The TAM for physical AI is measured in trillions – McKinsey estimates the market for agents and robots could be $2.9T by 2030. Going after this market are companies trying to deploy intelligent robots, with the vision to displace and transform massive amounts of human labor.
However, today these companies are hitting the same bottleneck trying to get to the level of intelligence and real-world reliability necessary for these robots to be deployed at scale: robotics training data. To put the criticality of robotics data in perspective, one startup said, “When we were in stealth for the first year, we spent literally no time working on the robot – the entire aspect that we figured out was how to create a flywheel for data.”
There are significant lessons to be learned from the evolution of the LLM data market. Given the sheer diversity of tasks, embodiments, and environments physical intelligence has to cover relative to language, the robotics data market is likely to end up even larger.
On who wins and becomes generational here, my belief is that the winning motion is a land-and-expand through the transition from development / early research into the deployment age: landing through specific differentiators (labor access, diversity, agility) at the robotics-intelligence inflection point, then expanding into evals-as-a-service, simulation, and deployment support as demand shifts from training data into deployment.
Hot Demand Now
Unlike LLMs, robotics does not yet have a standard, accepted data recipe, so different labs demand materially different types of data and have different willingness to outsource collection.
Human data collection is hot because of recent evidence (HumanScale Paper, EgoScale Paper, Dyna-2 Model) suggesting that pre-training on large-scale human data can lead to leaps in model gains. Conviction is spreading among researchers that trading off harder-to-scale, embodiment-matched data (i.e. teleop, UMI), in favor of larger volumes of easier-to-scale, less embodiment-matched data (i.e. egocentric), can be a critical component of the recipe for robotics intelligence – this is occurring simultaneously with 15+ new startups competing in the market.
The strongest demand for third-party data right now is from labs pursuing broad, general purpose intelligence and needing large volumes of diverse training data. In contrast, robotics companies optimizing around specific customer deployments tend to view data collection as part of their proprietary feedback loop and keep significantly more of it in-house. For example, Sunday Robotics (focused on home robot deployments), Field AI (general deployments), Skild, and Persona AI (welding/industrial deployments) share this belief and do all their data collection in-house.
For those who are outsourcing data collection (Generalist, Dyna, OpenAI, Nvidia, 1X) third-party data vendors create value primarily by taking on the operationally-intense collection problem. Insiders describe this as an “operational shitshow” – sourcing labor, scaling up collection on tight timelines, managing task and environment diversity, enforcing quality requirements on the task and downstream annotations, and adapting (re-training labor, re-distributing hardware and collection methods) as labs change their data recipes.
Commoditization Risk and Differentiators
But buyers increasingly struggle to distinguish between vendors, particularly for egocentric data, which is pushing the category toward commoditization. The CEO of one general intelligence humanoid startup said his inbox is flooded with vendors pitching “100k hours” of ostensibly differentiated data, while much of it is difficult to distinguish and likely has very low trainable yield. A researcher at a large foundation model lab similarly expects egocentric data to commoditize because it increasingly looks the same from the buyer’s perspective.
From conversations with buyers, labs care less about raw volume alone and evaluate their supply options on three things: diversity, yield, and agility. Their big questions are: Can the vendor deliver underrepresented tasks and environments? Does a higher percentage of their data actually train well? And can the vendor quickly shift collection methods if our recipe changes?
Scale AI appears to have built an early advantage around the diversity and agility vectors. On diversity, an insider there cites that a core reason labs buy from them is their ability to execute on environment-diversity operations, like ensuring a distribution of lighting conditions in their data. On agility, Scale collects using its own hardware, modifies it to meet customer specs, and operates using hardware supplied by the lab itself; one of the leading foundation model labs, for example, sends its proprietary grippers to Scale for collection. On the importance of agility, a Scale robotics operator estimated that the current egocentric-data wave may have only roughly a year of shelf life and explicitly warned against vendors making large fixed investments into a single modality, e.g. buying 10,000 cameras, where demand may fluctuate or rapidly disappear as the winning recipe emerges.
Longer-Term Winners
Longer term – the period after robots are proven functional/valuable, extending to scaling and proliferating deployments across our entire human labor force – buyer needs will shift downstream: from “give me more training data” toward “help me get robots reliably deployed in new environments.” Some leading data providers are already beginning to orient their roadmaps in this direction, including evals-as-a-service and deployment support.
Hence, what the long-term winner likely needs are:
- Access customer environments and proprietary workflows that others cannot reach
- Become the deployment-partner to robot companies and labs, that enables them to enter, fine-tune, simulate/eval-as-a-service, and rollout into net-new real-world deployments
- Prove through in-house research that their data is of quality and improves frontier robot performance
- Build benchmarks that steer labs and robot-builders towards economically valuable tasks
What would let a vendor expand from near-term into long-term value, and become a generational company?
| Long-term demand | What does it take? |
|---|---|
| Higher bar of data quality | In-house research chops and benchmarking, to prove that data improves performance |
| Specialized collection networks or hardware | Access to domain-specific expert operators/demonstrators; ability to co-design hardware for specialized data collection |
| Unblock more/new real-world deployments | Access to exclusive end-customer deployments; operations to collect training data from new end-customer deployments; simulations/testing to de-risk the actual deployment of robots in those environments |
| Benchmarking / evaluation infrastructure | Standardized tests/environments and hold-out environments |
What I'm Looking For
Here's how I'm evaluating startups in this space:
- Egocentric or generic data vendors with differentiated access to labor and diversity of tasks, exceptional agility to adapt to different modalities, or provable yield. These factors are key for them to win and differentiate in the near term amid the commoditization of this data type.
- Look for companies building deployment-accelerating infrastructure or access to proprietary environments / deployments.
- Look for companies building a specialized network of demonstrators or co-designing hardware for specialized data collection.
- Look for tooling that improves yield for in-house, vertically integrated collectors. These approaches – such as robotics factories or consumer robotics – have a very different near-term demand for data and services; focus on companies that make it easier to extract yield from in-house data.