Cursor reveals the reason for XAI training: the compute power is stuck, and SpaceX has separately secured acquisition options worth $60,000,000,000

AI programming editor Cursor’s parent company Anysphere officially disclosed its decision to pursue model training in partnership with xAI in its 4/22 official announcement. Cursor’s official wording spells it out directly: “We want to push the training further, but we’ve一直被算力给卡住（bottlenecked by compute）。” At the same time, SpaceX announced on 4/21 that it holds an option to acquire Anysphere for 60 billion USD; if it does not exercise the option, it will pay about 10 billion USD as partnership compensation. These two transactions move in tandem, redefining Cursor’s stakes along two lines: “who can train Cursor’s model” and “who can buy Cursor.”

Cursor’s stated rationale: training is bottlenecked by compute

Anysphere’s official announcement cuts in through a product-evolution timeline: the development of three generations of models—Composer, Composer 1.5, and Composer 2—gradually increased Cursor’s capabilities in multi-file editing and command-line workflows, but each generation’s training scale hit the ceiling of external compute rental. The announcement directly states that over the past few quarters, Cursor wanted to push the training further, but compute became the ceiling.

The corresponding partnership scope: Cursor will leverage xAI’s Colossus infrastructure (a large AI data center located in Memphis, Tennessee, USA) as the next-generation model training platform. The announcement does not reveal any financial details, the number of GPUs, or the contract term—this stands in contrast with the 60 billion USD acquisition option later disclosed on the SpaceX side, in terms of information completeness.

SpaceX’s 60 billion USD option and 10 billion USD cooperation fee

Item Content Acquisition option SpaceX can acquire Anysphere as a whole within 2026 for 60 billion USD Cooperation alternative paid by SpaceX if the acquisition is not executed: about 10 billion USD as partnership compensation between both parties Comparison: the valuation for the primary funding discussed earlier in April for Cursor 4 is about 50 billion USD (2 billion USD fundraising)

According to CNBC’s compilation, SpaceX’s deal structure is effectively a combination of two options: “betting 60 billion USD on Cursor’s year-end valuation” and “locking in partnership rights for 10 billion USD.” This sidesteps Cursor’s next round of primary funding that would otherwise open to external VCs—per a TechCrunch report, SpaceX’s bid directly pre-empts Cursor’s originally planned $2 billion fundraising, causing new external investors to lose the window to enter.

Entering a gray area with existing ties to OpenAI／Anthropic

Cursor’s product itself has long used OpenAI and Anthropic models as optional backends for users. Cursor previously signed “zero-data-retention” agreements with both companies, guaranteeing that Cursor users’ code would not be used to train the corresponding models. Once the ownership structure moves into the Musk camp (xAI training + SpaceX’s acquisition option), the renewal and trust basis for agreements like these enters a watch-and-wait period:

If SpaceX ultimately exercises the acquisition right, Anysphere becomes a Musk-camp company, and its relationship with OpenAI would face a structural conflict (Musk and Altman have repeatedly publicly clashed)

The relationship with Anthropic is also delicate—Anthropic recently overtook OpenAI in secondary-market valuations, making it a direct beneficiary on the model adoption side for Cursor

Cursor’s intention to train the Composer series itself is precisely to reduce long-term dependence on third-party models

A perspective for crypto readers

This case forms a complete set with yesterday’s story about FTX’s bankruptcy liquidator selling 5% of Cursor’s equity for 200k USD: 2022 Alameda bought for 200k USD, 2023 bankruptcy disposition sold it back at the original price, and in 2026 SpaceX values it at 3 billion USD (calculated as 5% of a 60 billion USD valuation); the mismatch in timing in between is a typical example of how the crypto industry redistributes AI assets. SpaceX itself is also preparing for an IPO. SoftBank’s recent move to secure a 10 billion USD loan with OpenAI equity as collateral similarly shows that equity in AI flagship companies has become a core target for large-scale capital operations.

For Cursor users, near-term changes are minimal: the product keeps an optional multi-model backend, with xAI added as a new option. The key in the medium to long term is whether SpaceX will exercise the acquisition right within 2026—if it does, Cursor will become part of a Musk-camp AI stack; if it gives up, the 10 billion USD cooperation fee would be enough to fundamentally change Cursor’s financial structure and competitive position.

This article that Cursor publishes on why it picked xAI for training—compute is bottlenecked, and SpaceX also holds the 60 billion USD acquisition option—first appeared on Chain News ABMedia.