Reflection AI will pay $150M monthly for Nvidia GB300 chips at Colossus 2, joining Anthropic and Google as tenants
Every month, Reflection AI will wire $150 million to SpaceX for access to Nvidia GB300 chips housed at the Colossus 2 data center. Over the full term of the agreement, which runs through 2029, that adds up to $6.3 billion, one of the largest compute commitments in the open-source AI ecosystem.
The deal gives Reflection immediate access to hardware at Colossus 2. Reflection is an Nvidia-backed open-weights lab that trains multimodal models spanning video and audio. The company argues that releasing model weights publicly improves AI safety by allowing outside researchers to inspect and probe systems for vulnerabilities, a position that puts it in philosophical opposition to closed frontier labs like OpenAI and Anthropic.
Anthropic, a closed-model company, is already a Colossus tenant. So is Google. SpaceX is now underwriting both sides of the open-versus-closed debate in AI, driven by the logic of recurring infrastructure revenue.
SpaceX as compute landlord
SpaceX's Colossus operation has become one of the more consequential infrastructure plays in AI. The company is not building models or deploying AI products. It leases raw compute capacity, Nvidia chips at scale, to organizations that need more than they can source elsewhere.
Anthropic, Google, and now Reflection represent three distinct AI strategies: safety-focused closed models, hyperscaler integration, and open-weights distribution. Their shared landlord is an aerospace company whose primary business remains rockets and satellite internet.
Get the latest model rankings, product launches, and evaluation insights delivered to your inbox.
Whether SpaceX's compute margins match its aerospace margins is not established by available reporting. The Colossus business generates revenue independent of launch cadence or Starlink subscriber growth, a diversification with its own strategic logic.
Nvidia's position across the deal
Nvidia sits on both sides of this transaction. It backs Reflection AI as an investor, and it supplies the GB300 chips that SpaceX is leasing to Reflection. That dual position, hardware supplier and equity stakeholder in the customer, is not unique to this deal, but the scale makes it visible.
Reflection separately secured more than $1 billion in computing capacity from Nebius, announced in mid-July 2026, also using Nvidia GB300 hardware. Reflection has now committed to well over $7 billion in compute access across two providers. For a startup competing against frontier labs with far larger balance sheets, the bet is that raw compute access, combined with an open-weights release strategy, can close the gap.
What open-weights AI looks like at this scale
Reflection's public argument for open-source AI centers on safety through transparency. The company contends that open weights allow broader scrutiny than closed systems permit. That argument has supporters among researchers, and critics who argue that open weights lower the barrier for misuse.
The $6.3 billion commitment does not resolve that debate. It establishes that Reflection is pursuing open-weights training at a scale previously associated only with closed labs, and that it has found infrastructure partners willing to support that approach at market rates.
The deal runs through 2029. Reflection's first major model releases on this hardware will be the earliest concrete test of whether the compute investment produces competitive open-weights systems.