SpaceX Signs $6.3 Billion Reflection AI Deal, Turning Colossus Into the World's Newest AI Cloud

SpaceX Signs $6.3 Billion Reflection AI Deal, Turning Colossus Into the World's Newest AI Cloud

A rocket company that has never sold a server just became one of the most aggressive landlords in artificial intelligence. On June 22, 2026, SpaceX signed a compute deal worth up to $6.3 billion with Reflection AI, the Nvidia-backed open-source lab founded by two ex-Google DeepMind researchers. Under the agreement, Reflection will pay $150 million a month beginning July 1, 2026, and running through 2029, in exchange for immediate access to Nvidia's latest GB300 chips at SpaceX's Colossus 2 data center near Memphis, Tennessee.

It is the third blockbuster tenant SpaceX has signed in roughly two months, and it pushes the company's committed compute revenue past a staggering $80 billion through the end of the decade. For a business with no prior cloud-computing operation, that figure now rivals the contracted backlogs of incumbents like Amazon Web Services and Microsoft Azure.

A neocloud is born in Memphis

The Colossus complex was originally built for a single purpose: training Grok, the chatbot from Elon Musk's xAI. Less than a year later, it has been recast as a commercial compute platform renting capacity to the very labs xAI competes with.

The roster is remarkable. Anthropic, maker of Claude, is paying SpaceX roughly $1.25 billion a month for the entirety of Colossus 1. Google committed about $920 million a month for access to 110,000 Nvidia GPUs through June 2029. Together those two contracts alone represent more than $26 billion in annualized revenue. Add coding-tool maker Cursor and now Reflection, and the picture is of a company monetizing silicon almost as fast as it can install it.

The Reflection contract is structured with unusual flexibility. After an initial three-month commitment, either side can walk away with 90 days' notice. That escape hatch reflects both the speculative nature of the AI buildout and Reflection's own status as a young company spending heavily before it has a proven revenue engine.

Nvidia on both sides of the trade

The deal also underscores how tightly Nvidia now sits at the center of the AI economy. The chipmaker has invested $800 million in Reflection, which is valued at roughly $25 billion. Reflection will now run its models on GB300 systems that SpaceX bought from Nvidia. In other words, Nvidia is simultaneously a backer of the tenant and, indirectly, the supplier to the landlord, capturing value at multiple points in a single transaction.

Reflection, founded in 2024 by Misha Laskin and Ioannis Antonoglou, has positioned itself as a champion of what it calls American open intelligence, with stated ambitions toward government and national-security work. Antonoglou was one of DeepMind's founding engineers in 2012 and contributed to landmark systems including AlphaGo, AlphaZero and MuZero. Laskin led reward modeling for DeepMind's Gemini effort. Some observers have begun calling the startup a DeepSeek of the West.

The company framed the compute deal as a vote of confidence in open models. "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," a Reflection spokesperson said.

Why it matters

The obvious tension is competitive. If Grok is starved for compute, why would SpaceX rent its newest GB300 capacity to rival labs? The answer appears to be cash and capital discipline. SpaceX did about $18.7 billion in total revenue across all of 2025. Signing more than $80 billion in compute commitments in two months effectively lets the company finance a multibillion-dollar data-center buildout with other people's money, locking in long-term contracts that de-risk the enormous upfront cost of chips, power and cooling.

There is a strategic logic, too. By renting capacity at Colossus 2 while keeping Colossus 1 and future sites in reserve, SpaceX can throttle who gets compute and when. The 90-day exit clauses mean Musk retains optionality: if xAI's own training runs demand more silicon, those leases can unwind quickly. For now, monetizing idle and incremental capacity is simply more valuable than hoarding it.

For Reflection, the calculus is existential. Open-source labs cannot win on model quality alone if they lack the hardware to train and serve at scale. A guaranteed $150-million-a-month pipeline of GB300 access buys the startup a seat at the frontier table that until recently was reserved for trillion-dollar incumbents.

What to watch next

Three things will signal whether this arrangement holds. First, the 90-day clauses: if either SpaceX or Reflection invokes them after the initial period, it will reveal how fragile these mega-deals actually are. Second, xAI's own compute appetite. Any sign that Grok's roadmap is being squeezed could force Musk to choose between renting to rivals and feeding his own lab. Third, regulatory and antitrust attention on Nvidia, whose dual role as investor and supplier in deals like this is drawing scrutiny as the AI infrastructure market consolidates.

What is already clear is that the definition of a cloud provider has changed. A rocket company is now a top-tier AI landlord, and the line between building models, building chips and building data centers has all but disappeared.