SpaceX Signs $6.3 Billion Compute Deal With Open-Source AI Startup Reflection

SpaceX just landed its biggest AI customer yet. On June 22, 2026, CNBC reported that Elon Musk’s rocket company has signed a computing power agreement with Reflection AI, the $25 billion open-source frontier lab, worth up to $6.3 billion if the contract runs through 2029. Reflection will pay SpaceX $150 million per month starting July 1, gaining immediate access to Nvidia’s top-of-the-line GB300 chips at the Colossus data center complex in Memphis, Tennessee.

The deal is the most visible signal yet that SpaceX’s ambitions in AI infrastructure are real—and that its Colossus supercluster has quietly become one of the most contested compute assets in the industry.

Colossus Is Now a Cloud Platform

When SpaceX built Colossus in 2024 to train Grok for xAI, the 100,000-GPU cluster was an internal project. By mid-2026, it has evolved into something that looks far more like a commercial cloud. Recent customers include Anthropic, Google, and Cursor—which SpaceX acquired in April—and now Reflection AI. The company appears to be monetizing its Colossus infrastructure by renting compute to AI labs that need fast access to frontier hardware, effectively competing with AWS, Microsoft Azure, and Google Cloud for the highest-margin AI training workloads.

SpaceX’s advantage is timing: it has physical GB300 clusters operational now, at scale, while hyperscalers are still ramping their Nvidia Blackwell and Vera Rubin deployments. For a startup that needs to train a frontier model quickly and cannot wait twelve months for AWS or Azure capacity, Colossus is an attractive option despite the lack of managed services.

The Reflection contract also carries flexibility unusual for enterprise compute agreements: either party can terminate with 90 days’ notice after the initial three months. That structure acknowledges the volatility of frontier AI development—training runs get cancelled, architectures change, competitors leapfrog—while locking in near-term revenue for SpaceX regardless.

What Is Reflection AI?

Reflection AI was founded in March 2024 by Misha Laskin and Ioannis Antonoglou, two researchers with unusually strong pedigrees. Laskin led reward modeling for Google DeepMind’s Gemini series; Antonoglou co-created AlphaGo and was one of DeepMind’s longest-serving researchers. Their thesis, stated explicitly from launch, is that the world needs an American open-source frontier lab—one that can match the capability of closed models like GPT-5.5 or Claude Fable 5 while releasing weights that governments, defense contractors, and enterprises unwilling to route data through proprietary APIs can actually use.

The company emerged from stealth in March 2025 with $130 million at a $545 million valuation. Seven months later, in October 2025, it closed a $2 billion round at an $8 billion valuation, led by participation from Nvidia—a strategic investment that simultaneously solved Reflection’s hardware access problem and signaled GPU ecosystem endorsement. By March 2026, the valuation had reached $25 billion on the back of early government and national security partnerships, including work with the Department of Energy’s Genesis Mission and broader Pentagon AI efforts.

The paradox of Reflection’s position is striking: the company has no public frontier model yet, no products a consumer can use, and revenues that remain undisclosed. Yet it commands a $25 billion valuation and has now committed to paying SpaceX $150 million a month for compute access. The bet being made—by investors including Nvidia—is that the demand for sovereign, open-weight US frontier AI is large enough to justify building the model first and worrying about monetization second.

The Compute Sovereignty Thesis

The strategic logic behind Reflection’s existence is closely tied to the geopolitical fracture in AI. After the Trump administration placed export controls on Claude Fable 5 and Mythos models earlier in 2026, restricting their deployment to allied governments without case-by-case licensing, the demand for capable open-weight alternatives from US labs surged.

The argument: Chinese enterprises, Gulf sovereign AI buyers, and parts of Southeast Asia need frontier-class models, but cannot use closed US systems for data sovereignty reasons and will not use DeepSeek or other Chinese open-weight models for security reasons. Reflection, if it delivers weights that match frontier closed-model quality, would be the only product that satisfies both constraints simultaneously.

That is a large and currently unserved market. It is also, however, a market that assumes Reflection can actually build a frontier model—something it has not yet done publicly. The GB300 chips at Colossus are the hardware necessary to attempt it at the required scale.

SpaceX as AI Infrastructure Kingmaker

The broader implication of the Reflection deal is the accelerating transformation of SpaceX’s business model. The company that went public on Nasdaq in June 2026 under the SPCX ticker—with a first-day close that implied a $350 billion market cap—has positioned itself as more than a launch provider and Starlink operator. It is becoming an AI infrastructure company, with Colossus as its flagship product.

For investors and policymakers, this raises questions that were not on the agenda when SpaceX was primarily a space company. The concentration of AI compute in a facility owned by Elon Musk—who also controls xAI, one of the primary AI labs competing with Anthropic and OpenAI—is a conflict of interest structure with no precise precedent. SpaceX has not disclosed whether Colossus customers receive identical terms, or whether xAI’s access to the infrastructure confers advantages in training speed or chip priority.

What is clear is that Reflection AI has decided the access is worth $150 million a month. With the GB300 cluster operational by July 1 and a team of DeepMind alumni at the helm, the company will have the hardware, the capital, and the competitive pressure to ship something. The open-source AI race just got another serious contender.

What to Watch

Reflection’s first public model release is the most important near-term event to track. The company has signaled that its initial model targets the 400-billion parameter range with a mixture-of-experts architecture—comparable in scale to Llama 4 or Kimi K2.7—but with a specific focus on long-horizon reasoning and government-grade security evaluations.

If the model lands in Q3 2026 and benchmarks competitively with GPT-5.5 and Fable 5, the $25 billion valuation will look prescient. If it underperforms, or if the company cannot keep pace with the release cadence of well-funded competitors, the $150 million monthly commitment to SpaceX becomes a very expensive infrastructure experiment. For Elon Musk’s balance sheet, either outcome is lucrative. For the open-source AI ecosystem, the stakes are considerably higher.