Reflection AI SpaceX Compute Deal Puts Open Models On Colossus Capacity

Reflection Buys Into Colossus Compute

Reflection AI has signed its first compute deal with SpaceX, committing to pay $150 million a month for access to Nvidia GB300 AI chips and supporting hardware at the Colossus 2 data center near Memphis, Tennessee.

The arrangement begins July 1, 2026 and runs through 2029, with a maximum value of $6.3 billion.

After the first three months, either company can leave the contract by giving 90 days’ notice, so the headline value depends on how long both sides keep the capacity agreement in place.

The deal gives Reflection immediate infrastructure for training or serving frontier-scale open-weight models.

It also shows how SpaceX is turning AI chip holdings into a leasing business after Colossus was originally built by xAI for its own AI work.

Open Models Need Reserved Chips

Reflection is using the deal to support its open-weight strategy.

Open-weight models release trained parameters publicly, and Reflection has positioned that approach as an alternative to closed frontier labs such as Anthropic and OpenAI.

The company was founded in 2024 by two former Google DeepMind researchers.

It described the SpaceX agreement as one of the largest announced open AI infrastructure commitments to date, tying its model ambitions directly to reserved accelerator supply rather than only to research claims.

The contract is smaller than SpaceX’s agreements with Anthropic and Google.

Those customers pay $1.25 billion per month and $920 million per month, respectively, and their contracts also run through July 2029.

Elon Musk has publicly emphasized that the longer-term contracts can be canceled at any time.

That comparison is important for the economics of the site.

Reflection is not taking the largest Colossus commitment named in the available details, but its payment is still large enough to make monthly infrastructure cost a central part of the company’s open-model plan.

Compute Access Is The Business Constraint

Reflection has an open-model argument, but it still needs enough chips, power and data center capacity to turn that strategy into competitive systems.

For SpaceX, the deal extends a pattern in which Colossus capacity is rented to outside AI labs.

For Reflection, the contract converts compute from a future fundraising or procurement problem into a near-term monthly cost.

The Memphis site therefore becomes more than a host facility; it becomes the capacity base for testing whether an open-weight lab can compete while buying infrastructure from a Musk-controlled platform.

The unresolved issue is utilization.

The contract names the chips, location, start date, monthly payment and cancellation terms, but it does not provide Reflection’s training schedule, customer commitments or revenue that would show whether the Colossus capacity can support the cost of the agreement.