Frontier AI labs need rights-cleared multimodal training data at scale, and the most accessible source — the public web — has been substantially exhausted for the modalities that matter most.^[2] Most available datasets fail production requirements: unclear rights, weak provenance, missing consent, inconsistent metadata.^[1] The shortcuts that worked five years ago no longer work.

At the same time, the cost of using shortcuts has gone up. The NYT v. OpenAI suit, the YouTube transcript controversy, and the EU AI Act have collectively pushed legal, brand, and procurement teams to demand explicit licensing and audit-ready documentation before a dataset can enter pre-training.^{[3] [4] [8]}

The remaining option is to commission real-world data: egocentric video from licensed contributors, professional dialogues with consent, robotics teleoperation footage shot to spec. That is operationally hard. It involves recruiting and vetting contributors, capturing consent, running multi-stage QA, and shipping artifacts the procurement team can sign on. Most labs don't want to build the operation in-house — and the vendors that built it at scale have new problems of their own. After Meta took 49% of Scale AI in June 2025, neutrality itself became a procurement question: OpenAI wound down its Scale work and Google moved to cut ties, sending labs shopping for independent partners.^{[24] [25] [26]}

Data exhaustion is no longer hypothetical. Epoch AI's projections put high-quality public text data on a finite curve, with constraints intensifying as training compute scales.^[2] Multimodal demand makes the gap larger, not smaller — real-world video, audio, and robotics footage was never on the web in usable quantities to begin with.

Lawsuits and licensing deals reshaped the buyer's risk model. The NYT v. OpenAI litigation, OpenAI's deal with Stack Overflow, Google's licensing arrangement with Reddit, and Axel Springer's partnership with OpenAI together signal that the era of "scrape now, apologize later" is closing.^{[3] [17] [18] [19]} Buyers now want a paper trail.

The EU AI Act's GPAI obligations are now in force. Final approval in May 2024 codified provenance and consent expectations; since August 2, 2025, GPAI providers must publish training-data summaries on a mandatory Commission template and maintain a copyright-compliance policy.^{[8] [33]} Provenance paperwork moved from best practice to legal requirement — and it effectively exports globally for any lab serving European customers.

Neutrality became a buying criterion. Meta's $14.3B stake in Scale (June 2025) triggered an immediate customer exodus: OpenAI phased out its Scale work, Google planned to walk from roughly $200M of 2025 spend, and Scale's competitors reported an influx of labs seeking neutral partners.^{[24] [25] [26]} Independent vendors are catching budgets that were locked up a year ago.

Multimodal and robotics need bespoke collection. State-of-the-art robotics systems like RT-1 rely on hand-collected teleoperation datasets, not scrapes.^[5] RLHF practice has converged on the conclusion that quality, diverse human data — not sheer volume — is the bottleneck.^[6]

Bespoke collection. The high-margin, high-touch motion: a frontier lab needs gemstone manufacturing footage, or patient-doctor dialogues, or egocentric data from cooks in a commercial kitchen. Luel turns the request around in days with the full paper trail. Customers pay for speed and for the procurement-ready artifacts.^[1]

Off-the-shelf catalog. The compounding motion: collections that were funded by a bespoke job become re-licensable inventory. The customer gets a fast start; Luel gets a margin profile closer to stock-media than to services. Inigo's published Ego-Realm dataset on Hugging Face is an early demonstration of the catalog motion.

Compliance is the product, not a wrapper. Standard license templates, usage scopes, consent revocation flows, and documentation aligned to EU AI Act risk categories are not features bolted on — they are why the buyer chooses Luel.^[8]

Interoperability. APIs, SDKs, and delivery formats are designed to drop into the lab's existing pre-training, fine-tuning, and evals pipelines so the dataset doesn't sit in a slow review queue.

In May 2026, Luel announced a $31.2M seed co-led by General Catalyst and Lightspeed — one of the largest seed rounds in Y Combinator's history.^{[22] [23]} Additional backers include Paul Graham, SV Angel, Human Capital, and Orange Collective.^[22]

The traction preceded the capital: $2M ARR within roughly six weeks of demo day, over a million submissions processed through the QA pipeline, and 40+ dataset campaigns running concurrently.^[22] The customer base already spans generative AI labs, robotics companies, speech research teams, major social platforms, universities, hospitals, and banks — broader than the frontier-lab-only wedge we expected at memo time.^{[22] [23]}

Lightspeed's stated rationale is the "data wall": models exhausting public web data and requiring massive net-new, human-generated, rights-cleared data across modalities and geographies.^[23] That is the same thesis as this memo — now underwritten at institutional size.

The AI training datasets market sits at roughly $2.82B in 2024 and is projected to reach ~$9.58B by 2029 at ~27.7% CAGR, with multimodal as the fastest-growing segment.^[9]

The broader data collection and labelling services market is roughly $3.0B in 2023 heading to ~$29.2B by 2032 at ~28.5% CAGR.^[10] Publisher and platform licensing deals validate willingness to pay at the upper end of that range: Google pays Reddit a reported $60M a year, OpenAI's Reddit deal is estimated at ~$70M a year, and News Corp's OpenAI agreement is worth up to $250M over five years.^{[30] [32] [31]}

Two structural tailwinds compound on top of the headline numbers. First, modern robotics and multimodal systems require real-world egocentric and device-specific data that scraping cannot provide.^[5] Second, public-data exhaustion drives a premium on curated, re-licensable corpora — and pushes more of the spend toward bespoke collection rather than pre-existing dumps.^[2]