Our intention is to charge on the "dataset under management." ie the number of unique datapoints in your latest labeled dataset that's stored in Aquarium. In practice, you can think about datapoints as "the number of labeled examples that we index / generate embeddings for."

Some examples:
- In a classification task, we only index embeddings for each example (say, each image / audio clip / row), so we just count # examples.
- In an object detection / instance segmentation task, we index embeddings for each example (say, each image / pointcloud) and each box / polygon in that example. So we count # examples + # boxes / polygons.
- In an image semantic segmentation task, we index embeddings for each image. So we count # images only.

This means we do not charge for metadata fields attached to each example (ie device id, timestamp, etc.). We do not charge for model inferences on these datasets. And we currently do not charge for unlabeled datapoints, though we may revise pricing for this in the future.

Aquarium offers a suite of pretrained models that can automatically generate embeddings on your own dataset.

Yes. Aquarium's client API allows users to upload their own model embeddings. In fact, we prefer it! Embeddings extracted from your own model tend to be better suited for your specific dataset and task.

Yes, users can upload URLs to data hosted outside of Aquarium with a little bit of setup. All of Aquarium's functionality functions the same as long as authentication is configured correctly.

Aquarium works without needing to store or host sensitive raw data on Aquarium servers. We support SSO and other enterprise-grade authentication schemes to interface with hosted data. In addition, we have a special "Anonymous Mode" which allows users to upload metadata to Aquarium without ever exposing their raw data.

We currently do not offer on-prem or VPC deployments. We find that Anonymous Mode satisfies the security needs of our most restrictive customers while also scaling well to massive datasets. However, we may examine supporting VPC deployments in the near future. If you're interested, talk to us.