The intelligence of a model is only as rich as the data it learns from. We believe the next leap in AI capability will not come from architecture alone — it will come from data that captures the full texture of how the world actually works.

Midcentury is a multimodal data research lab building the data layer for real-world AI. We develop and structure large-scale, high-fidelity datasets across voice, egocentric vision, gameplay, and robotics — and build the infrastructure to make this data directly usable in training and evaluation.

Most training data is assembled opportunistically. We work the other way: we form a perspective on where model capabilities are heading, design data shapes that capture the behaviors those models will need, validate them through internal R&D against real performance, and scale what works.

This research-driven process lets us build structured multimodal data aligned with real-world interaction and long-horizon tasks — data that is not just large, but shaped with intent.

• Built one of the largest available egocentric industrial datasets — over 1,000,000 hours of first-person, real-world interaction data for embodied AI and world models.
• Partnered with several AAA-quality game studios to license high-quality 3D interaction and gameplay data at scale.
• Developed a large-scale conversational dataset suite spanning underrepresented languages including Hindi, Arabic, and Finnish — 10,000+ hours with rich dialect metadata.

Our data powers frontier voice systems, world models, and embodied agents. We work closely with many of the world's leading AI labs.

We are researchers and operators who have built and scaled some of the most consequential AI work of the last decade. Our founding team brings together experience from Stanford AI Lab, Scale AI, and DeepMind — and we tightly integrate science and product development in everything we do.

If you are training frontier models and want access to data that actually reflects how the world works, we would like to hear from you.