TL;DR
- Compute Deal: Reflection AI has secured SpaceX Colossus 2 access for Nvidia GB300 AI chips.
- Contract Caveat: The agreement is set at $150 million per month and could reach about $6.3 billion only if it runs through 2029.
- Open-Model Test: Reflection must turn the rented hardware into model releases or customer deployments that justify its open-model pitch.
Open-model startup Reflection AI has reportedly secured GB300 Nvidia AI chips and other hardware from SpaceX’s Colossus 2 data-center capacity near Memphis, giving it high-end infrastructure for training and running frontier AI systems. Wider contract terms remain unclear. The agreement moves Reflection deeper into one of artificial intelligence’s costliest resource races.
Reflection is reportedly set to pay SpaceX $150 million a month beginning July 1, with access running through 2029 if the agreement reaches the end of its term. Monthly payments could total about $6.3 billion.
Either company can end the compute agreement after the first three months with 90 days’ notice. An exit right keeps the maximum value tied to contract duration rather than certain SpaceX revenue.
Why the Compute Access Matters
Compute, in this context, means rented access to specialized AI chips, memory, networking, power, and data-center operations. Reflection has reportedly secured immediate access to chips and hardware inside SpaceX’s Colossus 2 facility, including Nvidia GB300 systems used for advanced reasoning inference workloads.
Reflection AI is trying to build models whose trained parameters are available publicly, a contrast to closed models from OpenAI, Anthropic, and Google Gemini. A Reflection AI spokesperson said governments and enterprises are paying closer attention to open systems because dependence on closed models can create cost and control risks.
So far, Reflection is still training its models, so the SpaceX access is still mainly a development runway. Open-model training with public weights and no distillation can demand more original compute than adapting a closed provider’s outputs.
Nvidia has invested $800 million in Reflection, while Reflection is now gaining access to Nvidia chips purchased by SpaceX, which is another case of what critics call “circular financing”.
Colossus Becomes a Customer Platform
SpaceX is also broadening its AI compute capacity beyond internal workloads. Originally built by xAI for its own AI efforts, the Colossus data center has now become a compute capacity backbone that SpaceX can rent to outside labs. Its infrastructure currently houses more than 220,000 Nvidia GPUs, creating a bridge between AI developers and the hardware needed to train large models.
SpaceX recently also signed compute-related deals with Anthropic, Google, and Cursor. Anthropic’s Colossus arrangement and Google’s compute arrangement are reported at $1.25 billion and $920 million per month, respectively.
For SpaceX, the customer mix turns Colossus into a recurring-revenue infrastructure business rather than a single-company training cluster. Deployed Colossus capacity gives AI labs a shortcut around AI chip scarcity because customers can rent hardware with power, cooling, networking, and operations already attached. At the same time, renting AI infrastructure limits expenditure risks than building up own costly data centers.
Reflection’s Next Execution Test
Reflection AI was founded in 2024 by former Google DeepMind researchers. Its founders include Misha Laskin and Ioannis Antonoglou, now leading the company as chief executive and chief technology officer.
Founder pedigree helps explain why investors and potential customers are watching the company which was last valued at $25 billion.
For Reflection, the practical test now is whether GB300 access can shorten the gap between its open-model pitch and deployed systems that customers trust.
