SpaceX targets over 2 million Nvidia Rubin GPUs in 5x compute buildout by 2027

SpaceX is planning a compute expansion so massive it could require more than two million of Nvidia’s next-generation Vera Rubin GPUs, Elon Musk disclosed during the company’s Q2 earnings call. The target: push total capacity from today’s 1.4 gigawatts to somewhere closer to ten gigawatts by the end of 2027. That’s a more than fivefold increase, and Musk made it clear the entire buildout will ride on Nvidia silicon. “Vera Rubin is the best architecture available,” he said on the call, signaling a full-throated bet on the chipmaker’s roadmap.

The math gets staggering fast. Current Colossus clusters run on roughly 540,000 GPUs split across H100, GB200, and GB300 systems, with at least 220,000 more GB300 units already planned. If SpaceX were to use Vera Rubin for the entirety of the new capacity, the fleet would blow past two million GPUs. The company hasn’t confirmed the final chip mix, but the trajectory is unambiguous. This is industrial-scale AI infrastructure on a level that makes most nation-states look like hobbyists.

What’s actually paying for this? Not Grok, at least not directly. SpaceX’s AI segment — which lumps together xAI and X after their February 2026 merger — posted $2.56 billion in Q2 revenue, nearly triple the previous quarter’s $818 million. But the segment still bled $1.26 billion in operating losses. Most of that revenue surge came from cloud contracts leasing out Colossus capacity, with a single unnamed customer — almost certainly Anthropic — accounting for roughly $1.52 billion. That’s a revealing detail. SpaceX is building the factory, but its competitors are helping foot the bill.

Musk’s long-term vision is even more outlandish: data centers in orbit. For now, the ground war is expensive enough. SpaceX booked $3.73 billion in AI segment operating losses for the first half of 2026 alone. Meanwhile, rivals are spreading their bets. Anthropic splits workloads across AWS Trainium, Nvidia GPUs, and Google TPUs. OpenAI has lined up six gigawatts of AMD capacity and at least ten gigawatts of Nvidia. The AI infrastructure race stopped being about who has the best model a long time ago. It’s about who can physically power the next one.