SpaceX Turns Colossus 2 into Cloud Business for the New AI Race

SpaceX has signed an computing infrastructure agreement with Reflection AI that could generate around $6.3 billion by 2029. The deal positions Elon Musk’s company in a vastly different field beyond space launches and Starlink: leasing computing capacity to train and run frontier AI models.

According to Reuters, Axios, Quartz, and The Wall Street Journal, Reflection AI will pay $150 million per month starting July 1, 2026, for access to NVIDIA GB300 chips and associated hardware at the Colossus 2 data center. The contract includes a clause allowing either party to terminate with 90 days’ notice after the first three months. This detail is significant because it introduces flexibility in a market where GPU demand shifts rapidly.

Computing Becomes the New Bottleneck

This news confirms a concept already dominant in the cloud sector: the strategic scarcity of AI isn’t just about models, talent, or data. It’s in computing capacity. Those with early access to cutting-edge GPUs, energy, networking, cooling, and data center space can train larger models, serve more users, and accelerate the pace of updates.

Reflection AI fits into this landscape with a profile distinct from other major infrastructure clients. The company works on open models and presents itself as an alternative to the closed systems of large laboratories. Its move comes at a time when governments, corporations, and public agencies are paying closer attention to risks associated with dependence on proprietary models, especially concerning control, transparency, and deployment in private environments.

To compete at that level, an open-model startup requires something that can’t be improvised: sustained, large-scale access to advanced hardware. NVIDIA’s GB300 chips within a facility like Colossus 2 would give Reflection a much stronger foundation for training and deploying next-generation models. The agreement doesn’t guarantee that Reflection will release a competitive model, but it significantly lowers one of the main entry barriers.

SpaceX is No Longer Just Launching Rockets: Leasing AI Infrastructure

The deal also highlights SpaceX’s transformation into a broader tech infrastructure provider. For years, their story relied on reusable rockets, space contracts, Starlink, and satellite deployment. Now, a new business line emerges: AI data centers and computing services.

The Colossus infrastructure originated within Musk’s tech environment to support internal loads linked to Grok and xAI. But its evolution points toward a commercial platform for third parties. SpaceXAI had previously announced a computing agreement with Anthropic to provide access to Colossus 1, described as one of the largest and fastest deployable AI supercomputers.

The Reflection deal expands this movement. SpaceX isn’t just utilizing its own infrastructure for AI projects anymore; it’s beginning to sell capacity to other labs. Essentially, acting as a specialized cloud provider — offering very specific products: access to densely packed GPU clusters.

This strategy can make financial sense. Building AI data centers requires enormous investments in chips, energy, networking, land, cooling, and operation. Leasing capacity to multiple clients can turn this infrastructure into a steady income stream. The agreement with Reflection alone could generate around $1.8 billion annually if maintained for a full fiscal year.

Nonetheless, it introduces operational tension. The same capacity leased to external clients could be needed for internal projects. Musk had already noted in May that some computing agreements should be approached with caution due to termination clauses and the need to maintain flexibility against potential future capacity restrictions.

NVIDIA, Reflection, and the Circular AI Supply Chain

This agreement exemplifies another characteristic of the current industry: increasingly circular relationships among AI companies, chip suppliers, and infrastructure providers. Axios reports that NVIDIA invested $800 million in Reflection AI, and now Reflection will use NVIDIA hardware operated by SpaceX. In this loop, NVIDIA acts as investor, tech supplier, and indirect beneficiary of the computing demand that drives growth for its clients.

This pattern repeats across the market. Major labs sign cloud deals, cloud providers purchase GPUs, chipmakers invest in startups, startups secure compute contracts, and investors value companies based on their infrastructure capacity. Generative AI has created an economy where GPU access is almost treated as a strategic raw material.

For open models, this dynamic is especially sensitive. The promise of a more open ecosystem depends on publishers being able to train models with resources comparable to those of closed labs. Without sufficient compute, openness risks remaining limited to smaller, less competitive models or relying on external funding. Agreements like SpaceX’s with Reflection aim to position the company on a different scale.

This operation may also be attractive to governments and regulated sectors. Concerns about dependence on closed models, access restrictions, trade policy shifts, and control over sensitive data have grown recently. An open-model provider with native training capacity, or at least secured through long-term contracts, could appeal to organizations seeking greater technological sovereignty.

An Increasingly Specialized Cloud Market

The clear takeaway for the cloud sector is that AI is creating a new category of infrastructure providers. Not everyone competes like AWS, Microsoft Azure, or Google Cloud. Some offer GPU-as-a-service, others inference capacity, full clusters, or sovereign data centers—now SpaceX aims to position itself with large-scale facilities tied into its own tech ecosystem.

This could accelerate market fragmentation. AI frontier companies often prefer long-term contracts over building their own hardware and data centers from scratch, because delays in GPU access can be more costly than leasing. In a fast-changing technological landscape where six months can shift industry leadership, access availability is critical as much as price.

Conversely, this trend may lead to increased concentration. If AI computing capacity becomes confined to a few providers with preferred access to chips, energy, and funding, new entrants will find it harder to break in. Even well-funded startups will depend on existing infrastructure agreements. Reflection’s case shows that compute might become a prerequisite to compete, rather than just a line item in budgets.

SpaceX has not announced a new GPU or chip architecture. Instead, it’s done something arguably more impactful from a business perspective: renting out part of the “factory” that builds models. In the AI economy, that factory isn’t a traditional manufacturing plant anymore. It’s a data center with GPUs, energy, networking, cooling, and contracts capable of moving billions before the final model even reaches the end user.

FAQs

What have SpaceX and Reflection AI signed?
Reflection AI has signed an agreement to access compute capacity in Colossus 2, involving NVIDIA GB300 chips and related hardware. They will pay $150 million monthly starting July 1, 2026, according to the publicly available information.

What is the value of this deal?
The potential value is approximately $6.3 billion through 2029, though the contract includes a termination clause with 90 days’ notice after the first three months.

Why is this significant for the cloud market?
Because it shows that AI infrastructure is becoming a standalone business. SpaceX is beginning to act as a capacity provider for third parties, not just an internal user of its data centers.

What is Reflection AI’s relationship with open models?
Reflection AI works on open AI models. The agreement with SpaceX allows them to access more training and inference capacity at a time when interest in alternatives to closed models is increasing.

via: mexicobusiness