The demonstration data is messy. It's privacy-laden. And it's scattered across warehouses, assembly lines, and surgical suites in ways that make most data engineers wince. But this footage—of human hands manipulating objects, course-correcting when things slip, narrating the task as they go—has become essential fuel for the next generation of physical AI.
Which is where a company called Cepta enters the picture, claiming to have built the infrastructure to capture and clean demonstration data—with capabilities that suggest potential for commercialization.
The startup's Praxis platform, live as of 2026, positions itself as purpose-built data infrastructure for robotics and embodied AI systems. The underlying bet? That the real bottleneck in physical AI isn't some breakthrough in foundation models. It's earlier in the pipeline: getting the training data right in the first place.
It's an unglamorous thesis. But it might be the correct one.
The Grunt Work No One Wants to Talk About
At its core, Praxis does what most robotics teams would rather not: it turns operator demonstrations into training-ready datasets. The platform captures multimodal data—video, depth, inertial measurement units, haptics—through what Cepta calls "operator-side rigs." These are portable hardware kits designed to collect demonstrations in actual industrial environments, not just pristine lab settings with perfect lighting and controlled variables.
Then comes the structuring work. Praxis processes that raw capture into something usable: per-hand action sequences tagged with grasp points, contact timing, and natural-language descriptions of each task segment. And here's where it gets interesting—the platform keeps the failures. The slips. The retries. The moments when the operator realizes mid-motion that the angle is wrong and adjusts on the fly.
Those imperfections, it turns out, matter more than anyone initially expected for training robust models.
The system also bakes in compliance from the outset, which is less sexy but arguably more important for enterprise adoption. Praxis redacts personally identifiable information, captures consent per episode, and maintains what the company describes as a "provenance and audit trail from raw frame to exported episode." For companies in regulated industries—manufacturing, logistics, healthcare—that kind of audit capability isn't a nice-to-have.
Enterprises as Data Vendors, Maybe
Cepta's positioning hints at an emerging business model that could reshape how we think about industrial operations: what if companies that already perform complex manual tasks could capture those workflows, govern them properly, and license the resulting datasets to robotics developers?
The warehouse that's been packing boxes for twenty years suddenly has a new revenue stream. The surgical center training residents has a data product. The assembly line becomes, in effect, a training ground—not just for workers, but for the robots that might eventually assist or replace them.
Praxis theoretically enables that motion. Export functionality, governance controls, privacy protections—these create the prerequisites for data commercialization. Though the specifics of marketplace mechanics and revenue-sharing models remain undisclosed, and the company is apparently working with only a handful of design partners. No customers named. No pricing disclosed.
Which raises the obvious question: is this vaporware dressed up as infrastructure, or simply a company moving cautiously through early partnerships?
The Infrastructure Land Grab
Whatever Cepta's current state, it's entering a suddenly crowded field. Encord announced a significant Series C earlier this year—$60 million, bringing total funding to $110 million—for AI-native data infrastructure targeting physical AI. Foxglove launched what it calls a unified data search and curation platform for robotics data. Physicl emerged from stealth at NVIDIA's GTC conference with physics-grounded 3D data infrastructure.
Then there's a cluster of smaller providers—Seed Physical, FieldMesh Data, EgoData, and others—all advertising egocentric data collection services. Even large integrators like Deloitte and Cognizant have launched enterprise offerings in recent months, with heavy emphasis on data governance and provenance in regulated environments.
The data layer beneath physical AI is becoming its own distinct market, with all the land-grab dynamics that implies.
A recent Forbes piece framed the emerging bottleneck this way: "Physical AI hits a data labeling wall that only cash can fix." The analysis outlined competing infrastructure strategies and underscored that collection-to-curation pipelines are becoming both a competitive requirement and a potential spending sink—the kind that can quietly drain budgets while producing little immediate value.
What We Don't Know
Here's the part that's harder to parse: Cepta's launch has been unusually quiet for a company claiming to solve a major infrastructure problem. No formal press release. No named team members on the website. No disclosed funding. References to a possible Y Combinator affiliation, but nothing verified. The company's website carries recent copyright dates and the product pages appear functional, but specifics remain scarce.
That opacity might be strategic—perhaps the company is deliberately working through design partnerships before a louder market push. Or it might simply be early-stage caution, the kind where founders are still figuring out if the thing actually works before making bold claims. Or, less charitably, it could suggest a product that's more pitch deck than platform.
Either way, Praxis is now ostensibly in the market with a pitch that enterprises have valuable training data sitting dormant in their operations. With the right infrastructure, they could govern it, structure it, and potentially sell it.
The Training Data Has to Come From Somewhere
Whether this thesis unlocks a new data economy for physical AI—or simply adds another expensive layer to an already complex stack—remains an open question. The robot arms perfecting your Amazon package did have to learn from somewhere, after all. And someone wearing a head-mounted camera did show them how.
The question is whether companies like Cepta can build the infrastructure that turns those demonstrations into a sustainable business model, or whether this becomes yet another venture-backed experiment in selling picks and shovels to a gold rush that may or may not materialize.
For now, the platform is reportedly live. The robots still need their training data. And the messy, privacy-laden, scattered footage of human operators doing their jobs has quietly become one of the most valuable datasets no one's talking about.
Yet.
