OpenAI wants your data — and it's offering two very different deals

OpenAI just rolled out a formal pitch for its Data Partnerships program, and the structure reveals a lot about where the company thinks its next training advantage will come from. They’re not just scraping the web anymore. They’re asking organizations to bring them large-scale, real-world datasets that aren’t already publicly available, spanning text, images, audio, and video. The stated goal is to build models that “deeply understand all subject matters, industries, cultures, and languages.”

There are two tracks on offer, and the distinction matters. The first is an Open-Source Archive, where partners contribute data that gets released publicly for anyone to use in AI training. That’s a direct nod to the open-source community and a signal that OpenAI wants to be seen as an ecosystem player, not just a walled garden. The second track is for Private Datasets, where a company’s proprietary data gets used to train—and presumably fine-tune—OpenAI’s own models under strict access controls. If you’re a business that wants GPT to actually understand your domain, this is the door you walk through.

The company is also signaling serious technical muscle to sweeten the deal. They name-drop their in-house optical character recognition (OCR) for digitizing old PDFs and automatic speech recognition (ASR) for transcribing spoken content. They’ll even help clean up messy data. The message is clear: format isn’t a problem. As they put it, they’re “particularly looking for data that expresses human intention,” like long-form writing or conversations, rather than disconnected snippets.

Existing partnerships give a sense of the playbook. OpenAI worked with the Icelandic Government and a local tech firm to dramatically improve GPT-4’s Icelandic capabilities, and they tapped the Free Law Project’s massive legal document archive to sharpen the model’s legal reasoning. This isn’t a vague call for scraps—it’s a targeted hunt for high-signal data that teaches models how the world actually operates, one domain at a time.