Cursor, Notion Exit Fireworks; Still 3 of 21 Case Studies

Fireworks AI’s customers page still features Cursor and Notion, two customers that say they have moved off its platform after its security incidents. Of the 21 case studies listed on the page when we read it on 15 November, two are Cursor’s and one is Notion’s: 3 of 21, or about 14%, by our count. The 21 entries cover 17 companies.

Cursor’s update dated 13 November says it “Moved all inference traffic off Fireworks”. Notion’s advisory says it disconnected its systems from Fireworks by the morning of 3 November and is “no longer serving open-weight models through Fireworks”. A quote from Cursor’s research lead also runs near the top of the page: “We work with Fireworks to push on inference.”

Business Pill · Embedded Dependency

An evergreen Business Engineer explainer on embedded dependencies: what is inside what you buy. It teaches the general idea only and does not discuss Fireworks, Cursor, Notion or these incidents.

What the Cursor Case Study Says

Fireworks published “Cursor Composer 2 + Fireworks” on 26 June. It says Fireworks “powers distributed rollout and inference infrastructure across multiple clusters” for Cursor’s reinforcement learning, and that “Cursor runs RL across 3–4 distributed global clusters unified through Fireworks infrastructure”. The page lists “6–10x lower inference cost” against comparable frontier coding models, a Fireworks claim.

The case study quotes Federico Cassano, Cursor’s research lead: “We have finite engineers like everybody else. We would prefer to have engineers make training more efficient and more precise rather than spin up an inference effort.” Cursor’s update, 140 days after that case study by our arithmetic, describes Fireworks as “the third-party training and inference vendor”. The second Cursor entry on the page is “How Cursor built Fast Apply using the Speculative Decoding API”.

What the Notion Case Study Says

Notion’s entry is titled “How Notion cuts latency 4x and scales enterprise AI workflows with Fireworks”. Notion’s advisory, updated 12 November, says it used Fireworks “as a general inference provider to serve open-weight models inside our product” and for classification jobs behind its Insights feature, which it has disabled. It says the incident affects less than 1% of Notion workspaces. Our Notion piece sets out its timeline.

The Other 15 Companies

The other 18 entries cover 15 companies, including Vercel, Sourcegraph, Factory AI, Genspark, Upwork and Cresta. The page’s claims include Zeta, which “meets the closed frontier at 94% lower cost and 6x speed” with Fireworks-trained models, and Paper Instruments, which “trains a frontier-level model for knowledge work at ~98% lower inference cost”. A featured section outside the numbered list says Fireworks helped Juicebox by “cutting annual inference costs by $3.2M”.

CNBC reported in July that other Fireworks clients include Elastic, GitLab and MongoDB. None of the three has an entry in the page’s numbered list, by our count. CNBC also reported that about half of Fireworks’ revenue came from Cursor as of last year; our piece on the 121 days covers that figure.

What Fireworks Has Said

The Fireworks security bulletin we read on 15 November does not mention Cursor or Notion. It says it has not detected any impact to its Inference, Training SDK and Nexus products, “which are all zero data retention (ZDR) by default”, and that it has disabled self-serve dataset uploads and downloads used by Batch Inference and dataset-based training while it investigates. Our piece on Cursor’s update covers Cursor’s account of the breach.

The Structural Read

As we read the Cursor case study, the tie ran deeper than serving answers: Fireworks says its infrastructure unified Cursor’s reinforcement learning across 3–4 clusters, and Cursor’s own update calls Fireworks its training and inference vendor.

As we read the rest of the list, the remaining 15 companies carry the page’s cost claims, from Zeta’s 94% lower cost to Paper Instruments’ roughly 98% lower inference cost, both figures from Fireworks’ page.

Three Implications

FOR FIREWORKS Its bulletin says it has not detected any impact to its Inference, Training SDK and Nexus products, which it says are zero data retention by default.

FOR BUYERS READING CASE STUDIES Cursor’s and Notion’s own posts, dated 13 and 12 November, are the record of where they stand with Fireworks.

FOR READERS OF THE COVERAGE Our piece on the 121 days covers CNBC’s revenue figure and the round.

The Business Engineer Lens

This story maps onto the Business Engineer framework The State of The Inference Economy.

The framework says: “Inference is a revenue engine: continuous, distributed across millions of applications and agents, incurring marginal cost with every query.”

As we read it, the case-study page is where an inference provider shows that revenue engine to buyers, which is why the named applications on it matter. The framework does not discuss Fireworks, Cursor, Notion or these incidents.

What Is Not Established

Most entries on the page carry no date, and a case study describes a past project, not necessarily a current contract. Among the documents we read, only Cursor and Notion state that they moved off Fireworks; we did not check the other 15 companies. Cursor’s account of how its key was exposed is its own. We did not contact Fireworks, Cursor or Notion for this piece.

The Bottom Line

Fireworks’ customers page still lists Cursor in two of its 21 case studies and Notion in one, 3 of 21 by our count. Cursor’s update dated 13 November says it moved all inference traffic off Fireworks, and Notion’s advisory says it is no longer serving open-weight models through Fireworks.