SpaceX has agreed to provide Reflection AI, the open-source AI startup backed by Nvidia, with access to Nvidia GB300 chips and supporting hardware at its Colossus 2 data center near Memphis, Tennessee, beginning July 1, 2026. Reflection will pay $150 million per month through 2029 in a contract worth up to $6.3 billion at full term, with either party able to exit with 90 days’ notice after the first three months. The agreement is the fourth major compute deal that SpaceX has announced for its Colossus infrastructure – following arrangements with Anthropic at $1.25 billion per month, Google at $920 million per month, and coding tool developer Cursor, which SpaceX is separately acquiring – and it is the most strategically distinctive of the four. We in NEWSCENTRAL note that the Reflection agreement is not simply another compute sale: it is SpaceX’s first deal with an open-source AI lab, and it arrives at a moment when the case for open-source models has never been more commercially and politically compelling.
Reflection AI was founded in 2024 by Misha Laskin and Ioannis Antonoglou, both former researchers at Google DeepMind, where Antonoglou was a founding engineer who worked on AlphaGo, AlphaZero, and MuZero. The company has been valued at approximately $25 billion in connection with a reported $800 million investment from Nvidia, and has positioned itself as the American answer to DeepSeek – an open-weight frontier AI company whose models can be downloaded, inspected, and run on private infrastructure by any organization that chooses to do so. The company describes its mission as building American open intelligence, a framing that is deliberately nationalistic and commercially astute: it positions Reflection as the alternative to closed model dependency for governments and enterprises who watched Anthropic’s Fable and Mythos models get suspended by export control directive and concluded that the risk of closed-model reliance had just become undeniable.
The timing of the announcement is not coincidental. Anthropic’s export control shutdown generated more commercial urgency for open-source alternatives than any amount of theoretical argument about model sovereignty could have produced. Governments and enterprises that were previously indifferent to the closed-versus-open distinction are now actively reassessing their AI infrastructure strategies, and Reflection has built its pitch directly around that reassessment. The company has established early government credibility through partnerships with the Department of Energy’s Genesis Mission and Pentagon AI programs, giving it a buyer profile that extends well beyond the conventional AI startup customer base. Nvidia’s $800 million investment in Reflection – combined with the company now receiving access to Nvidia chips purchased by SpaceX – creates a circular relationship that illustrates the interconnected capital flows defining the AI infrastructure market in 2026. Freddy Miller, Senior Analyst at NEWSCENTRAL, observes that Nvidia is effectively funding its next generation of customers through strategic investments in companies like Reflection, while simultaneously selling the chips those customers will use to train the models that will generate demand for more Nvidia chips in the next cycle.
The Colossus data center was originally constructed by xAI, Elon Musk’s AI company, which became part of SpaceX when the two entities merged in February 2026. Colossus was built to power Grok, Musk’s AI assistant and competitor to ChatGPT, but as xAI’s internal AI pursuits have encountered commercial headwinds, SpaceX has pivoted to monetizing the infrastructure as a compute platform for outside AI companies. The Colossus complex has expanded to a planned 2 gigawatts of total power capacity across multiple buildings, with 555,000 Nvidia GPUs purchased at a cost of roughly $18 billion. At that scale, the economics of maintaining compute capacity for internal use only would be wasteful; converting it into a multi-tenant platform at the pricing levels Anthropic and Google are paying generates revenue that makes the infrastructure investment commercially defensible regardless of xAI’s own model performance.
The Reflection deal is notably smaller than either the Anthropic or Google agreements, which reflects both the company’s earlier stage and the commercial reality that a startup committing $150 million per month from July 2026 is making an extraordinary bet on its own ability to generate the revenue or capital raises required to sustain those payments. Reflection has not yet publicly released a frontier open-weight model, and its code research agent Asimov remained on a waitlist at the time of the announcement. The compute deal gives the company the GPU capacity to train models at frontier scale, but Nathan Clark, Enterprise IT and Systems Architecture Analyst, points out that compute access is necessary but not sufficient: Reflection still needs to ship models that enterprise and government buyers find competitive with closed alternatives, and the $150 million monthly payment creates a time pressure on that delivery that the contract’s 90-day exit clause does not fully resolve.
For NEWS CENTRAL, the most revealing variable will be whether Reflection ships a publicly available frontier open-weight model before the first exit window opens – and whether that model is good enough to justify the infrastructure commitment it has made.