SpaceX Expands Its AI Compute Business With New Colossus 2 Deal

SpaceX has reportedly signed a new multibillion dollar agreement to rent AI computing capacity from its Colossus 2 data center, adding another major customer to its growing infrastructure business. The deal is expected to give AI startup Reflection access to NVIDIA GB300 hardware from July 2026 through June 2029.

The agreement is reportedly worth about $6.3 billion over its full term and would bring SpaceX roughly $150 million per month. It follows earlier compute agreements involving Colossus facilities, showing that SpaceX is moving quickly to monetize large scale AI infrastructure alongside its aerospace, satellite, and communications operations.

Reflection is expected to use the capacity for open AI model development. Access to modern GB300 systems could give the company far more room to train, test, and run large models without building its own massive data center footprint.

Colossus 2 Adds Another Customer as AI Compute Demand Grows

Colossus 2 is described as a large AI data center built around NVIDIA GB200 and GB300 accelerators. The facility is expected to support more than 550,000 GPUs, making it one of the largest AI compute projects currently under development.

Large AI companies are facing intense pressure to secure more computing capacity. Training advanced models requires huge amounts of power, networking, cooling, storage, and GPU hardware. Many companies cannot build this infrastructure quickly enough, so renting capacity from large operators has become an important alternative.

SpaceX has already taken a similar approach with Colossus 1, which has been linked to major compute agreements involving other AI companies. The newer deal suggests that Colossus 2 may also become a shared infrastructure platform rather than being reserved only for internal AI projects.

The Deal Raises Questions About Grok’s Long Term Compute Access

The expansion of SpaceX’s compute rental business also raises questions about how much capacity will remain available for Grok and other internal AI projects. Building competitive frontier models requires a large and reliable supply of high performance GPUs, especially as rivals continue to expand their own data center networks.

Renting out capacity can provide strong revenue and help justify the cost of building massive GPU clusters. It can also reduce the financial risk of maintaining expensive AI hardware at a time when demand is high.

However, there is a tradeoff. Every large block of compute capacity reserved for external customers is capacity that cannot be used immediately for internal training or inference workloads. That does not necessarily mean SpaceX is reducing its AI ambitions, but it suggests the company is balancing model development with the opportunity to earn recurring infrastructure revenue.

The reported Reflection agreement is another sign that AI computing has become a major business on its own. Companies no longer need to own every GPU they use, while operators with enough power, chips, and data center space can turn their infrastructure into a high value service.

For SpaceX, the long term challenge will be deciding how much capacity to keep for internal AI development and how much to rent to outside companies while demand for advanced GPU systems remains extremely high.