Fireworks CEO Lin Qiao explains data retention policy, distinguishing ordinary inference from batch processing
Fireworks CEO Lin Qiao posted a long response on X on November 14, 2026, addressing the recent data retention controversy. Earlier, Fireworks was breached, and chat records of some Notion AI users were leaked. Notion said that Fireworks had promised in the contract not to retain this data.
Qiao explained that Fireworks’ services are divided into two modes. The first is ordinary AI inference (ephemeral processing): a user sends a question, the model generates an answer and returns it, and the platform does not store the input and output for the long term. Fireworks’ Chat Completions, Nexus and serverless training services use this mode and support zero data retention (ZDR). The second is batch processing (customer-managed retention): the customer first uploads a dataset, and then lets the platform run tasks in batches. The platform does not know when the customer no longer needs these files, so it keeps them until the customer deletes them proactively. Qiao said that OpenAI, AWS and Google’s batch processing interfaces use a similar design.
She explained with an analogy: uploading files to a cloud drive and entering digits into a calculator are two different things. The former requires the user to manage and delete the files themselves, while the latter only requires calculating and returning the result. Qiao said Fireworks will not continue to retain copies after customers delete the data; for ordinary inference and serverless training, the platform provides zero-retention guarantees without requiring customers to do any additional cleanup.
Fireworks’ official documentation shows that enterprise customers can enable a whole-account zero-retention restriction. Once enabled, the platform will directly reject batch tasks and file uploads that require data to be stored. This restriction is off by default, and it will not automatically delete files saved previously. Qiao believes that, as long as the business allows it, priority should be given to processing methods that do not require data to be stored. Every additional copy of data retained means one more burden to protect and clean up.
Original source:
Title: I want to clarify the confusion recently raised about our data retention policy.
Posted by: Lin Qiao (@lqiao)
Publication date: 2026-11-14
URL: x.com/lqiao/status/2121434466082242896
Source type: company CEO original post
Verification status: I actually read the full original post this time, verification date 2026-11-13
Supporting content: distinction between the two data control modes, differences between ordinary inference and batch processing, explanation of ZDR, analogy with OpenAI/AWS/Google, cloud drive and calculator analogy, statement that copies are not retained after deletion, and recommendation to prioritize non-retention methods
ABAB AI Insight
Qiao’s post turns the controversy into two standard API design patterns: customer-managed upload-use-delete lifecycle processing (batch tasks) and ephemeral send-compute-return processing (ordinary inference). In the former, the customer is responsible for deletion, and the platform no longer retains the data after a deletion request; in the latter, the content is not persisted.
The post makes clear that Fireworks’ Chat Completions, Nexus and serverless training fall into the latter category and provide ZDR guarantees, while batch interfaces fall into the former. Enterprise accounts can enable an account-wide zero-retention policy to reject tasks that require storage, but this option is off by default and does not automatically clean up already stored data. This explanation came after Notion disclosed that data used by its Insights feature had been retained and did not match the contractual commitment.
This distinction separates batch file storage from real-time inference paths and emphasizes that each retained piece of data adds to the burden of protection and cleanup. Follow-up points include whether enterprise customers actually enable the account-wide zero-retention restriction and how the deletion process for batch task data is carried out.
ABAB News · Cognition Law
An AI platform can distinguish services into two modes: ephemeral processing and customer-managed retention.
Batch upload tasks can require customers to delete proactively, and the platform does not retain copies after deletion.
An account-wide zero-retention restriction can reject tasks that require storage, but it is off by default and does not automatically clean up existing data.