SpaceX has secured another massive compute tenant for its growing data center empire. Reflection AI, an open-source artificial intelligence startup barely two years old, will pay $150 million per month to rent Nvidia chips at Colossus 2, SpaceX's data center facility in Memphis, Tennessee. The payments are scheduled to run from 1 July 2026 through 2029, totaling roughly $6.3 billion, according to reports from CNBC and The Information, which first broke the news.
The deal structure is notable for its circularity. Nvidia, the world's leading AI chipmaker, provides the hardware that powers the entire transaction. But Nvidia is also an investor in Reflection AI, having poured around $800 million into the startup last year, as reported by Tech Funding News. This means the same company supplies the picks while owning a piece of the mine, a pattern that is becoming increasingly common in the frenetic AI infrastructure buildout.
Who is Reflection AI?
Reflection AI was founded in 2024 by Misha Laskin and Ioannis Antonoglou, both alumni of Google DeepMind. Antonoglou was part of the team that built AlphaGo, the program that defeated world champion Go player Lee Sedol in 2016. The startup has not yet shipped a public frontier model, but its pitch is built around open weights and transparency, a deliberate contrast to closed laboratories like OpenAI and Anthropic.
The company raised $2 billion in October 2024 at an $8 billion valuation, with backing from Nvidia, Sequoia Capital, and Lightspeed Venture Partners. It is now in the process of raising additional funds at a $25 billion valuation, according to Bloomberg. Reflection frames its work as "American open intelligence" and has ties to the U.S. Department of Energy's Genesis Mission, as well as Pentagon AI projects. "More compute means more runway to build the world's best open models at scale," the company has said.
SpaceX's compute landlord strategy
SpaceX originally built Colossus for its own AI model, Grok, developed by xAI. However, when the data center proved inefficient for its own purposes, SpaceX pivoted to renting out capacity to third parties. This has turned the Memphis site into a lucrative income stream. Reflection is the latest tenant, but certainly not the largest. SpaceX already rents capacity to Anthropic for about $1.25 billion per month and to Google for $920 million per month.
Together, these contracts give SpaceX more than $80 billion in committed compute revenue through 2029, according to Tech Funding News. SpaceX is also moving up the stack: it holds an option to buy the AI coding startup Cursor for $60 billion after Cursor's record initial public offering.
Inside Colossus 2
Colossus 2 is the newer of SpaceX's two Memphis campuses. The original Colossus 1 already houses more than 220,000 Nvidia GPUs. SpaceX aims to make the wider complex reach two gigawatts of power capacity, which would render it one of the largest concentrations of AI compute in the world. Such scale is necessary to serve the voracious demands of frontier AI training and inference.
The location in Memphis was chosen partly due to available power infrastructure and favorable tax incentives. The Tennessee Valley Authority has been working with SpaceX to upgrade the regional grid to support the massive energy needs. Environmental groups have raised concerns about the carbon footprint, but SpaceX has stated it is exploring renewable energy sources to offset the consumption.
Nvidia's double role
The Reflection deal is another example of Nvidia's strategy to embed itself deeply in the AI ecosystem. Not only does Nvidia sell the chips that power the data centers, but it also invests in many of the startups that end up renting those chips. This creates a virtuous - or circular - cycle: Nvidia's funding helps startups grow, those startups need more compute, and they buy Nvidia chips to meet that demand.
Nvidia's investment in Reflection was part of a larger $800 million round. The chipmaker has been prolific in making strategic investments across the AI landscape, including stakes in Inflection AI, Cohere, and CoreWeave. For Nvidia, these investments provide not just financial returns but also influence over the direction of AI development.
Market and investor reactions
Despite the steady stream of compute deals, SpaceX's stock fell sharply after its June 2025 initial public offering. Investors have been skeptical of the company's valuation, which priced at $180 per share but quickly lost ground. The compute contracts, while lucrative, have not fully calmed concerns about SpaceX's core rocket business margins and the regulatory risks surrounding its Starlink satellite internet division.
Some analysts argue that the data center rental business could become SpaceX's most profitable segment, exceeding even its launch services. The company's ability to attract tier-one tenants like Anthropic, Google, and now Reflection suggests strong demand for turnkey AI infrastructure. However, the capital expenditure required to build out Colossus to its full 2GW capacity is enormous, and any slowdown in AI investment could leave SpaceX with idle assets.
The open-source angle
Reflection AI's commitment to open weights puts it in direct competition with the leading proprietary models. The startup's backers believe that open models can rival closed ones in performance while offering greater transparency and customizability. The deal with SpaceX gives Reflection the compute needed to train these models at the frontier.
Open-source AI has gained significant traction in the developer community, especially after Meta's release of Llama models. Companies like Mistral AI and Aleph Alpha have also championed open approaches. Reflection's focus on "American open intelligence" adds a geopolitical dimension, positioning itself as a counterweight to Chinese open-source models and European efforts.
The Pentagon connection indicates that Reflection's models may also find use in defense applications. The Energy Department's Genesis Mission aims to leverage AI for nuclear security and energy grid optimization. Such government ties could provide Reflection with stable funding and priority access to classified computing resources.
Comparison with Anthropic deal
At $150 million per month, the Reflection contract is roughly one-eighth the size of the Anthropic deal, which reportedly costs about $1.25 billion per month. This reflects the different scales of the two companies: Anthropic is a leading frontier lab with models like Claude competing directly with GPT-4, while Reflection is still in its early stages. Yet the fact that SpaceX can attract multiple tenants at various price points demonstrates the flexibility of its data center business model.
SpaceX has structured all its compute contracts with standard escape clauses. Either party can walk away with 90 days' notice after the first three months. This provides tenants with flexibility if their needs change, but it also introduces revenue risk for SpaceX. So far, no tenant has exercised the exit clause, suggesting that the demand for compute remains robust.
Technical specifications and future plans
Colossus 2 runs on Nvidia's latest GB300 systems, which combine Grace CPUs with Blackwell GPUs. These systems are designed for large-scale AI training and offer significant improvements in energy efficiency compared to previous generations. The GB300 is part of Nvidia's aggressive roadmap that includes annual architecture upgrades, ensuring that SpaceX can offer cutting-edge performance to its tenants.
SpaceX is also exploring liquid cooling solutions to manage the heat generated by tens of thousands of GPUs running simultaneously. The company has partnered with cooling specialists to implement direct-to-chip and immersion cooling technologies. This is essential for maintaining reliability and performance in the dense server clusters.
Looking ahead, SpaceX has indicated it may expand Colossus to additional locations, including potential sites in the Midwest and Europe. The company's rocket launch capabilities could also be leveraged to send specialized compute modules to orbital platforms, though such plans remain speculative.
Founders and team background
Misha Laskin and Ioannis Antonoglou bring deep expertise from Google DeepMind. Antonoglou's work on AlphaGo demonstrated the power of reinforcement learning, while Laskin focused on large-scale training algorithms. Their combined experience gives Reflection a strong technical foundation. The startup has also hired several researchers from OpenAI and Meta AI, further bolstering its talent pool.
The company's culture emphasizes open collaboration, with plans to release model weights and training code publicly. This approach has attracted a community of developers who contribute to fine-tuning and optimization. Reflection hosts hackathons and research workshops to foster ecosystem growth.
Broader implications for AI infrastructure
The SpaceX-Reflection deal is emblematic of a broader trend where infrastructure providers become critical gatekeepers for AI development. Companies like CoreWeave, Lambda Labs, and Azure have also built massive compute clusters for rental. The competition for power and land has intensified, with data center operators scouting locations with cheap electricity and favorable regulations.
Memphis has emerged as a surprising hub due to its central location and willing local government. The city has granted tax abatements and expedited permitting to secure SpaceX's investment. However, the rapid growth has strained local water and power resources, leading to community meetings and sustainability concerns.
For the AI industry, the availability of massive compute on flexible terms lowers the barrier to entry for new startups. Reflection's ability to secure $150 million per month in rented compute without having a public model yet shows that investors and hardware suppliers are willing to bet on potential rather than proven results.