SpaceX–Reflection AI $6.3B compute deal at Colossus 2

Strip away the names and a clearer business emerges: SpaceX is quietly running a multi-tenant AI cloud, and Reflection is its third tenant. The monthly rents tell the story. Anthropic holds exclusive access to all of Colossus 1 at $1.25 billion per month ^[4]. Google signed on to pay $920 million per month for roughly 110,000 GPUs from October 2026 through June 2029 ^[6]. Now Reflection AI commits $150 million per month from July 2026 through 2029 for GB300 capacity at Colossus 2 ^[3]. Stacked together, SpaceX's combined compute revenue runs to roughly $2.17 billion per month — on the order of $26 billion a year ^[8]. That is the silent pivot: infrastructure SpaceX originally built to train Grok for its xAI unit is now a merchant cloud line serving the very labs xAI competes with. CNBC's framing captured it bluntly — Colossus is becoming its own business line rather than internal plumbing.

What is genuinely unusual is the buyer. Reflection AI, founded in 2024 by two former Google DeepMind researchers and valued at $25 billion ^[1], has committed up to $6.3 billion to compute before shipping a single public model. The bet is a positioning bet. Reflection calls this one of the largest announced open AI infrastructure commitments to date ^[2], and the timing is deliberate: open-weight models gained fresh momentum after the U.S. government's ban of Anthropic's closed models, opening a lane for a domestic, open answer to DeepSeek ^[2]. The company's own pitch is that nations and enterprises increasingly distrust depending solely on closed models, and that 'more compute means more runway to build the world's best open models at scale' ^[2]. In this frame, locking in scarce GB300 capacity is not the reward for success — it is the precondition for it.

The deal only makes sense against the backdrop of GPU scarcity. Advanced Nvidia chips remain the single biggest constraint on training and serving frontier models, so labs pay to lock in long-term capacity and providers monetize the shortage. Gartner's Arun Chandrasekaran frames compute as 'still the biggest constraint and source of advantage in frontier AI,' with access to it a 'key differentiator for AI innovation' ^[5]. Forrester's William McKeon-White reads SpaceX's move as opportunistic and rational, calling compute leasing 'one of the immediate growth areas for SpaceX through their xAI unit' ^[5]. Nvidia sits at the center of the flywheel as both the supplier of the GB300s being leased and a backer of Reflection itself ^[3]— a dual role that powers the deal and, as the next section shows, also fuels the skepticism around it.

Not everyone buys the flywheel narrative. The deal's structure leaves a notable gap between headline and reality: either party may terminate with 90 days' notice after an initial three-month period, which guarantees SpaceX only its first quarter — roughly $450 million — while the full $6.3 billion is contingent on the agreement running all the way through 2029 ^[3]. That short cancellation window is the load-bearing detail most coverage buried. Community reception has been markedly more cynical than the press releases, with the dominant frame being one of 'circular financing' — Nvidia backs Reflection, Reflection rents SpaceX compute, SpaceX buys Nvidia chips — alongside open doubts about whether a startup that has raised a fraction of $6.3 billion can actually cover the bill. A more sympathetic counter-read circulating in those same discussions holds that Nvidia bankrolling a domestic open-source lab is a rational strategic hedge in the post-DeepSeek landscape, not a shell game. The reported pressure on SPCX stock around the announcement, amid a separate large bond sale, only sharpened the question of how much of this revenue is durable ^[7].