whether by optimizing, cooling, generating, or storing, sit upstream of the AI economy.

Founder Archetypes

• Scholar to Founder: Born out of top academic labs and deep-tech research institutions, these founders push the boundaries of embodied intelligence and physics-informed models. They understand that off-the-shelf software won’t cut it when interacting with the real world. They champion hardware and software co-design: building novel compute, sensors, or robotics architectures in tandem with algorithmic breakthroughs.
• Industry Veteran: This founder has spent years living in the trenches of brownfield manufacturing plants, data center facilities, power grids, or logistics hubs. They hold an unshakeable expertise moat earned through exposure to the messy, unglamorous realities of legacy operations: edge cases, safety protocols, and stubborn hardware retrofits. Rather than building tech in search of a problem, they work backward from acute operational pain points, acting as the translator that brings offline, analogue physical machinery into the modern era.
• Capital & Talent Magnate: Because Physical AI requires heavy R&D, specialized hardware prototypes, and massive compute clusters, it is uniquely capital-intensive. The Magnate archetype is a grand-scale visionary, capable of raising continuous, large-scale funding rounds to sustain long deployment cycles before commercial inflection. Beyond capital, their superpower is heterogeneous talent aggregation: they possess the gravitational pull needed to align radically different work cultures under one roof, uniting elite ML researchers, mechatronics engineers, supply chain leads, and enterprise sales leaders.
• Hyperscaler Alum: Forged inside tech-native physical companies like Waymo, SpaceX, Tesla, and Google X, this founder has already productionized autonomous hardware at scale. Unlike the Industry Veteran’s brownfield domain knowledge, their moat is execution velocity, plus the alumni networks to recruit proven teams overnight. They take hard-won lessons from a category-defining program and compressing years of trial-and-error into their first eighteen months.

Risks & Open Questions

Funding R&D Until the ROI Threshold

The uncomfortable truth about many physical AI companies today, particularly in robotics, are are still venture-funded research programs. Venture dollars pays for model development, data collection, and hardware iteration before a customer sees materialized ROI. In addition, reliability and auditability remains a black box. Below a 99% task success and accuracy level, a robot still requires human supervision and manual intervention. The core bet is that two curves cross in time: hardware costs continuing to decline to justify deployment, alongside model reliability to production-grade. Companies that cross that intersection convert R&D spend into revenue, while companies that don’t remain science projects with burn rates.

This is why, outside of structured environments like goods-to-person picking and palletizing, robotics is still barely penetrating most of the sectors on these market maps. Agriculture, construction, and healthcare remain largely in the pilot phase — not because demand is absent (the labor shortage data says the opposite), but because the technology is still mid-research-curve. Enterprise procurement requires uptime guarantees, safety certification, and service-level economics that most companies cannot yet underwrite. The open question for each vertical is not “will robots work?” but “how many more years of until it fully performs with a repeatable sales motion?”

Lessons from the Self-Driving Era

Between 2015 and 2021, VCs invested ~$45 – $55B over 7 years into autonomous vehicles. Physical AI & Robotics Era from 2023-2026 has well surpassed that clip in 3.5 years at $85 – $95B in funding. However, the AV trend followed similar logic used today of capturing labor spend, replacing drivers, and having software margins on a physical service. The timelines slipped by a decade. Argo AI shut down after $3.6B in backing; Cruise was wound down by GM after roughly $10B; the survivors were those with patient, strategic-scale capital (Waymo inside Alphabet) or those who narrowed in scope. The potential lesson here is that the last 9's of reliability are the entire game and take far longer than demos suggest. Capital intensity structurally favors strategics and sovereign-scale backers over traditional venture funds. And lastly, the “picks and shovels” vendors: chips, compute, sensors, and models might monetize faster than the application layer.

---

How to Play It

Entry Price Discipline for an Early Stage Firm

The model-layer race has been pricing out early-stage venture firms. Seed and Series A rounds for foundation-model robotics now routinely clear $1B+ valuations: Mind Robotics reached $3.4B within months of spinning out, and Physical Intelligence raised at multi-billion marks pre-revenue. A company seeded at a $1B+ post needs a $10B+ exit just to deliver venture returns to its earliest backers, and only a handful of acquirers (hyperscalers, NVIDIA, defense primes, auto OEMs) can write that check. Some outcomes like Covariant’s absorption into Amazon was closer to a licensing deal than a return event. For a Seed/Series A firm, the discipline is to avoid competing where megafunds have set the clearing price and instead enter at at layers where capital efficiency still exists, exloring areas like inference and compute optimization, developer tooling, and the energy stack.

The Hard Proof Points

A company showing three or more of these is past the research phase regardless of sector; a company showing none of them is asking investors to fund science. The evidence that separates a fundable inflection from an extended pilot:

• Production hours: Robots running 12–16 hours a day in a paying customer’s facility, with utilization holding after the novelty wears off
• Declining intervention rates: Task success above ~99%, with teleop interventions per hundred hours trending downward
• ROI: Sub-24-month customer payback at current hardware costs, with a margin structure that improves as components cheapen
• Pilot-to-production conversion and expansion: The same customer buying more units, where every deployed unit measurably lowers the cost of the next

The Most Practical Near-Term Bets

Weighing every layer of this thesis against those criteria, pockets of innovation like the data center and energy stack is a practical, investable corner of physical AI today. Demand is contracted rather than hypothetical; hyperscalers are projected to spend $600B+ on capex in 2026 alone, with forecasts approaching $1T by 2027. The buyer has an urgent, quantifiable problem on power, with ROI is measured in timelines and units (PUE, stranded megawatts, cooling efficiency) rather than future labor substitution. Software and thermal innovation (DCIM, autonomous cooling control, chip-level heat removal) fit venture check sizes and timelines today (versus power generation and project-finance profiles), while robotics applications slowly mature behind them.