The past few months of robotics demos have taught us a lot. Robots can do more than ever, from driving cars to folding laundry to threading data center cables. But as capabilities compound, the gap between a compelling demo and a viable business remains stubbornly large. We believe robots will unlock enormous economic value for society, but the world's smartest physical AI researchers still hold conflicting views on how we’ll get there.
One camp believes that embodied foundation models are headed for a ChatGPT moment: absorb enough data, cross a capability threshold, and outperform every purpose-built system. It's what frontier LLMs did to the digital world. The other camp believes the physical world is too varied for that playbook, and that deployment-first companies will own their categories before any foundation model catches up. And a third camp is trying to do both simultaneously with a full-stack approach. Billions of dollars have been raised across all three bets.
We believe all three can be successful, but on different timelines.
The Stakes
The physical AI market is 7x the digital AI market.
Digital AI automates cognitive work: coding, customer support, legal, etc. Physical AI goes after something far bigger: the work that happens with hands and machines: manufacturing, logistics, construction, data centers, and infrastructure maintenance.
Estimates from NVIDIA and Salesforce put the opportunity for physical AI at roughly $40 trillion, about seven times the $6 trillion opportunity in digital AI.
In many ways, digital AI could just be the tip of the iceberg as machine intelligence expands into every corner of the economy. The winners in this race will unlock the greatest technological transition in modern history.
Physical is harder than digital
Language models received an extraordinary gift: the internet.
GPT-3 trained on decades of human writing that was readily available at near-zero marginal cost. Will Depue, formerly of OpenAI, called it a "one-time subsidy." That subsidy is what made the ChatGPT moment possible: a model good enough at everything, trained on data nobody had to contrive. And an API that put it in front of hundreds of millions of people almost instantly. The ChatGPT moment was built on free data and (nearly) free distribution. Robotics has neither advantage.
There is no data subsidy in robotics. No archive of physical demonstration data sits waiting to be scraped. According to Science Robotics, the field is almost five orders of magnitude behind modern vision-language models. And the data that does exist pools poorly. Text is text, wherever it came from. Physical data is bound to the robot that produced it: its morphology, its sensors, its environment. A dataset collected on one arm in one warehouse transfers imperfectly to a different arm in a different warehouse.
The distribution challenge compounds the data problem.
Software ships at zero marginal cost to anyone. A robot has to be manufactured, which often takes months. Then it has to be shipped, installed, and integrated into a customer's workflow. When something breaks, someone drives to the site. Every deployment requires hours of human labor and deep customer relationships. Not to mention the complexities around hardware system integrations, reliable power and internet, and broader capex challenges. The feedback loop that took digital AI companies weeks could take robotics companies years, and unfortunately there is no API shortcut around it.
The consequence is that there is likely no singular ChatGPT moment coming for robotics, and that is not a bearish take. Progress is real and accelerating; it will just arrive task by task rather than all at once. In digital AI, one overnight moment concentrated value in a handful of labs. In physical AI, a $40 trillion transition will be won deployment by deployment. The window belongs to builders.
The physical AI stack
While the industry is still early, we broadly see three emerging approaches between the underlying data layer and the end customer:
- Foundation Models
- Deployment-First Companies
- Full-Stack Companies
Each is attacking the same fundamental problem from a different direction: how do you create the data flywheel required to make physical intelligence reliable enough for the real world?
The model-first approach
Foundation model companies believe one model will generalize and win it all. Physical Intelligence (PI) and Generalist have raised billions on this logic, running OpenAI's GPT-3 playbook: co-train on web-scale, vision-language data and deploy early to accelerate research. These models have taken longer than expected to demonstrate production results. The most recent benchmarks are sobering: in a recent real-robot evaluation, the best model averaged under 50% task success, and on new benchmarks covering long-horizon manipulation, state-of-the-art models succeed just 13% of the time on tasks where human teleoperators score 100%.
At the same time, the generalization these labs are betting on is starting to appear. Generalist's newly released GEN-1.5 is the best demonstration yet: it is a one-shot learner. Show it a single 3–12 second demonstration of a new task, and it attempts the task immediately, with no retraining. This is a new capability that emerged from pretraining on physical data at scale. One-shot success on simple tasks still sits around 59%, while few-shot learning is rising to 83%. Generalization is arriving, but the last mile to production reliability still runs through the real world.
*Experts disagree on what data to collect and on what architecture to build to close the performance gap.
*
On data, every lab hedges across every strategy, but most started in one modality. Physical Intelligence began with teleoperation, the highest-fidelity data at exorbitant cost. Generalist began with handheld tools like UMI grippers, an approach that is more data-efficient but hard to scale beyond narrow tasks. Everyone uses egocentric video, abundant and cheap but missing the embodied details a robot needs. Simulation scales infinitely but remains stuck crossing from sim to real. No consensus has formed on which mix gets to production quality.
On architecture, VLAs, or vision-language-action models, are the commercial workhorse today: they translate visual observations and language instructions directly into motor actions, typically after post-training on demonstration data. World action models (WAMs) instead pre-train on video to learn how the physical world behaves, betting that simulating the world dramatically reduces the robot-specific data needed to climb the scale curve. In reality, most models today are a mix of both.
Beneath these open questions sits an unfortunate truth: the gap between 95% and 99.9% task success is the gap between a demo and a business. The only agreed-upon way to reach 99.9% is robots doing real work with real customers. That is the opening for deployment-first companies, and why, for now, the two camps' successes are inextricably linked.
The deployment-first approach
Deployment-first companies start with the customer. Their thesis is straightforward: real deployments generate the highest-quality data and compel a company to build what customers will pay for.
Waymo is the best early example of the deployment-first approach.
Waymo collected ~300 million lidar-driven miles and proved that a data flywheel built on real-world miles in a constrained environment is one successful path to autonomy. It also paid full price: predating LLMs and commodity hardware, Waymo built every layer of the stack from scratch, over 15 years and billions of dollars.
Companies starting today inherit a much stronger foundation. Our investment in Atoms reflects this belief. Atoms is bringing commercial applications across food, mining, and transport in a fraction of the time and cost by combining advances in AI with an intense focus on real-world data collection, operations, and manufacturing.
Research partnerships are forming between deployment companies and foundation models to catalyze the data flywheel.
For example, Ultra Robotics deploys robots for pick and pack in 3PL warehouses. Ultra partners with PI for intelligence: it installs custom hardware at customer sites, teleoperates the robots, and sends the resulting data back to PI. PI fine-tunes on that data and returns a model checkpoint Ultra can run autonomously on simpler tasks. Each cycle automates more of the work, while PI's base model improves on real-world data it could not have collected alone. Foundry Robotics has a similar research relationship with Generalist, paired with a more aggressive business model: rather than charging per robot deployed, Foundry operates as a robotic contract manufacturer and charges for finished output. It takes on more cost and operational ownership, and in exchange captures far more of the value of every factory it runs.
Deployment becomes its own form of defensibility.
Over time, some deployment-first companies may adopt open models. Others may accumulate enough proprietary data to train models themselves. Either way, the customer, the integration, and the data stay with whoever deployed.
The full-stack approach
The third camp believes model development and deployment cannot be separated. These companies want to own the entire loop: train the model, build the hardware, deploy the robot, collect real-world data, improve the model, and repeat. Many begin with a vertical or use case rather than attempting to solve general-purpose robotics immediately.
Advantages lie in access to a real operating environment and large-scale, relevant data from day one.
Rhoda Robotics is building a video-predictive world action model, pretrained on internet-scale video and a controlled closed-loop, with the stated philosophy that robots must prove themselves in factories and warehouses, not labs. Sunday Robotics, the creators of the UMI grippers, are bringing the full stack into the home, starting with laundry and dishes. Mind Robotics and Walden Robotics have kickstarted their data flywheel by signing up captive and deeply aligned customers, Rivian and Toyota, respectively. This creates an obvious advantage: access to a real operating environment and large-scale dataset from day one. It also creates risk, namely over-indexing on a single customer's problem set and struggling to generalize beyond them.
Model companies are moving into deployments and becoming full-stack. For example, Skild AI, initially funded as a pure foundation model lab, ultimately acquired Zebra's robotics division to own warehouse deployments end-to-end. Its models now run on Foxconn assembly lines building Nvidia server racks. When one of the best-funded model labs in the field starts buying deployment businesses, it is a good indication of where the data and the revenue actually live.
Where we see opportunity
High-ROI use cases
Regardless of the approach, we are most excited by use cases with high ROI and highly repetitive tasks. A robotics company incurs costs a software company never had to absorb. The top robotics companies are targeting a longer-term bill of materials (“BOM”) of $15k - $50k per robot (including on-device compute). However, when you factor in shipping and installation, local operations teams, hardware repair costs, and the likely continued cost of teleoperation support to maintain production quality, we expect the costs to be more than 2x the BOM. The task has to be valuable enough to cover those costs and still leave attractive margins, yet repetitive enough that a single robot can run it thousands of times.
We believe the most attractive opportunities initially will be where companies can charge $100k + per robot per year, which is the pricing foundation a hardware business needs to recoup at least 2x its own cost structure. Watney Robotics is a great example in the data center operations space, where the severe technician labor shortage and surging data center buildout demand have created a willingness to pay $500k + per robot per year. We see compelling opportunities in industries like data centers, manufacturing, construction, or warehousing and logistics - just to name a few.
The difficulty of physical businesses becomes part of their moat.
A software user can switch from Codex to Cursor to Claude Code on a whim. A customer that has installed a physical fleet and built business-critical processes around it cannot. That friction compounds alongside the deployment data, and we believe deployment-first companies will build deeper moats and bigger businesses than their digital counterparts, in markets measured in trillions rather than billions.
Different approaches are on different timelines.
For deployment-first and full-stack companies, the window is especially attractive today. General-purpose models cannot yet solve every physical task, giving these companies time to build customer relationships, accumulate proprietary deployment data, and embed deeply within specific workflows.
Over time, foundation models will generalize, moving up the complexity curve and unlocking use cases out of reach today. We expect the lines between the three approaches to blur: model companies will deploy, deployment companies will build more intelligence, and full-stack companies will expand beyond their initial use cases.
In a $40 trillion market, there is room for many approaches to win. If you are building in this space, we would love to meet you. And if you disagree with our perspective, even better. We would love to hear how you see this emerging world.
Thanks to my co-authors @loubohan and @nmfaulkner_
