SpaceX's prospectus lists Cursor's "data and datasets" as payment for GPU capacity. Cursor's own pages disagree on whether Privacy Mode covers Grok Bot.
Cursor staff investigated whether confidential data reached SpaceX for AI training, according to The Information on 15 October. The report concerns the weeks before SpaceX's AI unit launched Grok Bot in August. The outcome of that check is not public. The contract behind the concern is.
The contract lists Cursor's data as payment for computing power. Cursor's own pages disagree on whether its privacy setting covers Grok Bot. The written promise is weaker than the one customers know from the editor.
SpaceX's prospectus, dated 11 June, describes a compute agreement signed on 19 April. SpaceX provides GPU cluster capacity, meaning racks of chips that train AI models. In return, "Cursor will contribute certain personnel, data and datasets, documentation, technical know how, workflows, prompts, specifications and software code." Models the two develop jointly are jointly owned. The arrangement ties the companies together beyond one purchase of chips.
The same document explains why SpaceX wants the data. It says "the depth of Cursor's integration with a high frequency coding workflow generates valuable developer interaction data, including coding generation prompts, iteration cycles, and software architecture decisions." It adds: "We expect that access to this data will enhance our model training and inference, including with respect to Grok."
The clause does not mention customer code. It lists what Cursor contributes. The interaction-data passage shows what SpaceX values most: how developers use the editor. Cursor said in April that it had been bottlenecked by compute. The prospectus shows what it paid with to get past that limit.
The clause is no longer just a term between two companies. The deal closed. Cursor announced on 14 August that it "has officially been acquired by SpaceX" and has access to "the largest fleet of GPUs in the world." The data is now an internal asset of a parent that told investors it expects to train Grok on it.
The prospectus also flags the risk. SpaceX tells investors that the arrangement "may create operational complexity relating to" cybersecurity, data governance and commercialisation strategy.
Privacy Mode is Cursor's setting for keeping code out of AI training. Cursor's privacy help page says it "ensures your code is never used for training by Cursor or other AI model providers." It also says Privacy Mode "works the same for Grok 4.5 as for every other model."
The same page narrows the promise when it addresses xAI training on Grok Bot data. It says Privacy Mode "in the Cursor editor" prevents training and that "Grok Bot runs on a separate product surface with its own data flows."
A Cursor forum reply of 12 August says the opposite. It says Privacy Mode "does apply to Grok Bot" and that with it on the data "isn't used for training by Cursor or any model provider (including xAI)." A forum post is not an official term, so a security review cannot cite it. Cursor has not put that sentence on its own help page.
SpaceX's own Grok Bot FAQ is narrower again. Training opt-out "follows the applicable Cursor account and privacy settings." That describes a choice tied to account settings, while the editor page promises "never" without conditions.
The Grok Bot security page, last updated 12 October, adds that "privacy and data-sharing choices are managed through Cursor account settings and, when required, the Grok Bot access flow." That second clause implies a consent step outside the Privacy Mode toggle that administrators already enforce. A security lead who audited the editor settings has not necessarily audited that flow.
The wording gap matters because of who sits on the other side. The company that listed Cursor's data as payment also expects interaction data to improve Grok. When that party says its product follows account settings, the settings text decides the case.
Two official pages use different verbs, "prevents" in one place and "follows" in the other. Both agree that Grok Bot requires data storage and does not support Legacy Privacy Mode, the no-storage setting.
Legacy Privacy Mode is the setup that keeps Grok Bot data flows off a team's repositories by design. An administrator who needs that guarantee can enforce it across the organisation. A team that wants Grok Bot's cloud agent has to accept storage and rely on the narrower promise.
The practical step for an enterprise or Teams buyer is specific. Ask Cursor to confirm in the data processing agreement or a signed addendum that Grok Bot inputs, files on its cloud computer and browser sessions are excluded from training by Cursor, SpaceX and xAI under Privacy Mode. Ask it to name the retention period too. Get the answer from Cursor, not from a forum.
The help page says that agreement governs personal data under enterprise contracts and that individual plans sit outside it. A team on individual seats has nothing to cite. The cheaper fix is a short written answer now. It should arrive before an incident review or a renewal assumes the editor promise travels with the product.
As of 15 October, the contract and documentation still read as they did in June: data is the stated currency, while the written promise on Grok Bot points to account settings. Until Cursor publishes a help-page sentence that matches the forum answer, customers with sensitive repositories should treat Grok Bot as a separate product with its own terms.