Fireworks CEO responds to data retention controversy: no data stored for regular inference, batch tasks require customers to delete data themselves.

Beating AI News Flash: Fireworks CEO Lin Qiao published a lengthy post responding to the recent data retention controversy. Previously, Fireworks suffered a hacker breach, leaking chat records of some Notion AI users. Notion pointed out that Fireworks had contractually promised not to retain this data.

Lin Qiao explained that Fireworks' services fall into two categories:

The first is standard AI inference. Users send questions, the model generates answers and returns them, and the platform does not retain inputs and outputs long-term. Fireworks' Chat Completions, Nexus, and serverless training services all adopt this model, supporting zero data retention (ZDR).

The second is batch processing. Customers first upload datasets, then have the platform run tasks in batches. Since the platform does not know when customers no longer need these files, it retains them until the customer actively deletes them. Lin Qiao emphasized that OpenAI, AWS, and Google's batch processing interfaces adopt similar designs.

She drew an analogy: uploading files to a cloud drive and entering numbers into a calculator are two different things. The former requires users to manage and delete files themselves, while the latter only requires computing and returning results.

Lin Qiao stated that Fireworks will not retain copies after customers delete data. For standard inference and serverless training, the platform provides zero-retention guarantees, requiring no additional cleanup from customers.

Fireworks' official documentation also shows that enterprise customers can enable an account-wide zero-retention restriction. Once enabled, the platform will directly reject batch tasks and file uploads that require data storage. However, this restriction is disabled by default and will not automatically delete previously stored files.

Lin Qiao believes that whenever business allows, processing methods that require no data storage should be prioritized. Every additional copy of data retained is one more burden to protect and clean up.