Reflection AI has signed a GPU-compute capacity agreement with SpaceX’s AI operation that could be worth up to $6.3 billion through 2029. Beginning July 1, the open-weight AI startup will pay $150 million a month for access to Nvidia GB300-based systems and related data-center hardware at SpaceX’s Colossus 2 complex in the Memphis area. [1]
The agreement gives a young model developer access to infrastructure normally available only to the largest AI companies. It also offers SpaceX another large external customer for the data-center business built around its former xAI operations, illustrating how advanced chips, power, cooling and networking have become as strategically consequential as model research itself.
A $6.3 billion headline with important caveats
The reported $6.3 billion figure is the maximum potential value of the arrangement, rather than a fully guaranteed four-year commitment. It reflects 42 monthly payments of $150 million from July 2026 through 2029.
Both SpaceX and Reflection may terminate the agreement after its first three months by providing 90 days’ notice. That flexibility means the announcement should not be read as $6.3 billion in booked revenue for SpaceX or an unconditional long-term spending obligation for Reflection. The companies have not publicly released the contract, and neither initially provided Reuters with additional details. [1][3]
Several of the information points needed to assess the commercial terms remain undisclosed. Public reporting does not specify how many GPUs or racks Reflection has reserved, how much electrical capacity is included, what utilization is guaranteed, or whether the monthly charge covers networking, storage, operations and facility services. Without those details, it is not possible to calculate an effective price per GPU, megawatt or unit of training compute.
What Reflection is getting
Reflection said the capacity will support its work on frontier open models. More precisely, Reflection is an open-weight AI company: it aims to make trained model parameters available, but that does not necessarily mean its training data, data licenses, training code and infrastructure are all open source. The distinction matters as the company positions itself as an alternative to fully closed model providers.
The startup was founded in 2024 by former Google DeepMind researchers Misha Laskin, its chief executive, and Ioannis Antonoglou. Reflection has argued that governments, enterprises and scientific institutions need models they can customize and inspect rather than systems accessible only through a provider-controlled API. [1][3]
That argument has particular relevance to the government market. In May, Reflection was selected as an AI-model provider for the U.S. Department of Energy’s Genesis Mission, where its models are intended to be customizable by national laboratories. [7] Still, as of July 1, Reflection has not publicly disclosed the architecture, parameter count, training data, benchmark goals, release schedule or production capabilities of the models it plans to train using the SpaceX capacity.
Some investors have promoted Reflection as a potential U.S. counterpart to China’s DeepSeek. The new compute deal gives that ambition meaningful infrastructure behind it, but access to hardware alone does not establish that a startup can match the capabilities, distribution or economics of leading closed frontier-model developers.
GB300 hardware at Colossus 2
The agreement identifies Nvidia GB300 systems, part of Nvidia’s Blackwell Ultra generation of AI hardware. Nvidia’s GB300 NVL72 reference platform is a liquid-cooled rack-scale design with 72 Blackwell Ultra GPUs and 36 Grace CPUs. Nvidia lists up to 20 TB of GPU memory, 130 TB/s of NVLink bandwidth, up to 800 Gb/s of networking per GPU through ConnectX-8 SuperNICs, and up to 360 FP16/BF16 Tensor Core petaflops per rack, subject to its configuration and sparsity qualifications. [6]
Those are platform specifications, not a description of Reflection’s allocation. Neither company has said that Reflection will receive full GB300 NVL72 racks, how many systems it will use, or whether its capacity is dedicated, shared or subject to other operational constraints.
SpaceX’s June prospectus describes Colossus and Colossus II as a connected, gigawatt-scale AI-training cluster spanning Memphis, Tennessee, and Southaven, Mississippi. The facilities support training and inference for frontier models including Grok, the model business inherited when SpaceX acquired Elon Musk’s xAI in February. [5]
For Reflection, leasing this infrastructure avoids the immense capital cost and lengthy construction timeline involved in building a frontier-scale cluster. For SpaceXAI, the transaction is evidence that the company can sell access to its AI infrastructure beyond its internal model-development work.
SpaceX turns AI capacity into a product
The Reflection contract is smaller than other reported SpaceX compute agreements, but it is significant for the type of customer it serves. TechCrunch reported that SpaceX has also struck much larger arrangements with Anthropic and Google, reportedly valued at about $1.25 billion and $920 million per month, respectively. [1] Those figures, like the Reflection headline value, need to be understood in light of contract terms and cancellation provisions that may not be public.
SpaceX’s broader AI strategy links several businesses that are usually separate: large-model development, GPU clusters, launch services, Starlink connectivity and, eventually, orbital computing. Its prospectus identifies chips, data-center construction and power generation as central constraints on that plan. The filing also says SpaceX expects to continue obtaining a substantial portion of its hardware from outside suppliers despite its plans for the Terafab semiconductor project. [5]
The deal also highlights the increasingly circular structure of AI financing and infrastructure. Nvidia is both the supplier of the GB300 technology being used and a major Reflection investor; Axios reported that Nvidia invested roughly $800 million in the startup’s financing. [2] Such relationships can help an emerging model company secure scarce capacity, while expanding the market for the chipmaker’s systems and software stack.
The bottleneck is bigger than GPUs
Advanced accelerators remain a crucial input to model training, but obtaining chips is only one part of deploying them. Frontier clusters require land, grid connections, substations, high-capacity networking, storage, liquid cooling, skilled operations and enough software engineering to keep large numbers of accelerators productive. Leasing from a specialized operator can therefore be more valuable to a startup than merely buying GPUs.
That is the strategic logic behind Reflection’s agreement. It provides the company with a path to large-scale training capacity immediately, while preserving the option to exit after the early period. The same optionality, however, makes the deal a less certain revenue signal for SpaceX than its $6.3 billion ceiling implies.
The transaction arrives as investors scrutinize whether the revenue opportunities attached to AI infrastructure justify the sector’s escalating valuations and capital spending. CFRA analyst Keith Snyder has questioned whether the aggressive AI-growth assumptions implied by SpaceX’s valuation are supportable. The criticism was directed at SpaceX’s overall valuation, not specifically at Reflection, but contracts such as this one are now part of the wider test: whether demand for leased frontier compute will translate into durable, profitable infrastructure revenue. [8]
For now, the agreement is best understood as a capacity lease, not a joint model-development program or an equity investment by SpaceX. It gives Reflection a major new route to train open-weight models, while leaving the key measures of its scale and economic value—including GPU volume, power allocation, guaranteed availability and eventual model output—unknown.
Editor’s Take
I see this less as a $6.3 billion model-company triumph than as a sophisticated capacity option. Reflection gets immediate access to a class of infrastructure that would otherwise take years, huge capital expenditure, and a serious power-development effort to assemble. For a startup pursuing open-weight models, that can be the difference between having a credible frontier roadmap and merely having one.
The headline number deserves restraint: a three-month exit structure and a 90-day notice period make this materially different from locked-in multiyear revenue. What matters next is not the nominal contract ceiling but the operational details still missing—reserved GPU count, dedicated versus shared capacity, delivered power, uptime commitments, and the effective cost of useful training compute. If Reflection produces capable, commercially deployable open-weight models from this allocation, the lease will look prescient; if utilization or model economics disappoint, the flexibility cuts both ways.
SpaceX’s more interesting move is turning its cluster into a product. The scarce asset is no longer just a GPU shipment; it is the integrated system of chips, networking, cooling, power, operations, and scheduling that keeps a large training run productive. That is a real business, but investors should avoid treating every large compute headline as durable infrastructure revenue until renewals, utilization, and margins become visible.
References
- TechCrunch – https://techcrunch.com/2026/06/22/spacex-inks-compute-deal-with-reflection-ai-an-open-source-ai-lab/
- Axios – https://www.axios.com/2026/06/22/open-source-ai-gets-more-compute-from-spacex
- Reuters via Investing.com – https://www.investing.com/news/stock-market-news/ai-startup-reflection-signs-computing-power-deal-with-spacex-4753536
- SpaceX 2026 prospectus – https://content.spacex.com/cms-assets/FINAL_Documents%20and%20Updates/SpaceX%20-%20EU%20Prospectus%20%28Approved%20by%20Bafin%29%20-%20June%205%2C%202026.pdf
- Nvidia GB300 NVL72 specifications – https://www.nvidia.com/en-sg/data-center/gb300-nvl72/
- Axios on Reflection’s DOE Genesis Mission partnership – https://www.axios.com/2026/05/22/reflection-ai-genesis-mission-energy-partnership
- Fortune on SpaceX valuation scrutiny – https://fortune.com/sites/jonmarkman/2026/06/15/spacex-valuation-2-trillion-how-much-ai-grok-elon-musk/
