Cursor's enterprise customers bought zero data retention. The SpaceX acquisition may have already made that promise moot.

SpaceX announced its binding merger agreement to acquire Anysphere (Cursor) for $60 billion in all-stock on June 16, 2026. The deal has not yet closed — expected Q3 2026. On the same day, xAI's Grok V9-Medium went live in the consumer Grok product on X and SuperGrok.

The story most coverage told: xAI used the Cursor acquisition to train a coding model on proprietary developer session data, building a data moat that Anthropic and OpenAI cannot replicate, and V9-Medium is going to close the benchmark gap against Claude Opus 4.8 (88.6% SWE-bench Verified) and GPT-5.5 (88.7%).

Three things are wrong with that story, and the one thing that is right has a different implication than the coverage suggests.

The benchmark gap closure is an unverified marketing claim

As of June 25, 2026, Grok V9-Medium has no published SWE-bench Verified score, no published HumanEval score, and no published LiveCodeBench score. It has been live in the consumer product for nine days. xAI has released no model card. The 1.5 trillion parameter figure — consistent across secondary sources — traces back to a single original reporting chain and has not been confirmed by xAI. (The figure also does not specify whether this is total MoE parameters or dense-equivalent; these are not comparable numbers.)

Current production Grok (version 4) sits at 72–75% SWE-bench Verified. Claude Opus 4.8 is at 88.6%. GPT-5.5 is at 88.7%. The gap V9-Medium is supposedly closing is 13–16 percentage points. That gap may close. But every article that stated it as fact ran before any evidence existed.

The coverage ran "xAI targets coding gap" as a factual description of a thing that has happened. The accurate description is: xAI has released a model it has positioned as targeting the coding gap, with no published evidence that it has done so. If V9-Medium's scores, when published, come in below Claude and GPT-5.5, every outlet that treated xAI's positioning as a factual result will have been doing PR.

The acquisition timeline is also wrong

Most coverage described the April 2026 date as the Cursor acquisition. It was an option agreement. SpaceX paid to secure the right to acquire Cursor for $60 billion or pay a break-up fee of approximately $10 billion. The binding merger agreement was filed June 16, 2026. The acquisition is not complete. Enterprise customers who evaluated Cursor's data policies in May and early June — after the April announcement but before the June binding agreement — did not have a finalized acquisition event to incorporate into their DPA risk assessment. Many will not have started that assessment at all.

The DPA affiliate-fencing gap is the story that matters

Cursor's enterprise offering is built on zero-data-retention as a foundational feature. The product documentation and enterprise privacy governance pages explicitly state that Privacy Mode prevents session data from being used to train Cursor's models or those of model providers. Enterprise customers sign DPAs with Cursor that are structured around this protection. A Cursor forum staff member confirmed in January 2026 that Privacy Mode and ZDR protections apply even to Grok model routing within Cursor — meaning xAI's model was considered a "model provider" subject to the same third-party protections as Anthropic and OpenAI.

That was the correct legal framing before June 16.

SpaceX confirmed before the acquisition announcement that it had been "jointly training a new model with Cursor for several months" and had "licensed a significant corpus of Cursor usage data." This is the training data for V9-Medium. There is no public statement from Cursor, SpaceX, or xAI specifying which users' sessions were included, under which privacy settings, or how the joint training arrangement was structured relative to the ZDR commitments.

Once the acquisition closes, xAI and Cursor will both be wholly owned SpaceX co-subsidiaries. xAI will no longer be a "model provider" or a "third party." It will be an affiliate.

Standard enterprise DPA templates — the templates that Cursor's 64% Fortune 500 customer base signed — are structured around third-party data sharing protections. They routinely require sub-processor agreements that list permitted external recipients of customer data. What they typically do not contain are explicit prohibitions on intra-affiliate data sharing. Affiliate sharing is considered internal, not third-party, and is either excluded from coverage or subject to different, often weaker, protections.

This means that after the acquisition closes, sharing Cursor session data with xAI may not constitute a breach of any current enterprise DPA — even for customers who signed under the explicit understanding that their sessions were zero-data-retention and would never be used for model training. The DPA terms wouldn't be violated. The understanding those terms were meant to protect would be.

No enterprise customer has received formal notice that this transition is occurring or that their DPA protections are being evaluated for continued adequacy under the new corporate structure. No Cursor or SpaceX statement has addressed the affiliate-fencing question.

The analyst quoted by InfoWorld put it directly: "Zero data retention survives only where it stays contractual, auditable, enforceable, and fenced from affiliate use." The affiliate use question is now open.

Why the "data moat" argument requires a distinction the coverage missed

The principle behind the V9-Medium training story is correct: session-level coding data captures developer intent (what was tried, accepted, rejected) that GitHub commit history does not. This is qualitatively different training signal. A model trained on Cursor sessions can theoretically learn to anticipate what a developer is trying to do before they express it in code.

The coverage treated this as an argument that V9-Medium will close the benchmark gap against Claude and GPT-5.5. But SWE-bench Verified measures a specific capability: read a GitHub issue, localize the bug in the codebase, write a patch, pass the test suite. It is a full-repo bug-fix task. Cursor session data is primarily autocomplete patterns, inline edit sequences, chat interactions, and suggestion accept/reject decisions. These are different capabilities. A model optimized on developer session data may produce better autocomplete and inline editing quality while performing similarly or worse on structured issue-resolution tasks.

The benchmark gap on SWE-bench Verified may or may not be where the Cursor data advantage shows up. If it shows up in user-perceived coding assistance quality and not in benchmark scores, that's a real competitive advantage — and one that doesn't show up in the published number that everyone is racing to.

Anthropic's position

Anthropic is the dominant Claude model backend for Cursor coding workflows. It is now simultaneously a competitor in the Cursor platform and a compute customer of the platform's acquirer: on May 6, 2026, Anthropic signed an agreement with SpaceX for access to 220,000+ NVIDIA GPUs at Colossus 1. The infrastructure Anthropic is renting from SpaceX may be the same infrastructure training the model built to displace Claude in Cursor.

Anthropic has made no public statement on the Cursor acquisition, its implications for developer data, or whether its compute agreement with SpaceX includes any data or non-compete provisions between tenants.

I think the data moat argument for V9-Medium is real if it delivers on the benchmark claim. I don't think there's any way to know that yet, and the coverage that treated the claim as established fact did readers a disservice. The DPA affiliate-fencing question is the one that has near-term consequences for real enterprise customers — and it's received one substantive article against hundreds of pieces on the benchmark claim that hasn't materialized yet.