Deepen AI launches Refinery to sell and verify Physical AI data

Deepen AI has unveiled Deepen Refinery and opened a waitlist for data partners and buyers ahead of a closed beta later this month. The marketplace is designed to help Physical AI teams buy real-world data with clearer provenance, licensing and payout terms.

Why it matters: - Physical AI teams are spending heavily to collect data that already exists elsewhere, while quality, provenance and licensing often remain unclear. - Deepen Refinery is meant to turn that data into a tradable asset with visible terms, validation and payment back to the original producers. - The launch targets a market where robotics investment is surging but usable training data has not kept pace.

What happened: - Deepen AI unveiled Deepen Refinery and opened a waitlist for both data partners and buyers. - The company plans an invite-only closed beta later this month. - The waitlist is live at the Refinery waitlist, with separate queues for data partners and buyers. - Model developers can sign up for access next month. - Companies and research labs with unused Physical AI data can contact Deepen AI at info@deepen.ai about listing it before the beta opens.

The details: - Refinery gives buyers access to real-world data locked inside other companies and labs. - Buyers can preview samples before purchase. - Each listing shows what a license allows. - Buyers can see which datasets have been validated through Deepen's pipeline. - The catalog will include Deepen's own egocentric video, motion capture and vehicle data, plus datasets from data partners. - Each listing shows how the data was captured, including sensors, location and capture date. - Partner identities can stay anonymous during the beta. - Deepen Refinery does not build foundation models, so it is not competing with the teams it sells to. - Data partners decide who can buy their data and on what terms. - Partners can set their own license terms. - Partners can block specific companies, including direct competitors. - Partners can list data as-is or send it through Deepen's validation pipeline. - Partners are paid on every sale. - A partner dashboard tracks requests and payouts. - Deepen's own datasets go through the same pipeline and validation as every other listing. - Deepen gives its own datasets no preference.

Between the lines: - Robotics startups raised $18.8 billion by mid-June this year, according to Crunchbase, already more than the $15 billion raised in all of 2025. - A June study by researchers at NVIDIA, MIT and three other universities said even the most generous count of public real-robot data is only about 20,000 hours. - That is roughly what a single lab may hold privately. - In the Open X-Embodiment study led by Google DeepMind, a model trained on combined data from nine robot arms was about 50% more successful on average across five robots than models trained on each lab's own data. - CEO Mohammad Musa said the industry should treat physical-world data as a shared asset with ownership and payments flowing back to contributors. - Chief Product Officer Ali Peracha said Refinery addresses both sides of the market: buyers can find data before collecting it, and owners can sell it with control over who gets access.

What's next: - Deepen AI plans to admit the first participants when the closed beta opens later this month. - The company expects model developers to gain access next month. - Deepen AI is also building a broader data economy for Physical AI through sensor calibration, annotation and validation services.

The bottom line: - Deepen AI is trying to make Physical AI data easier to find, safer to buy and more valuable to sell.