Deep Dive: SpaceX and Reflection AI Sign $6B+ Compute Leasing Deal

Deal Overview: A $150 Million Monthly Bet on Compute

Open-source AI lab Reflection AI has struck a remarkable compute leasing agreement with SpaceX. Under the terms of the deal, Reflection AI will pay SpaceX $150 million per month starting July 1, 2026 through 2029, in exchange for immediate access to computing power at SpaceX's Colossus 2 data center near Memphis, Tennessee.

SpaceX and Reflection AI Compute Deal

At the heart of this deal is NVIDIA's latest GB300 AI chips and their supporting infrastructure. Based on a rough calculation of the contract duration, the total value exceeds $6 billion — making it a mega-deal in the AI compute leasing space. To appreciate the scale of this transaction, consider the exponentially rising cost of training AI foundation models. GPT-3's training cost was estimated at around $4.6 million, GPT-4's reportedly exceeded $100 million, and the next generation of frontier models is expected to cost billions. At $150 million per month, based on current market leasing rates for top-tier GPU systems, this roughly corresponds to continuous access to tens of thousands of high-end GPUs — signaling that Reflection AI is gearing up for ultra-large-scale model training.

Colossus 2 Data Center: SpaceX's Compute Empire

From Rockets to Data Centers: A Cross-Industry Play

SpaceX is world-renowned for rocket launches and the Starlink satellite constellation, but its ambitions in AI infrastructure are equally bold. SpaceX's entry into the AI data center space is no accident — it's a natural extension of the synergies within Elon Musk's business empire. SpaceX has unmatched execution capability in rapidly delivering large-scale engineering projects, and the supply chain management, thermal management, and power systems engineering expertise accumulated from rocket manufacturing can be directly transferred to data center construction. Additionally, the global low-latency network connectivity provided by the Starlink satellite constellation could create future synergies with the data center business — providing network access for data centers in remote locations or serving as edge computing nodes for distributed AI inference. From a business model perspective, data center operations are a high-capex but stable cash flow business, complementing the high-risk, high-reward nature of SpaceX's launch business.

The choice of Memphis, Tennessee for the Colossus 2 data center was carefully considered. The Colossus data center was originally built by Elon Musk's xAI in Memphis in 2024. The initial version, Colossus 1, deployed approximately 100,000 NVIDIA H100 GPUs, setting a record at the time for the largest single AI training cluster. The construction speed was staggering — going from groundbreaking to operational in just about 122 days. Memphis offers several site advantages: affordable electricity from the Tennessee Valley Authority (TVA), relatively low land and construction costs in the mid-South region, and mature logistics and network infrastructure thanks to FedEx's headquarters being located there. Colossus 2 is the expanded version, expected to deploy next-generation Blackwell architecture GPUs with compute capacity far exceeding the original. SpaceX is extending its engineering prowess and capital advantages into the critical arena of AI compute supply.

The Strategic Significance of NVIDIA's GB300 Chips

NVIDIA's GB300 is its next-generation AI accelerator chip, representing the pinnacle of current AI training and inference hardware. Specifically, the GB300 is an advanced product within NVIDIA's Blackwell architecture, belonging to the GB200 series as a high-density rack-scale solution. Compared to the previous-generation Hopper architecture H100/H200 chips, the Blackwell architecture delivers multi-fold improvements in AI training and inference performance, with particularly notable breakthroughs in FP4/FP8 low-precision computation, Transformer Engine optimization, and NVLink interconnect bandwidth. The GB300 employs an NVLink72 fully-connected topology, linking 72 GPUs through high-speed interconnects into a single supercomputing node, dramatically improving the efficiency of distributed training for large models. In the current AI chip supply chain, NVIDIA holds over 80% market share in data center AI accelerators, with TSMC's CoWoS advanced packaging capacity being the key bottleneck constraining supply.

Reflection AI's decision to lock in GB300 access means they are making long-term plans for large-scale model training and iteration over the coming years. In a global environment where GPUs are in short supply, securing top-tier compute resources in advance has become one of the core competitive strategies for AI companies.

Reflection AI: Heavy Investment on the Open-Source Path

The Financial Muscle Behind an Open-Source AI Lab

Positioned as an open-source AI lab, Reflection AI's $150 million monthly compute expenditure indicates substantial financial backing. In today's AI industry, the competition between open-source and closed-source approaches is intensifying. The open-source AI space experienced explosive growth in 2024–2025: Meta's LLaMA series evolved from LLaMA 1 to LLaMA 3.1, making the leap from academic experiment to industrial-grade application, with its 405B parameter version approaching GPT-4-level performance on multiple benchmarks; France's Mistral AI became known for its lean and efficient model architectures, with its Mixtral mixture-of-experts model excelling in inference efficiency; and China's DeepSeek achieved remarkable performance at extremely low training costs. The core business logic of the open-source model is: build an ecosystem by open-sourcing model weights, then monetize through API services, enterprise customization, and technical support. This approach requires massive upfront capital investment to train foundation models, but once an ecosystem advantage is established, marginal costs become significantly lower than those of closed-source competitors.

Reflection AI's massive investment signals that the open-source camp is competing for compute resources with unprecedented intensity.

Strategic Rationale for Long-Term Compute Lock-In

The contract period from July 2026 to 2029 is worth noting. AI chip iterations happen at breakneck speed, yet Reflection AI chose a three-year lock-in period. The strategic logic behind this is clear:

Certainty of compute demand: They have a well-defined roadmap for future model training scale
Cost optimization: Long-term contracts typically secure more favorable pricing than the spot market
Competitive moat: In an era of compute scarcity, locking in supply is itself a strategic advantage

Of course, long-term lock-in carries risks as well. Currently, there are three main ways to acquire AI compute: building your own data centers (like Meta and Google), on-demand rental from cloud providers (AWS, Azure, GCP, CoreWeave), and long-term compute leasing contracts. The advantages of long-term contracts lie in price certainty and supply guarantees; the disadvantage is technology iteration risk — if a more efficient chip architecture emerges during the contract period (for example, NVIDIA's next-generation Rubin architecture is expected to launch after 2026), the locked-in compute could face declining cost-effectiveness. Reflection AI has clearly conducted a careful assessment, concluding that the GB300 will remain sufficiently competitive throughout the contract period, or that the sheer scale of their training needs justifies the investment.

Industry Impact and Trend Analysis

Shifting Dynamics in the AI Compute Market

This deal reflects several important trends in the AI industry:

First, compute supply is diversifying. Beyond traditional cloud providers (AWS, Azure, GCP), non-traditional players like SpaceX are entering the AI compute market, giving AI companies more options. Previously, AI-focused cloud providers like CoreWeave and Lambda Labs had already demonstrated the enormous potential of this niche — CoreWeave's valuation exceeded $19 billion in 2024. SpaceX's entry further intensifies competition on the supply side, which in the long run should help reduce AI training costs and give AI companies more bargaining power and supplier diversity in compute procurement.

Second, the compute arms race continues to escalate. A monthly expenditure of $150 million would have been almost unimaginable a few years ago. But in the era of foundation models, this level of investment is becoming standard for leading AI companies. For reference, Microsoft's capital expenditure exceeded $50 billion in fiscal year 2024, with the majority going toward AI infrastructure; Meta plans to invest $60–65 billion in AI infrastructure in 2025. The fact that Reflection AI, as an open-source lab, can reach this level of compute investment marks a significant expansion of the AI compute competition beyond the exclusive domain of tech giants to a broader set of participants.

Third, the business model for open-source AI is maturing. An open-source lab capable of sustaining compute investment at this scale must have a clear path to commercialization and strong fundraising capabilities.

Implications for the Global AI Industry

Against the backdrop of global compute competition, this deal serves as a reminder: the ability to build and secure compute infrastructure will largely determine the ultimate landscape of AI competition. Whether through building your own data centers or locking in compute resources through long-term contracts, early positioning is crucial. Notably, the compute competition has transcended the corporate level and risen to the level of national strategy. U.S. chip export controls on China, the EU's AI infrastructure investment plans, and Middle Eastern sovereign wealth funds' massive investments in AI data centers all indicate that compute is becoming the new strategic resource — the oil of the AI era.

Conclusion

The $6 billion+ compute leasing deal between SpaceX and Reflection AI is not just a business partnership — it's a landmark event signaling the AI industry's entry into a new phase. When an open-source AI lab is willing to commit billions of dollars in long-term contracts for compute, we can be certain: the race for AI at scale has only just begun, and the battle for compute resources will be the most critical variable in this competition. This deal also foreshadows a profound restructuring of the AI value chain — a rocket company becomes a compute provider, an open-source lab signs a blockbuster lease, and traditional industry boundaries are being shattered by AI's immense gravitational pull. In the years ahead, we may see more cross-industry players enter the compute market, and the ways in which compute is acquired and allocated will profoundly shape the direction and pace of AI development.