Deepen AI Unveils Refinery to Build the Data Economy Physical AI Will Run On

The top Physical AI labs are buying data without guarantees on quality, provenance, or usability. Deepen Refinery solves this by providing a unified marketplace

SANTA CLARA, CA, UNITED STATES, September 21, 2026 /EINPresswire.com/ — Deepen AI today unveiled Deepen Refinery™ and opened its waitlist to data partners and buyers, ahead of an invite-only closed beta later this month. Refinery gives teams building Physical AI access to real-world data that’s locked inside other companies and labs today, along with the checks they need before training on it. Buyers preview samples, see what each license allows, and can tell which datasets have been validated through Deepen’s pipeline. The people who produced the data get paid on every sale.

Investment in Physical AI is at a record. Robotics startups raised $18.8 billion by mid-June of this year, according to Crunchbase, already more than the $15 billion raised in all of 2025. The data behind those companies hasn’t kept pace. Even the most generous count of all public real-robot data comes to only about 20,000 hours, according to a June study by researchers at NVIDIA, MIT, and three other universities. That’s roughly what a single lab holds privately.

That’s a waste, because pooled data trains better models. In the Open X-Embodiment study led by Google DeepMind, a model trained on combined data from nine different robot arms was about 50% more successful on average, across five robots, than the models each lab had trained on its own data. The gain came where each lab’s own dataset was small.

“The best robots of the next decade won’t be trained on one company’s data. They’ll be trained on everyone’s, and everyone who contributed should get paid for it,” said Mohammad Musa, CEO of Deepen AI. “Right now the industry treats physical-world data as a one-time project cost and keeps paying to collect it again. Refinery turns that data into an asset, with clear ownership, license terms buyers can see, and payment going back to whoever produced it.”

For buyers

When the beta opens, the catalog will include Deepen’s own egocentric video, motion capture, and vehicle data, alongside datasets from data partners. [Each listing shows how the data was captured, including sensors, location, and capture date, even when the partner behind it stays anonymous.] Deepen Refinery™ doesn’t build its own foundation models, so it isn’t competing with the teams it sells to.

For data partners

Data partners decide who gets their data and on what terms. They set their own license terms, can block specific companies from buying, including direct competitors, and stay anonymous to buyers during the beta. They can list data as-is or have it validated through Deepen’s pipeline. Partners get paid on every sale and can track [requests and payouts] in a partner dashboard. Deepen’s own datasets go through the same pipeline and validation as every other listing and get no preference.

“Across more than 200 Physical AI programs, we’ve watched teams pay to collect data that already exists on someone else’s drive,” said Ali Peracha, Chief Product Officer at Deepen AI. “The buyer can’t find it, and the owner won’t sell it without control over who gets it. Refinery fixes both. Buyers look before they collect, and owners decide who buys.”

Waitlist

The waitlist is open at https://refinery.deepen.ai, with separate queues for data partners and buyers. The first participants will be admitted when the closed beta opens later this month. Model developers can sign up now for access next month. Companies and research labs holding Physical AI data they aren’t using, such as robot demonstrations, egocentric video, or driving logs, can talk to the Refinery team directly at info@deepen.ai about listing it before the beta opens.

About Deepen AI
Deepen AI is building a new data economy for Physical AI. Through sensor calibration, annotation, validation, and Deepen Refinery, it supports teams building robots, embodied AI, autonomous vehicles, and industrial systems.

Media contact: info@deepen.ai

Mohammad Musa
Deepen AI
+1 650-560-7130
info@deepen.ai

Media gallery

About The Author

Written By: EIN Presswire