Dropbox ditched AWS S3 for its own in-house system. Now it’s buying Tigris to sell S3-compatible storage to AI developers.

Dropbox is acquiring Tigris to accelerate its move into commercial storage infrastructure. The deal gives Dropbox an existing product, experienced team, developer relationships, and early customers. Dropbox already runs storage at huge scale, but selling infrastructure externally requires different capabilities.

Tigris provides S3-compatible object storage for AI applications. Its customers include Amplified, Railway, Runpod, and Replicate. The service helps developers keep data available across cloud and computing environments, while avoiding egress fees from Tigris storage.

Dropbox plans to use Tigris technology in DBX Agent Cloud, an upcoming service for developers building AI applications. It will initially sell standalone object storage, then may add filesystem and content-processing features. The acquisition still requires usual regulatory approvals, and Dropbox has not announced a launch date.

Object storage keeps data as individually addressable objects, such as datasets, images, videos, or model files. S3-compatible storage follows the interfaces and behavior developers commonly associate with Amazon S3. This gives applications a familiar way to save, retrieve, and manage large collections of data.

For example, an AI developer can store training data with Tigris while using computing resources from another provider. Existing software built to work with S3-style storage can generally connect through compatible operations instead of requiring a new storage integration. The article identifies Tigris as S3-compatible but does not detail every technical feature.

Compatibility can reduce switching friction and make storage easier to use across different cloud and compute environments. That matters as AI teams change model and compute providers. Dropbox plans to use Tigris’s technology for DBX Agent Cloud, initially offering object storage directly to developers.

Dropbox moved more than 90% of its user data from Amazon’s S3 service to an in-house system called Magic Pocket. The migration began in 2015. It marked a major shift from relying primarily on an external cloud-storage service toward operating Dropbox’s own storage infrastructure.

The reason was practical: Dropbox wanted better performance and lower costs. Moving that much user data required Dropbox to handle storage at substantial scale. Magic Pocket became evidence that the company could build and operate infrastructure for its own massive data needs.

That experience now supports Dropbox’s broader infrastructure ambitions, but the company distinguishes internal operations from selling storage to outside developers. Dropbox says external customers bring different requirements. Tigris adds a commercial product, existing customers, and developer relationships, helping Dropbox enter that market faster.

AI applications work with growing volumes of content, including images, videos, training datasets, and model artifacts. Their computing needs can also change. A developer may rent GPUs or other compute resources from providers in different clouds instead of keeping every workload inside one environment.

For example, a team might store a training dataset in one place but run model training on rented infrastructure elsewhere. If the data must travel repeatedly to reach that compute, transfers can add waiting time and charges. Tigris says it automatically keeps data close to where it is needed.

This approach matters because storage and compute do not always come from the same provider. Tigris also charges no egress fees for transfers out of its storage service. Dropbox sees consistent storage as a foundation for AI applications as models and compute providers change.

Egress fees are charges associated with transferring data out of a storage service. When an AI workload uses Amazon S3 but runs its computing outside AWS, the stored data may need to cross provider boundaries. Large datasets and repeated transfers can make that arrangement more expensive.

The same movement can slow an application or training job. AI systems often handle substantial datasets, images, videos, and model artifacts. Waiting for data to travel between storage and external compute can add delays, especially when workloads use different environments.

The article presents Tigris as an alternative to this problem. Tigris automatically keeps data close to where it is needed and charges no egress fees for transfers out of its storage service. That can make cross-environment AI workloads more predictable, although the article does not quantify savings or performance improvements.

Tigris’s multi-cloud approach separates storage from dependence on one cloud and compute environment. Developers can keep data accessible while using computing resources from multiple providers. This is useful for AI teams whose models, GPU needs, or deployment environments may change.

A developer could store a large training dataset with Tigris, then rent compute from another provider. Tigris says it automatically keeps data close to where it is needed. It also charges no egress fees for transfers out of its storage service, unlike the extra costs described for some Amazon S3 transfers.

The result is potentially more flexibility in choosing compute and deployment environments. Tigris already serves companies including Railway, Runpod, and Replicate. The article does not provide a direct price comparison or measured performance results, so its advantages are described as positioning rather than quantified savings.

Object storage is a way to store data as separate objects rather than as files arranged in a traditional folder hierarchy. Each object contains the data and information used to identify or manage it. This structure works well for large collections of varied digital content.

An AI team might store training datasets, images, videos, and model artifacts as separate objects. Applications can retrieve the needed items from storage while computing happens in another environment. Tigris builds on this model with S3 compatibility and automatic placement close to where data is needed.

The article highlights object storage as the first planned capability of DBX Agent Cloud. Dropbox says it may later add filesystem and content-processing features. Those additions would help applications work with different content types, track files, and resume activity across sessions as AI workloads grow.