A note from the Dhow Team: this edition of the Dispatch is a bit different from our typical founder deep dives; we want to showcase more of the level of insight we will provide through the Dhow platform, which will launch in April. It’s going to be a long one, but it’s well worth it. We sincerely appreciate your support and hope you enjoy the newsletter. Please hit the ad at the bottom of the newsletter; it helps us out a ton.
🚀 TL;DR
Physical Intelligence is building a general-purpose AI control system intended to power robots across tasks and hardware platforms
Advances in large-scale AI models may now allow robot intelligence to generalize beyond task-specific programming
The company has raised ~$1B at a reported ~$5.6B valuation, with no publicly disclosed revenue as of early 2026
Commercial traction appears to be at the pilot stage rather than scaled enterprise deployment
The upside depends on whether value accrues to the intelligence layer rather than hardware or integrators
The core risk is that real-world deployments remain customization-heavy, limiting scalability and margins
What to watch: whether deployments scale with declining integration costs, improving real-world performance, and hardware-agnostic adoption before incumbents close the gap
📇 Snapshot
Founded: 2024 (Co-founded by Adnan Esmail)
Founder highlights: Former leader at Tesla (worked on Autopilot tech) and Anduril Industries (SVP of engineering for defense systems and autonomous hardware/software integrations)
HQ: San Francisco, CA
Sector: Robotics Software / AI-driven foundational models for robotics
Latest raise/valuation: Raised $600M at a $5.6B valuation in November ‘25
Revenue: No public revenue estimates
Commercial Stage: Pilot deployments and early industry testing across logistics, grocery, and manufacturing contexts
Customers: No named enterprise customers publicly disclosed
Positioning: Building AI systems that allow robots to learn and perform multiple physical tasks without being manually programmed for each one
Source: Forge
Source: Forge
Elevator pitch: Physical Intelligence is building general-purpose AI models that allow robots to perceive, reason, and act across different tasks and hardware platforms without being manually programmed for each specific job.
🧠 The Investment Question
The core question is straightforward:
If robots become meaningfully more capable over the next decade, where will the economic value accrue, and can Physical Intelligence own that layer?
In robotics, there are typically a few layers to consider:
Hardware - motors, sensors, and arms
Low-level control - moves joints and executes precise movements
Intelligence layer - the AI system that interprets the sensory inputs of the robot (primarily sight) and decides how to interact with + adapt to a changing environment; we believe it’s the key infrastructure to scale
In simpler terms, the intelligence layer is the “brain” of the robot, turning perception into action. If that “brain” can be reused across a variety of tasks and robot bodies, it becomes a high-leverage position in the stack. However, if it requires heavy customization across every situational deployment, it ends up falling into the same category as traditional robotic software
The core concept to follow here is whether Physical Intelligence is building a reusable, scalable intelligence layer that compounds in value over time, or a sophisticated but deployment-specific robotics system.
👤 Founding Team and Backgrounds
Adnan Esmail
Adnan Esmail (Co-Founder) brings deep systems and engineering expertise from hardware-intensive technology environments, including senior engineering leadership roles at Tesla and Anduril. A Massachusetts Institute of Technology graduate, he has blended hardware and software systems thinking throughout his career and now focuses on building general-purpose robot intelligence.
Karol Hausman (CEO & Co-Founder) previously served as a staff research scientist at Google DeepMind and adjunct professor at Stanford, specializing in robotics manipulation and reinforcement learning.
Sergey Levine (Chief Scientist & Co-Founder) is a renowned robotics and deep reinforcement learning researcher, professor at UC Berkeley, and author of foundational work in robot learning.
Chelsea Finn (Research Lead & Co-Founder) is a computer scientist and Stanford faculty member known for work in meta-learning and sim-to-real knowledge transfer in robotics.
Lachy Groom (COO & Co-Founder) brings product leadership and go-to-market experience from Stripe and angel investing, focusing on scaling operations and partnerships.
Brian Ichter (VP Engineering & Co-Founder) has a background in robotics research and engineering at Google and related teams focused on control and experimentation.
Other early team members include researchers and engineers from top AI labs and robotics backgrounds, contributing to model development, systems integration, and real-world testing.
📈 Why PI, and Why Now
For the better part of the past two decades, robotics has been constrained less by hardware and more by intelligence. Robots were able to move precisely, but they struggled to adapt outside controlled environments, with each new task typically requiring new programming.
However, there are 3 main shifts right now changing that:
The first is that advances in large-scale AI models have improved the ability to map perception to action. Techniques developed for language and vision are now being applied to physical control systems, increasing generalization across tasks.
Second, simulation environments and synthetic data have improved meaningfully, allowing models to train on large volumes of interaction before deploying in the real world. This has historically been a bottleneck in robotics experimentation.
Third, labor pressures in logistics, manufacturing, and service sectors have increased willingness to experiment with automation in semi-structured environments.
The timing question is whether these shifts are sufficient to support a reusable intelligence layer, as opposed to the incremental robotics gains that we’ve seen over the years.
Physical Intelligence was founded in this moment, not before it. The company is attempting to build generalized robot control models at a time when capital, research talent, and computing are concentrated around large AI systems.
The growth opportunity exists because robotics may finally be able to borrow scaling laws from AI. The risk is that physical-world complexity resists that translation.
🤖What Physical Intelligence Is Actually Building
Physical Intelligence is developing a general-purpose AI control system designed to operate robots across a range of tasks and hardware platforms.
This reinforcement learning (RL) framework refines the robot’s "Action Expert" through real-world practice and human intervention. By closing the loop between perception and movement, Physical Intelligence enables models to master complex, autonomous tasks - like box assembly and laundry - with increasing precision.
Instead of programming robots for individual, task-specific workflows, the company is training large models that connect various sensory inputs (vision and language) with physical action. The goal is for a robot to interpret its environment, understand an instruction, and execute a multi-step task without bespoke code written for that specific scenario.
This architecture enables robots to perceive and act across diverse tasks and hardware platforms. By merging internet-scale AI with a specialized Action Expert, Physical Intelligence creates a reusable intelligence layer that generalizes skills - from folding laundry to bussing tables - without the need for task-specific programming.
In more practical terms, this means:
The same underlying model should be able to support different robot bodies
The system should generalize across tasks rather than requiring one model per use case
Performance should improve as more data is collected across deployments
It’s important to note that Physical Intelligence is not positioning itself as a hardware manufacturer. Their ambition is to sit above the mechanical layer and provide the intelligence that powers it.
This collage demonstrates "Cross-Embodiment" intelligence. The same underlying model is shown simultaneously controlling different robot hardware to perform distinct, non-programmed tasks—from folding laundry and assembling boxes to delicate kitchen work. This is the visual proof of Physical Intelligence’s "Generalist" thesis.
The critical question is whether this control system meaningfully reduces the engineering required per deployment, or whether customization remains necessary beneath the surface.
🛡 Market Structure and Economic Context
The robotics market is fragmented across hardware manufacturers, systems integrators, and application-specific solution providers. Historically, most economic value has accrued either to hardware suppliers with manufacturing scale or to vertically integrated automation providers serving specific industries. Software has played a role, but typically in narrow, task-specific contexts.
If Physical Intelligence succeeds, it is attempting to insert itself at a different point in the value chain: the shared intelligence layer that sits above hardware and below applications.
The economic logic of that position is attractive. A reusable control system that works across hardware vendors could benefit from:
Cross-deployment data improvement
Lower marginal cost per additional task
Platform-level integration into enterprise workflows
However, there are structural constraints to note:
First, robotics buyers are not early adopters in the way software buyers often are. Purchasing decisions in manufacturing and logistics are risk-sensitive and ROI-driven.
Second, hardware fragmentation complicates standardization. Unlike cloud infrastructure, robotics lacks uniform interfaces, which can increase integration cost.
Third, incumbents with capital and computing resources may attempt to internalize similar intelligence capabilities, especially if they control hardware distribution.
The economic question is therefore not just whether generalized robot intelligence is possible, but whether it can occupy a defensible and high-margin position within a fragmented and capital-intensive ecosystem.
📊 Technical Differentiation and Replicability
Physical Intelligence’s core claim is that large-scale AI models can be trained to generalize physical control across tasks and hardware platforms, reducing the need for task-specific programming.
The potential differentiation rests on three core pillars:
The first is model architecture. If the company has developed a vision-language-action model that meaningfully improves task generalization compared to prior robotics approaches, that creates an early technical advantage.
Second is data. In robotics, real-world interaction data is difficult and expensive to collect. If Physical Intelligence is building a proprietary dataset of multi-modal physical interactions that improves model performance across deployments, that could become a compounding asset.
Third is systems integration. Translating model outputs into reliable real-world motion requires tight integration between AI, sensors, and hardware constraints. Execution at this layer can create defensibility beyond pure model design.
Physical Intelligence’s research proves that as models scale, they "emerge" with the ability to translate human motion into robotic action. By bridging this gap, robots can learn complex tasks simply by observing humans—effectively turning infinite video data into a 2x performance gain without the need for manual programming.
However, it’s important to note that replicability risk remains high. Large incumbents with access to compute, robotics talent, and distribution could attempt to build similar models. Open research progress in robotics and reinforcement learning reduces the probability of long-term exclusivity at the algorithmic level.
The durability of the differentiation therefore, depends less on the novelty of the technology and more on the speed of iteration, data accumulation at scale, and depth of deployment integration
If those can compound, the advantage strengthens. If not, the technical moat narrows quickly.
💰 Business Model and Early Financial Signals
As of early 2026, Physical Intelligence has not publicly disclosed revenue, pricing structure, or recurring revenue metrics. The company appears to be in early commercialization, with pilot deployments rather than scaled enterprise rollouts. The expected business models likely fall into three categories:
First is licensing the intelligence layer to robotics manufacturers or integrators. This would resemble a software licensing or usage-based model, with higher gross margin potential and lower capital intensity.
Second is vertical integration, where the company embeds its models into full-stack robotics solutions. This increases control but raises capital requirements and compresses margins.
Third is hybrid partnerships, where revenue comes through structured deployments tied to specific industry use cases before broader platformization.
The absence of disclosed revenue suggests the company remains in a validation phase rather than revenue optimization mode. Capital raised to date, approximately $1 billion, indicates investor confidence in long-term platform potential rather than near-term cash flow.
For investors, this creates a familiar tradeoff. The upside profile is driven by control of the intelligence layer. The risk profile is shaped by capital intensity, commercialization speed, and time to repeatable revenue.
At this stage, the key financial signal is not ARR, but whether deployments reduce integration cost over time and demonstrate improving model generalization.
🏗️ Durability and Defensibility
The durability of Physical Intelligence depends on whether its control system becomes embedded infrastructure or remains advanced tooling.
The way we see it, there are three potential sources of defensibility:
First, data accumulation. If real-world deployments generate proprietary interaction data that meaningfully improves model performance across tasks, the company could build a compounding advantage. In robotics, high-quality physical-world data is scarce and expensive, which increases its strategic value.
Second, integration depth. If the intelligence layer becomes tightly integrated into customer workflows and hardware stacks, switching costs increase. Over time, this can shift the company from experimental vendor to operational dependency.
Third, model iteration velocity. In AI-driven systems, advantage often accrues to teams that iterate faster and translate research into deployment more efficiently than competitors.
However, each of these advantages is conditional. If hardware vendors internalize similar capabilities, switching costs weaken. If open research narrows technical gaps, algorithmic advantage erodes. If deployments remain shallow or pilot-stage, data accumulation may not scale meaningfully.
Durability, therefore, depends less on the novelty of the model and more on whether usage compounds. Without scaled deployment, defensibility remains theoretical.
The central question is whether Physical Intelligence can convert early technical progress into embedded infrastructure before competitors close the gap.
⚠️ Key Risks and Failure Modes
The primary risk is that technical progress in controlled environments does not translate into reliable real-world performance. Robotics introduces variability that does not exist in purely digital AI systems. If generalization breaks down outside pilot settings, commercialization slows materially.
A second risk is capital intensity. If deployments require significant hardware integration or custom engineering per customer, growth may remain linear rather than compounding. In that scenario, margins resemble traditional robotics businesses rather than scalable software platforms.
Third, competitive compression is real. Well-capitalized incumbents and research labs are investing heavily in similar model architectures. If foundational advances are broadly replicable, differentiation may narrow faster than deployment scale can compensate.
Fourth, buyer conservatism in industrial markets may delay adoption. Logistics and manufacturing operators prioritize reliability and predictable ROI. If integration risk outweighs labor savings, pilots may not convert to scaled contracts.
Finally, there is timing risk. If the broader robotics ecosystem matures more slowly than expected, capital may outpace revenue for longer than public markets ultimately tolerate.
However, it’s imperative to note that not all of these risks permanently impair value. Some delay outcomes. Others, particularly failure to achieve scalable generalization, would undermine the core thesis.
⏳Timing, Portfolio Fit, and What Would Change Our View
Physical Intelligence sits at the intersection of two key trends: rapid advances in large-scale AI models and renewed capital allocation toward robotics. The question is whether those trends meaningfully reinforce each other.
If recent AI scaling techniques translate into physical control systems, the intelligence layer in robotics could become a high-leverage position in the value chain. That would support platform-level economics and long-duration upside.
However, if translation proves harder than anticipated, progress may remain incremental and capital intensity may remain elevated. In that case, value accrues more slowly and margins compress.
From a portfolio perspective, this resembles a high-variance, asymmetric bet. The downside is prolonged capital deployment without durable defensibility. The upside is control of a reusable intelligence layer across multiple robotics markets.
What would change our view:
Evidence that deployments require heavy customization per environment.
Slower-than-expected improvement in model generalization.
Major incumbents demonstrating equivalent capabilities with broader distribution.
Clear signs that value is accruing primarily at the hardware or integrator layer rather than the intelligence layer.
The thesis strengthens if integration cost declines over time and performance compounds across tasks, but weakens if progress remains linear and deployment-specific.
AI in HR? It’s happening now.
Deel's free 2026 trends report cuts through all the hype and lays out what HR teams can really expect in 2026. You’ll learn about the shifts happening now, the skill gaps you can't ignore, and resilience strategies that aren't just buzzwords. Plus you’ll get a practical toolkit that helps you implement it all without another costly and time-consuming transformation project.
At Dhow, we back builders who chart new waters. Adnan Esmail and Physical Intelligence are trying to turn robotics into reusable intelligence by building a model that can see, reason, and act across different tasks and hardware platforms without requiring task-specific programming. If they succeed, integration friction declines, deployments become less brittle, and the same intelligence can be reused across environments. In a multi-trillion-dollar automation market, that leverage would sit at the control layer, where infrastructure standardizes and value compounds over time. Join the movement, share this with a friend (or two).
Want to get the most out of ChatGPT?
ChatGPT is a superpower if you know how to use it correctly.
Discover how HubSpot's guide to AI can elevate both your productivity and creativity to get more things done.
Learn to automate tasks, enhance decision-making, and foster innovation with the power of AI.
If you enjoyed this, you’d be doing us a HUGE FAVOR by just clicking the ads above - we appreciate it!
Sources
TechCrunch, “Physical Intelligence: Stripe veteran Lachy Groom’s latest bet is building Silicon Valley’s buzziest robot brains”, January 30, 2026.
The Robot Report, coverage of Physical Intelligence funding rounds:
Series A coverage (2024): https://www.therobotreport.com/physical-intelligence-raises-400m-for-foundation-models-for-robotics/
Series B coverage (2025): https://www.therobotreport.com/physical-intelligence-raises-600m-advance-robot-foundation-models/
Physical Intelligence Official Website
Adnan Esmail - LinkedIn Profile
General robotics and AI model scaling context drawn from public reporting and industry research as of February 2026.
Anduril Industries, “The Heartful Engineer: Adnan Esmail”
Karol Hausman - Personal Website / Academic Profile
Sergey Levine - UC Berkeley Faculty Profile
Chelsea Finn - Stanford Faculty Profile
Wired, coverage of Physical Intelligence and Lachy Groom