SpaceX signs $6.3B compute deal with open-source AI lab Reflection

Key Takeaways

• Reflection AI will pay $150 million monthly for access to Nvidia GB300 chips at SpaceX's Colossus 2 facility
• The deal runs through 2029 and is worth up to $6.3 billion, though either party can exit with 90 days notice
• SpaceX now rents compute to Anthropic ($1.25B/month), Google ($920M/month), and Reflection ($150M/month)

Reflection AI, an open-source artificial intelligence startup, will pay SpaceX $150 million per month for access to Nvidia's GB300 chips at the Colossus 2 data center in Memphis, Tennessee. The contract starts July 1, 2026, runs through 2029, and carries a maximum value of $6.3 billion. Either company can terminate with 90 days notice after the first three months.

This is the third major compute deal SpaceX has announced in recent months. Anthropic pays $1.25 billion monthly for Colossus 2 capacity; Google pays $920 million. Reflection's deal is smaller, but for an open-source lab founded just two years ago by former Google DeepMind researchers, it's a significant capital commitment.

Why is SpaceX renting out AI chips?

The Colossus data center was originally built by xAI, the AI company Elon Musk founded and later folded into SpaceX. xAI intended to use the facility for its own model training. That hasn't gone as planned. As SpaceX's internal AI work stalled, the company pivoted to leasing its hardware to outside labs.

The shift made business sense. Nvidia's latest chips are in short supply, and labs will pay premium rates for immediate access. SpaceX now collects more than $2.3 billion monthly from Anthropic, Google, and Reflection combined. The rocket company has quietly become one of the largest AI infrastructure landlords in the U.S.

What does Reflection AI get from the deal?

Access to Nvidia's GB300 chips. The GB300 is Nvidia's current flagship AI accelerator, and demand far exceeds supply. Startups often wait months for hardware or pay steep markups. Reflection's contract guarantees immediate capacity at a fixed monthly rate.

For an open-source lab, compute is the bottleneck. Training competitive models requires thousands of GPUs running continuously. Without this deal, Reflection would struggle to match the training runs that OpenAI, Anthropic, and Google routinely execute. The company framed the announcement as proof that open-source AI can attract serious infrastructure investment.

“Recent events highlight how important open source is to the AI ecosystem, with more nations and enterprises recognizing the risks and costs associated with exclusively depending on closed models. Our deal with SpaceXAI signals Reflection's strategic importance within the frontier AI ecosystem, and more compute means more runway to build the world's best open models at scale.”

— Reflection AI spokesperson

The U.S. government ban on closed models

Reflection's pitch gains credibility from a recent policy shift. The U.S. government banned Anthropic's closed models, known as Fable and Mythos, citing national security concerns. The decision rattled enterprise buyers who had built workflows around those systems.

Open-weight models, which release their trained parameters publicly, sidestep some of those risks. Organizations can inspect, modify, and host the models themselves. They're not locked into a single vendor's API or subject to sudden policy changes. Reflection is betting that recent bans will accelerate enterprise interest in open alternatives.

How does Reflection compare to other SpaceX tenants?

Reflection's monthly spend is roughly one-eighth of Anthropic's. That's not surprising. Anthropic and Google are well-funded incumbents training some of the largest models in existence. Reflection is a two-year-old startup that has raised far less capital. The company is punching above its weight class by securing any Colossus 2 capacity at all.

The flexible termination clause matters too. Musk has publicly downplayed the three-year terms, noting that any party can exit at any time. This protects SpaceX if chip prices drop or new hardware makes GB300s obsolete. It also gives Reflection an exit if the open-source model doesn't pan out financially.

Can open-source AI compete at the frontier?

That's the central question. Open-weight models like Meta's Llama series have proven capable, but they typically trail the latest closed models by several months. The gap exists because closed labs have more compute, more data, and more researchers.

Reflection's bet is that $150 million monthly in compute closes part of that gap. If the company can train models that approach GPT-5 or Claude 4 performance, enterprises gain a credible alternative. If not, Reflection burns through capital without matching the incumbents.

The deal doesn't guarantee success. It guarantees hardware. What Reflection builds with that hardware determines whether open-source AI becomes a real competitor or remains a promising also-ran.

Frequently Asked Questions

What is the SpaceX Colossus 2 data center?

Colossus 2 is a data center near Memphis, Tennessee, originally built by xAI (now part of SpaceX) to train AI models. It houses Nvidia's GB300 chips and is now leased to outside AI labs including Anthropic, Google, and Reflection AI.

What are open-weight AI models?

Open-weight models publicly release their trained parameters, allowing anyone to inspect, modify, and deploy them. This contrasts with closed models from companies like OpenAI and Anthropic, which keep their weights proprietary.

Why did SpaceX start renting AI compute?

SpaceX's internal AI efforts stalled after xAI was absorbed into the company. Rather than let expensive Nvidia chips sit idle, SpaceX began leasing capacity to AI labs willing to pay premium rates for immediate access.

How much is Reflection AI paying SpaceX?

Reflection will pay $150 million per month starting July 2026. The contract runs through 2029 with a maximum value of $6.3 billion, though either party can exit with 90 days notice after the first three months.

Why did the U.S. ban Anthropic's models?

The U.S. government banned Anthropic's Fable and Mythos models, citing undisclosed national security concerns. The ban has increased enterprise interest in open-source alternatives that don't depend on a single vendor.

Source: TechCrunch / Kirsten Korosec