SpaceX Lands $6.3 Billion Compute Deal With Reflection AI for Open-Source GPU Access

Reflection AI just bought its way into the big leagues. The open-source AI lab agreed to pay SpaceX $150 million per month starting July 1, 2026, through 2029 for access to Nvidia's latest GB300 AI chips and supporting hardware at SpaceX's Colossus 2 data center near Memphis, Tennessee, according to TechCrunch. The total contract value could reach $6.3 billion.

This isn't a small startup renting scraps. Reflection AI is committing real capital to train and run open models at a scale previously reserved for closed labs like OpenAI and Anthropic. The deal signals that the open-source camp believes it can compete on raw training compute, not just distributed enthusiasm. CNBC notes this makes Reflection the latest outside company to tap Elon Musk's Colossus facility, following similar moves by bigger names.

The economics are stark. $150 million monthly means Reflection is betting it can monetize open-source models fast enough to justify burn rates that would sink most startups.

SpaceX has quietly transformed its Colossus supercomputer from an private training cluster into a commercial computing platform. Bloomberg reports the Reflection deal is the latest in a string of agreements as Musk's firm vies to become a major cloud computing provider. The company has already landed compute deals with Anthropic and Google, turning what Musk once termed a "gigafactory of compute" into a revenue engine.

The strategy mirrors what Amazon did with AWS, except SpaceX is starting with the most scarce resource in tech: access to cutting-edge Nvidia GPUs. By building massive data centers clusters and renting capacity to AI labs, SpaceX captures value from the infrastructure bottleneck rather than the models themselves. Mezha characterizes this as a deliberate move to dominate AI infrastructure, with Colossus 2 positioned as a hub for companies desperate for GB300 access.

The Memphis facility gives SpaceX geographic proximity to cheap power and cooling, plus room to expand. For customers, it offers something cloud giants can't always deliver: immediate access to the newest hardware without waiting in Nvidia's allocation queue.

Reflection AI is joining a club with deeper pockets. Anthropic signed a multi-year deal in May 2026 to rent the entire compute capacity of Colossus 1, gaining access to more than 300 megawatts and over 220,000 Nvidia GPUs including H100s, H200s, and GB200s, according to Chamath's newsletter citing New Street Research. Google reportedly agreed to pay approximately $920 million monthly for AI compute capacity through June 2029, with one report valuing that arrangement at nearly $30 billion total.

The pricing tier is telling. Anthropic and Google are paying premium rates for dedicated, large-scale capacity. Reflection's $150 million monthly sits below those figures but still represents a massive commitment for an open-source lab. The deal structure suggests SpaceX is segmenting its customer base: hyperscalers and well-funded closed labs at the top, ambitious open-source projects in the middle tier.

For Reflection, the comparison cuts both ways. It gets to play in the same infrastructure sandbox as Anthropic, but it also faces the same brutal economics of renting versus owning. Unlike Google, it can't fall back on a trillion-dollar balance sheet if training runs go over budget.

The specific chip matters. Reflection AI is getting access to Nvidia's GB300, the latest iteration in the company's data center GPU lineup. TechCrunch reports this is part of the package at Colossus 2, while Anthropic's earlier deal included H100s, H200s, and GB200s. The GB300 represents Nvidia's newest architecture, meaning Reflection will train on hardware that few competitors outside the largest labs can access.

This creates a hardware moat that open-source projects rarely enjoy. Traditionally, open models have been trained on older or less powerful chips, or distributed across volunteer compute. By securing GB300 access, Reflection can potentially match or exceed the training efficiency of closed labs, closing a gap that has limited open-source model quality.

The scarcity dynamic works in SpaceX's favor. Nvidia's newest chips are allocation-constrained, and demand from cloud providers, enterprises, and AI labs far outstrips supply. SpaceX's ability to secure large quantities and resell access makes it a critical intermediary, even as Nvidia itself tries to control distribution through preferred partners.

Reflection AI's bet could reshape competitive dynamics. Open-source AI has historically suffered from a compute gap, with closed labs spending billions on training while open projects relied on academic partnerships or crowdfunding. A $6.3 billion infrastructure commitment changes that calculation entirely.

The risk is concentration. Open-source advocates have long celebrated decentralization, yet Reflection is centralizing its compute through a single provider, SpaceX, which is controlled by Elon Musk, who also runs xAI, a direct competitor. Mezha raises questions about "open models scale and market influence" in this arrangement. If the open-source movement's compute backbone depends on a rival's infrastructure, independence becomes complicated.

The next 18 months will test whether Reflection can convert GPU hours into differentiated models and sustainable revenue. Open-source AI has proven it can distribute and improve models. Whether it can outspend closed labs at the training stage, and do so profitably, remains unproven. This deal gives it a real shot.