Reflection AI commits $150 million a month to SpaceX compute before shipping its first open model

A downloaded checkpoint is the start of an integration project. The customer chooses the infrastructure, integrates its private data and maintains the system it builds ^[5]. The source's example is a hedge fund that keeps sensitive information on its own infrastructure while adapting a general model to research or trading workflows ^[6]. Reflection's product for that buyer is what it calls an "AI factory", a package of models, software, computing capacity and engineering built around the customer's own data and deployment requirements ^[3]. The idea is to sell the assembly work as a product, so each customer does less of it ^[19].

Axios's comparison with Chinese open-weight models is the only capability claim on the record, and it describes a model that has not been released ^[1][2]. Axios also reported, citing people familiar with the model, that a model behind the frontier could still be useful for a narrower task ^[7]. For that ranking to transfer to a given buyer, the evaluation tasks would have to look like the buyer's tasks. The ranking would also have to hold up after the buyer's own fine-tuning. Fine-tuning on private data is the main reason to pick open weights at all.

For a buyer whose reason for open weights is keeping data in-house, I'd put the deployment package ahead of the leaderboard position. At release, that means reading the license and checking whether the customization software runs on hardware the buyer controls. Runtimewire lists what enterprise buyers will want as a date, a license, pricing and proof the system performs on their workloads ^[17]. The October 4th report gives no release date ^[15].

Reflection has bought capacity first. Axios reported that the SpaceX agreement, signed in June, calls for payments of $150 million a month from July 1st, 2026, through 2029 ^[8]. That is $1.8 billion a year ^[1]. If the payments run to the end of 2029, the 42 months come to $6.3 billion ^[2]. The Nebius deal, signed in July, is worth more than $1 billion ^[9]. Both give Reflection access to Nvidia AI servers ^[10]. Nvidia is also a Reflection backer ^[12]. Runtimewire treats the scale of the deals as a measure of how much capacity Reflection is assembling before its first public model can be evaluated ^[16].

Laskin, the chief executive, has acknowledged that reaching the highest capability levels takes time. "To build a big rocket ship, it takes time," he said in an interview with CNBC ^[14]. The SpaceX payments run monthly in the meantime ^[8].