The assembly line at BMW's Spartanburg plant hums along as it has for years, but something has changed. Those robotic arms aren't just welding and lifting anymore. They're learning. Every motion, every adjustment, every moment when a human operator steps in to demonstrate a tricky maneuver—all of it gets captured, cataloged, and potentially sold.
Welcome to the peculiar economics of Physical AI, where the real money isn't in the robots themselves. It's in the data they generate.
This wasn't the plan, exactly. When companies first deployed humanoid robots and automated systems, they were chasing efficiency gains and labor cost reductions. But as the technology matured, a secondary market emerged—one that some observers say could dwarf the original automation business. The logic is almost embarrassingly simple: robots need to learn how the physical world actually works, and the best teacher is messy, unglamorous reality. Factory floors. Warehouses. Construction sites. Anywhere humans perform complex physical tasks that can't be replicated in a pristine research lab.
Which brings us to Praxis AI, a startup that graduated from Y Combinator in the summer of 2026 with what might be the most straightforward value proposition in recent memory: "Turns every company into a data vendor." The three-person team—Rohan Seelamsetty, Dev Karpe, and Tommy Li—is building the infrastructure to capture, package, and sell the kind of ground-truth data that robotics companies are scrambling to acquire. According to the company's own materials, they've already embedded their systems across 150+ environment types, capturing egocentric video and multimodal data spanning everything from publicly traded industrial giants to fast-growing logistics operators.
Whether those numbers hold up to scrutiny remains to be seen. But the thesis behind them reflects something real.
Compute Isn't the Problem Anymore
For the better part of a decade, the AI industry obsessed over compute. More GPUs. Bigger clusters. Whoever could afford the longest training runs would win. That made a certain kind of sense when the challenge was getting language models to sound coherent or image generators to produce something other than nightmare fuel.
Physical AI is different.
A robot operating in a warehouse doesn't just need raw processing power. It needs to understand how a human reaches for an oddly shaped box, how to navigate around a pallet that wasn't there yesterday, how to recover when a conveyor belt jams. That knowledge doesn't come from scaling up a data center in Virginia. It comes from watching thousands of hours of real people doing real work in environments that resist standardization.
In their January 2026 analysis, Andreessen Horowitz flagged this shift, arguing that data—not compute—would determine the winners in Physical AI. The research community had already started connecting the dots. NVIDIA released EgoScale in February 2026, a dataset containing over 20,000 hours of action-labeled video captured from a first-person perspective. The accompanying paper demonstrated what researchers call a log-linear scaling law: feed the model more human demonstration data, and validation loss for dexterous manipulation tasks drops in a predictable, steady curve.
By May, another research team published HumanEgo, showing something even more striking. They achieved zero-shot transfer from human video to robot control without using any robot data at all. Let that sink in. Show the AI enough footage of humans performing tasks, and it can figure out how to translate those movements to a completely different physical form.
A comprehensive survey published in July—titled "Data Pyramid for Embodied Manipulation"—laid out the emerging hierarchy and identified the gaps. Large-scale tactile datasets? Missing. Scalable collection pipelines? Still largely theoretical.
The message was blunt: if you want robots that can handle the long tail of real-world tasks, you need data from real-world environments. Preferably with humans in the loop, demonstrating edge cases and recovery strategies that no engineer would think to program explicitly.
And who has that data? Not the AI labs. The companies actually running factories and warehouses.
When BMW Became a Data Vendor (Whether It Knew It or Not)
Consider BMW's partnership with Figure AI. In June 2026, the automaker announced deployment of the Figure 03 humanoid robot at its Spartanburg facility. The previous generation had already contributed to production of more than 30,000 X3 vehicles over a ten-month stretch. Toyota Manufacturing Canada struck a similar deal with Agility Robotics on February 19, 2026.
On the surface, these look like traditional automation deployments. Dig deeper and the economics get more interesting.
Every robot rollout requires extensive demonstration data upfront. Human operators show the system how to perform tasks. Engineers capture edge cases and failure modes. The physical environment itself—lighting conditions, part variations, spatial layouts that change shift by shift—becomes part of the training corpus. As NVIDIA's Cosmos platform (announced in stages through May and June 2026) made clear, world models for Physical AI don't just need volume. They need diversity. Different factories. Different workflows. Different failure modes.
But most industrial companies lack the infrastructure to capture this data systematically, let alone monetize it. They're optimized for making cars or fulfilling orders, not for running what amounts to a data collection operation with rigorous provenance tracking and licensing frameworks.
That's the gap Praxis is targeting.
Their flagship product, marketed as "SilkRoute," ostensibly provides real-time guidance and verification for physical work in warehouse settings. The company claims it reduces order processing time from a 30-45 minute baseline to under five minutes, with roughly 40 percent fewer fulfillment errors. Those are vendor-provided metrics, and the company has not disclosed the methodology behind these results, so apply the appropriate grain of salt.
But efficiency gains may not be the real product. SilkRoute captures time- and location-stamped evidence of every action. That data—cleaned, annotated, and packaged—can be licensed to robotics companies and foundation model developers who need exactly this kind of material but can't easily generate it themselves.
It's a clever inversion of the traditional model. Instead of robotics companies deploying expensive data collection operations in environments they don't control, they buy access to streams from companies already doing the work. And those companies get a new revenue line from an activity they were performing anyway.
Whether it actually works at scale is another question entirely.
The Picks-and-Shovels Crowd Arrives
Praxis isn't operating in a vacuum. An entire ecosystem has materialized around Physical AI data, and some of the players bring serious infrastructure advantages.
Scale AI—the annotation powerhouse that's become something of a default vendor for large language model training—launched a dedicated Physical AI offering in 2026. They're building what they call "robotics data factories" and distributed collection networks. In March, Scale partnered with Universal Robots on the "UR AI Trainer," an imitation learning system designed to capture systematic data as robots transition from lab prototypes to factory deployment.
Encord rolled out infrastructure specifically for embodied, egocentric, and sensor data. Defined.ai expanded its Physical AI portfolio to include LiDAR, radar, force-torque, and teleoperation data, announcing the expansion in January 2026.
Open-source efforts popped up too. AXIS, released in July, provides a community-driven robot data engine with browser-based teleoperation and automated quality checks. MobileEgo Anywhere offers an open mobile app that converts egocentric video into standard formats used by robotics researchers.
The market itself, if you believe the secondary research firms, is heading somewhere between substantial and explosive. Grand View Research estimated the AI training dataset market at roughly $3.9 billion in 2026, projecting growth to $16.3 billion by 2033. Fortune Business Insights pegged it at $4.44 billion in 2026, climbing to $23.18 billion by 2034. These are analyst projections, not hard revenue figures, but the directional trend is consistent: double-digit annual growth as AI developers pay premiums for high-quality, domain-specific data with clear provenance.
Meanwhile, Physical AI deployments themselves are accelerating. Agility Robotics disclosed a SPAC path in June and opened a Fremont production facility to scale manufacturing. UK-based Humanoid raised $152 million at a $1.35 billion valuation that July. Figure AI closed a $675 million Series B at a $2.6 billion valuation in February. Google announced Gemini Robotics 2 in July, bringing what they described as enhanced dexterity and control to their AI-first robot stack.
Each of these deployments creates data. Each company that figures out how to license that data back to the ecosystem potentially opens a new revenue stream.
Perhaps more importantly, they create competitive pressure. If your competitor is monetizing operational data and you're not, you're effectively leaving money on the table.
The Privacy Problem Nobody's Solved
Turning workplaces into data sources raises some uncomfortable questions. Start with privacy.
Meta's Ray-Ban AI glasses—which Meta claimed had millions of daily users by mid-2026—triggered a wave of privacy concerns that the industry still hasn't fully addressed. In March, Ars Technica reported workers reviewing footage that had been captured in bathrooms. The U.S. Air Force banned the glasses over operational security concerns that February. These aren't edge cases or theoretical risks. They're glimpses of what happens when egocentric capture becomes ubiquitous without clear boundaries.
Regulation is starting to catch up, though whether it's catching up fast enough is debatable. The EU AI Act's transparency obligations took effect in August 2026. California's CPRA workforce data risk assessments became required at the start of that year. Illinois BIPA—the biometric privacy law that's generated more litigation than perhaps any other state statute—saw amendments and a Seventh Circuit ruling in April that reduced per-scan liability exposure but maintained strict consent requirements.
Data vendors operating across multiple jurisdictions have to navigate GDPR lawful bases, data protection impact assessments, and evolving AI Act requirements for systems deployed in EU workplaces. That's a lot of compliance overhead for a startup that's supposed to be moving fast.
Then there's the question of who actually benefits from this data economy. Hivemapper, a decentralized mapping network, updated its contributor terms in May 2026 to license data directly from individuals to developers. By July, community forums were filling up with complaints about reduced payouts. The lesson, perhaps, is that data vendor models need sustainable economics for the people actually generating the data—whether they're gig workers, employees, or automated sensors embedded in equipment.
If the value flows overwhelmingly to the platform and the buyers, with little trickling down to the people being recorded, you're building a system with unstable foundations. Workers notice when their movements become a monetized product.
Intellectual property presents yet another wrinkle. A February 2026 survey on 3D Gaussian Splatting IP protection outlined legal and technical gaps in protecting 3D scene assets. Enterprise terms from providers like Luma AI, updated in April, can grant the provider training rights on customer data unless explicitly negotiated otherwise. Companies contributing data need clear licensing frameworks, or they risk inadvertently surrendering their competitive moat.
Betting on a Future That May or May Not Arrive
The forecasts for Physical AI paint a remarkable picture—if you squint hard enough to ignore the uncertainty bands.
Goldman Sachs projected the humanoid robot market at $38 billion by 2035 in a February 2024 report, with a "blue-sky" scenario reaching $154 billion. Bank of America exhibits from early 2026 projected annual humanoid shipments hitting 1.2 million by 2030 and 10 million by 2035.
Those are enormous numbers. They're also, it bears noting, projections based on assumptions about technology maturation, cost curves, and market adoption that haven't been validated yet. The autonomous vehicle industry spent the better part of a decade chasing similarly optimistic forecasts before reality imposed some discipline.
Still, if even a fraction of those numbers materialize, the appetite for training data will be substantial.
Praxis is making that bet. When the founders brought the company out of stealth in July 2026, they positioned it as capturing "the physical-world data humanoid robots and frontier models learn from." They're seeking buyers—companies training robots and world models—and environment partners looking to monetize their operations.
The model is unproven at scale. The unit economics aren't public. The competitive landscape includes better-funded players with existing customer relationships. Open-source projects may commoditize pieces of the stack. Regulation could constrain what's permissible to capture and sell.
But the broader shift feels inevitable, or at least highly probable. As world models like NVIDIA's Cosmos—released as open weights in June 2026—become standard infrastructure, the bottleneck shifts to the training data. Companies that control access to real-world environments, sensor streams, and human demonstrations will be sitting on something valuable.
Whether Praxis becomes the dominant aggregator or just one node in a fragmented ecosystem remains an open question. Scale AI brings incumbent advantages and capital reserves. Vertical-specific players might capture niche segments. The data itself might become commoditized faster than anyone expects, driving margins toward zero.
What seems less debatable is the underlying dynamic. Factories really are becoming data generation sites, whether their owners planned it that way or not. The strategic question is who captures that value and how it gets distributed.
If history is any guide, the workers generating the data probably won't see much of it. But that's a different problem, and one the industry seems content to defer for now.
