SpaceX Signs $6.3B Compute Lease With Reflection For GB300 Access

Reflection has signed a major compute agreement with SpaceXAI, giving the Nvidia-backed open-source AI startup access to GB300 chips and other hardware inside the Colossus 2 data center.

Under the compute agreement reported by Axios, Reflection will pay SpaceXAI $150 million per month starting July 1, 2026, after an initial ramp period. The deal runs through 2029 and is worth about $6.3 billion if maintained through the full term.

Either party can terminate the agreement with 90 days’ notice after the first three months. That structure gives Reflection access to high-end infrastructure without building a full hyperscale data center itself, while giving SpaceXAI another large customer for its rapidly expanding compute business.

The deal matters because open-source AI developers are trying to compete in a market where compute access has become one of the hardest constraints. Reflection is still training its models, and Colossus 2 access gives the company a larger hardware base as frontier model development increasingly depends on full clusters, power, cooling and deployment speed rather than chips alone.

SpaceXAI Adds Another Compute Customer

Reflection joins a growing list of AI customers tied to Elon Musk’s compute infrastructure. Anthropic previously tapped SpaceX’s Colossus 1 for large-scale Claude capacity, while Google signed a separate agreement to pay SpaceX $920 million per month for cloud compute.

Reuters reported that Google’s agreement gives it access to about 110,000 Nvidia GPUs, CPUs, memory and related compute infrastructure from October 2026 through June 2029. SpaceX’s compute agreements with Anthropic and Google were already valued at more than $70 billion in aggregate if held to term, before the Reflection deal added another multibillion-dollar customer.

The Reflection agreement also fits the newer customer mix. Anthropic needs capacity for Claude. Google wants large-scale AI infrastructure. Cursor entered SpaceX’s orbit through a separate AI coding push after the Anysphere acquisition. Reflection brings the open-source model angle, putting SpaceXAI capacity behind a company trying to compete with closed frontier labs.

Compute Becomes The AI Market Bottleneck

The deal reinforces how AI competition is moving from model announcements into infrastructure access. Nvidia chips remain scarce, but the bigger constraint is often complete deployed capacity: data-center space, power, networking, cooling, cluster reliability and the financial ability to reserve hardware for years.

Reflection’s arrangement also shows how Nvidia’s ecosystem is feeding its next wave of customers. Axios reported that Nvidia invested $800 million in Reflection, which is now leasing access to Nvidia chips purchased by SpaceXAI. That circular structure is becoming more common across the AI market, where chipmakers, cloud operators, model developers and infrastructure owners are increasingly suppliers, investors and customers to one another.

SpaceX’s AI infrastructure expansion has already drawn capital-market scrutiny. The company has been linked to a possible post-IPO bond sale as investors weigh how much debt and compute revenue can support its AI buildout beside launch, Starlink and satellite infrastructure.