The real bottleneck in AI has shifted from compute to data. Protege is building the licensed marketplace for the real-world kind.
Public web data built the first generation of large AI models. Then the well ran dry. The open internet, as Protege's founders like to point out, represents only a small fraction of the world's usable data - the rest sits behind the walls of hospitals, media archives, recording studios and motion-capture stages, locked up by privacy law and intellectual property.
Protege is the company trying to open those doors without breaking them. It runs a two-sided data platform: on one side, owners of private, proprietary data; on the other, the AI labs and enterprises that need real-world data to train, fine-tune and evaluate their models. Protege licenses that data, curates it into AI-ready datasets, and layers on the unglamorous but essential machinery - de-identification, provenance tracking, rights preservation and compliance - that lets both sides say yes.
The pitch is that this middle layer is where AI's next leap actually lives. Bigger models help, but a model is only as good as what it eats. By aggregating access to billions of data points across healthcare, video, audio and physical motion, Protege positions itself as core infrastructure for a phase of AI development that has moved past scraping the web.
It is a New York company, founded in 2024, and it has moved quickly - from a stealth launch to roughly $65 million raised and a customer roster it says includes a majority of the industry's largest AI labs.
Three things gate AI progress: compute, models and data. The first two have armies of companies attacking them. The third - responsible access to real-world data - had no obvious owner. That gap is Protege's entire reason for existing.
The public internet powered early progress, but modern models need real-world data from regulated environments that was never online to begin with.
Synthetic data can approximate patterns but can't fully replicate the complexity of the real world and actual human behavior.
The most valuable data - clinical, audiovisual, behavioral - is bound by consent, IP and compliance rules that make casual use impossible.
Protege organizes its offering around the AI lifecycle - from the first pre-training run to the final benchmark - and around the industries where real-world data is richest.
Massive, diverse real-world datasets across industries to train foundation models at scale.
Narrower datasets for supervised training and human feedback that align model behavior.
Curated, domain-specific datasets that adapt general models to specialized use cases.
Uncontaminated, real-world data for honest testing - including a benchmarking collaboration with Vals.ai.
An analytics workspace for exploring and working with Protege's catalog of data.
De-identification, consent tracking and provenance built into every dataset, not bolted on after.
Across four verticals, Protege has aggregated access to billions of data points. The scale is the story - and the moat.
FIG. 1 - RELATIVE CATALOG SCALE BY VERTICAL. FIGURES APPROXIMATE, PER COMPANY DISCLOSURES.
Plenty of companies label and annotate data. Fewer solve the harder problem: legally and ethically sourcing data that was never meant to leave its owner's hands.
Protege's differentiation is structural. Rather than scraping or generating data, it builds licensing relationships with the organizations that own it - and designs for privacy, governance and rights preservation from the start. That approach favors curated, use-case-specific datasets over raw bulk, and it lets data owners keep their rights while getting paid. Since its seed round, Protege says it has generated tens of millions of dollars in revenue for its data partners.
The team's pedigree reinforces the pitch. Co-founder Travis May previously co-founded LiveRamp and Datavant, both built on neutral data infrastructure in regulated markets. In that world, being the trusted, neutral party is the product.
Frontier AI labs and startups on the demand side - reportedly a majority of the industry's largest players - and 100+ hospitals, media libraries, audio owners and motion-capture studios on the supply side. Enterprise customers include Siemens Healthineers.
AI builders license curated datasets; data owners are compensated through structured agreements that preserve their IP. Protege earns by facilitating governed access and by selling value-added services - curation, de-identification, benchmarks and analytics.
Protege aims to be the central platform for licensed, real-world data in AI - a market where, it argues, no dominant player yet exists. It sits alongside labeling firms and synthetic-data vendors but competes on sourcing real, regulated data responsibly.
Alternatives range from AI labs sourcing data in-house to data-labeling companies like Scale AI and Surge AI, marketplaces such as Defined.ai and Appen, and synthetic-data providers. Protege's wager is that none of them own the trusted, licensed, real-world layer - and that this layer is where the next decade of AI value accrues.
Bobby Samuels and Travis May launch Protege to unlock private, real-world data for AI.
The company emerges from stealth with a CRV-led round and opens its data platform.
Adds Audio & Speech and Motion Capture; signs 100+ data partners.
Footwork leads to deepen the product and expand verticals.
Andreessen Horowitz leads, bringing total funding to ~$65M.
| Round | Amount | Date | Lead |
|---|---|---|---|
| Seed | $10M | Sep 2024 | CRV |
| Series A | $25M | Aug 2025 | Footwork |
| Series A1 | $30M | Jan 2026 | a16z |
| Total | ~$65M | 2024-26 | — |
Other backers across rounds: Bloomberg Beta, Flex Capital, Shaper Capital, Liquid 2 Ventures.
Career at the intersection of data, privacy and infrastructure, with time at Datavant and LiveRamp.
Founder & CEO of Shaper Capital; co-founder and former CEO of LiveRamp and Datavant.
Leads the scientific rigor behind Protege's datasets and methodology.
Owns the engineering and platform that powers Protege's data infrastructure.
Interviews and demos where the founders explain the thesis in their own words.