Luel and the End of Free Data: Building the Marketplace for Rights-Cleared AI Training
For a decade, the fuel of AI progress was effectively free.
The recipe was simple: scrape the public internet, filter it, and train on it. Every webpage, every image caption, every code repository, every forum post became raw material for the models that now write our emails and answer our questions. The web was a vast, pre-existing corpus that nobody had to pay to create. Scale the crawl, scale the model, watch the capabilities emerge.
That era is ending, for two reasons at once.
First, the labs have essentially read the whole internet. The highest-quality public text has been consumed. What's left is lower quality, duplicative, or increasingly polluted with AI-generated content that degrades the next model trained on it. The free corpus is tapped out.
Second, the frontier has moved somewhere the web was never going to help. The models that matter now aren't just text predictors. They're voice agents that need to understand accented, emotional, real-world speech. They're robots that need to learn how a human hand grips a wrench. They're video models that need to understand how the physical world actually moves. None of that lives in a webpage.
This is the problem Luel is built to solve: sourcing the massive, net-new, human-generated data that the next generation of AI needs—collected with consent, documented for rights, and delivered at scale.
The Data the Web Doesn't Have
It's worth being precise about what "running out of data" actually means, because it's not that there's no data left. It's that the data that matters most no longer exists to be scraped.
Consider what a frontier lab actually needs today. A medical AI company might need thousands of hours of natural doctor-patient conversations in German, recorded with consent, spanning real clinical vocabulary. A robotics team might need egocentric video of craftspeople and factory workers using their hands—footage captured from the wearer's point of view, with sensor streams and device pose aligned to every frame. A speech team might need recordings of dozens of low-resource languages that barely appear online at all.
This data has three properties that make the old scrape-and-train playbook useless:
It has to be created, not found. Nobody has already uploaded ten thousand hours of consented clinical dialogue. It has to be collected on purpose, to specification.
It's multimodal and physical. The valuable signal isn't just what's in the frame. For embodied AI, it's the physics layer underneath—sensor data, hand-object interaction, the pose of the device—that turns a raw video into something a robot can learn from.
It has to be legally clean. A model trained on data of murky provenance is a liability waiting to surface. Enterprises deploying AI into regulated markets can't afford training sets whose rights they can't document.
The old data-labeling industry wasn't built for any of this. It was built to draw bounding boxes around cars in photos that already existed. Generating net-new, consented, multimodal human data at frontier scale is a fundamentally different operation.
What Luel Actually Does
Luel operates as a neutral, two-sided marketplace for rights-cleared multimodal training data.
On one side are the buyers: generative AI labs, robotics companies, and speech teams that write a specification for exactly the data they need. On the other side is supply—a global contributor network of hundreds of thousands of people across dozens of countries who can be mobilized to capture that data. Luel sits in the middle, matching the spec to the right contributors, running every submission through a multi-stage quality-assurance pipeline, and delivering audit-ready datasets.
The workflow compresses something that used to take months into a matter of weeks:
Specify. A lab describes the dataset it needs—modality, language, scenario, volume, quality bar.
Collect. Luel activates the relevant slice of its contributor network to capture the data to spec, in the geographies and languages required.
Clear and QA. Every submission carries consent and provenance records, and passes through quality control before it's accepted—so the buyer receives data that is both usable and legally defensible.
Deliver. The lab gets an audit-ready dataset, with the documentation trail intact.
Crucially, Luel builds along two axes. There's custom collection—bespoke datasets built to a buyer's exact specification. And there's a growing catalog of pre-cleared datasets, from patient-doctor conversations to manufacturing footage, that can be licensed off the shelf.
The Marketplace Economics
That two-axis structure is where the business gets interesting, and it's worth dwelling on.
A pure services business—we'll collect whatever you ask for—is valuable but linear. Every dollar of revenue requires a new project, new collection, new labor. It doesn't compound.
A marketplace with a re-licensable catalog behaves differently. When Luel builds a dataset for one buyer, the rights-cleared asset it produces can, where the terms allow, be re-licensed to others. The first customer funds the collection; every subsequent license is high-margin. Over time, the catalog becomes an appreciating asset—a proprietary library of exactly the net-new, consented, multimodal data the market is starving for, that grows more valuable and more defensible with every project.
This is the classic shape of a great marketplace: liquidity on both sides, a growing inventory that no competitor can quickly replicate, and unit economics that improve as the catalog deepens. The comparison the company's backers draw is deliberate—rights-cleared training data as the next great data marketplace, in the way that stock photography, market data, or app stores each became durable, high-margin businesses by aggregating supply the demand side couldn't assemble on its own.
The contributor network is the other half of the moat. Mobilizing hundreds of thousands of people across dozens of countries to reliably produce spec-compliant, consented data is an operational feat—part logistics, part incentive design, part quality control. It's not something a competitor spins up over a weekend, and it gets stronger with scale.
Why Provenance Becomes Non-Negotiable
There's a regulatory and legal current running underneath all of this that makes Luel's timing sharp.
The first wave of large models was trained under a "collect first, ask questions later" ethos. That's now generating exactly the questions you'd expect—lawsuits over copyrighted training material, regulatory scrutiny of data sourcing, and rising enterprise nervousness about deploying models whose training data can't be accounted for. Consent and provenance are shifting from nice-to-have to prerequisite, especially for anyone shipping AI into healthcare, finance, or other regulated domains.
In that environment, "we can document exactly where this data came from and that everyone in it consented" stops being a compliance checkbox and becomes a core feature of the product. A dataset with clean, auditable provenance is simply worth more than one without—because it's usable in situations where the alternative is legally radioactive.
Luel is building for the world regulators and courts are pushing the industry toward, not the one it's leaving behind. That's the difference between selling a commodity and selling an asset with a clean title.
The Team and the Bet
Luel came out of Y Combinator's Winter 2026 batch, founded by William Namgyal and Inigo Lenderking—repeat builders with backgrounds in machine learning and computer vision, and a working relationship that predates the company. The kind of operationally intense, network-driven business Luel is building rewards founders who move fast and sweat logistics, and the early trajectory reflects it: the company has drawn a substantial seed round with participation from leading venture firms and operators from across the AI industry.
But the more important bet is the thesis. Luel is wagering that the constraint on AI progress has moved. For years, the binding constraints were compute and algorithms—could you afford the GPUs, could you design the architecture. Those are still hard, but they're increasingly solved by capital and talent that the frontier labs already have. What they don't have, and can't simply buy off a shelf, is the net-new, consented, multimodal human data that the next capabilities require.
Whoever supplies that data at scale, with clean rights, sits at a chokepoint in the AI value chain.
The Picks-and-Shovels Layer
There's an old line about gold rushes: the people who got reliably rich weren't the miners, but the ones selling picks and shovels.
The AI boom has its share of miners—the labs racing to build the best models, an extraordinarily expensive and uncertain contest. Luel is selling shovels. It doesn't need to win the model race. It needs the race to keep happening, and it needs every serious contestant to require data that only a marketplace like it can supply.
As the free corpus of the public web gives out and the frontier moves into voice, video, and the physical world, the demand for created, consented, rights-cleared data goes up and to the right. The models will keep changing. The names at the top of the benchmarks will keep shuffling. But underneath all of it, the appetite for high-quality human data that can be deployed without legal risk only grows.
Luel is building the marketplace for exactly that. In an industry that spent a decade treating data as a free resource to be scraped, it's making a bet on the opposite: that the most valuable data is the kind you have to create on purpose, with permission—and that the company which organizes its supply becomes essential infrastructure for everyone building what comes next.