Reflection AI Secures SpaceX GB300 Cluster for $150M Monthly

On June 22, 2026, Reflection AI finalized a $150 million per month infrastructure agreement with SpaceX. This multi-year contract grants the open-source research lab immediate access to Nvidia GB300 hardware hosted at the Colossus 2 data center in Memphis, Tennessee. The commitment represents a massive scaling effort for the lab, trading distributed cloud rentals for centralized, next-generation silicon.

Agreement Terms and Infrastructure

The contract takes effect on July 1, 2026, and extends through the end of 2029. Totaling approximately $6.3 billion over the 3.5-year term, the deal fundamentally shifts the financial requirements for operating a competitive open-source AI laboratory.

Contract Parameter	Detail
Buyer	Reflection AI
Provider	SpaceX
Hardware	Nvidia GB300
Monthly Rate	$150 million
Duration	July 2026 through December 2029
Location	Colossus 2 (Memphis, Tennessee)

The arrangement also marks a strategic pivot for SpaceX. By leasing out infrastructure capacity originally developed for internal xAI and Starlink simulation workloads, SpaceX is solidifying its position as a major compute provider, building on previous momentum from its enterprise GPU contracts.

The Colossus 2 GB300 Architecture

The hardware underpinning this deal is the Nvidia GB300, the successor to the Blackwell architecture. The GB300 features improved FP4 precision and demands significant power and thermal management. SpaceX constructed the Colossus 2 facility specifically to handle these requirements, expanding on the original 100,000 H100 Colossus cluster.

To sustain the thermal output of the GB300 chips, the Memphis facility utilizes SpaceX’s proprietary high-density liquid cooling loops and power management systems. Supporting this hardware footprint requires an estimated 150MW to 200MW, prompting new grid investments by the Tennessee Valley Authority. For engineering teams scaling PyTorch training, running workloads on infrastructure of this density eliminates the network latency bottlenecks typical of fragmented cloud deployments.

Scaling Open-Source Model Training

Reflection AI is utilizing this compute allocation to train a new generation of open-source models, tentatively named Reflection-3. The lab previously built its reputation on “Reflection Tuning”, a technique that allows models to detect and correct their own reasoning errors during inference.

Moving from decentralized rentals to a dedicated GB300 cluster provides the synchronous bandwidth required for massive pre-training runs. The $1.8 billion annual expenditure indicates that Reflection-3 is targeted directly at the capability class of proprietary frontier models, requiring vast hardware resources to achieve parity in reasoning and evaluation benchmarks.

If your team builds applications on open-weight models, this infrastructure deal signals that the performance gap between open and closed models will continue to narrow. Prepare your deployment pipelines to handle heavily parameterized, reasoning-optimized models by late 2026.