Every AI breakthrough rests on three things: models, compute, data.
Two of those have had a decade of infrastructure investment. The third is still mostly hand-built.
Deep learning, LLMs, now Physical AI — each wave has needed bigger data systems than the last. And every time, the teams that got the data right are the ones who shipped.
We started Eventual after years of building petabyte-scale data systems for model training in self-driving, where we watched data become the bottleneck again and again. The best researchers we knew spent 80% of their time wrestling with data workflows (ETL, hard-example mining, dataloading) instead of training models.
So we gave sensor data the love it was missing. Traditional big-data systems fall apart on GPUs, video, lidar, and the high-frequency, unstructured logs a fleet emits. We built Daft, an open-source data engine that runs in production at exabyte scale inside Amazon, Mobileye, and Together AI.
MultiBase is what comes next: data infrastructure for teams building Physical AI, built natively for video and high-frequency sensor data. MultiBase indexes perception data at a deep semantic level beyond just embeddings, so you can run complex temporal and action-oriented queries on petabytes of fleet data in plain English: “left-arm grasp failures on deformable objects.” Then it feeds curated data to your GPUs at line rate, maximizing MFU. Everything runs directly on normal open formats (mp4, jpeg) in storage you own. No custom formats. No ETL. No dataloading pipelines to agonize over.
Good engineering and a new model every morning.