SpaceX’s ambitious compute goals could require over two million Nvidia Rubin GPUs
Quick Answer
SpaceX aims to increase its compute capacity over fivefold by 2027, targeting over two million Nvidia Rubin GPUs.
Quick Take
Currently at 1.4 gigawatts, the expansion will leverage Nvidia's Vera Rubin architecture, with significant revenue growth from cloud contracts despite substantial operating losses.
Key Points
- SpaceX's compute capacity is currently 1.4 gigawatts, aiming for over ten gigawatts by 2027.
- Expansion plans include at least 540,000 Nvidia GPUs, with potential for over two million Rubin GPUs.
- The AI segment reported $2.56 billion in Q2 2026 revenue, despite a $1.26 billion operating loss.
- A single unnamed AI customer contributed approximately $1.52 billion to the revenue spike.
- Competitors like Anthropic and OpenAI are diversifying their compute resources across multiple providers.
DeepSignal Analysis
What happened
SpaceX plans to increase its compute capacity by over fivefold by the end of 2027, targeting more than two million Nvidia Rubin GPUs. Currently, the company operates at 1.4 gigawatts and aims to reach closer to ten gigawatts. The expansion will utilize Nvidia's Vera Rubin architecture, although the exact number of Rubin GPUs to be deployed remains unspecified.
Key evidence
- SpaceX currently operates at 1.4 gigawatts and aims to exceed ten gigawatts by the end of 2027, as stated by Elon Musk during a Q2 earnings call.
- The planned buildout includes approximately 540,000 GPUs from existing Nvidia models, with the potential for the new capacity target to exceed two million Rubin GPUs.
- In Q2 2026, SpaceX's AI segment reported $2.56 billion in revenue but also faced an operating loss of $1.26 billion, indicating significant financial challenges despite revenue growth.
Why it matters
The ambitious compute expansion reflects SpaceX's strategic focus on AI and cloud computing, potentially positioning the company as a significant player in the industry. However, the substantial operating losses raise questions about the sustainability of this growth model. The reliance on Nvidia's technology also highlights the competitive landscape, where other companies are diversifying their compute resources across multiple providers.
📖 Reader Mode
~2 min read
Aug 5, 2026
Nano Banana Pro prompted by THE DECODER
SpaceX wants to grow its compute capacity more than fivefold by the end of 2027, finishing the year above two gigawatts of power, according to Elon Musk. The entire buildout will run on Nvidia chips.
Musk shared the target during SpaceX's Q2 earnings call. By the end of 2027, capacity should be closer to ten gigawatts than five, he said. The company currently sits at 1.4 gigawatts, according to its quarterly report. SpaceX plans to build on Nvidia's upcoming Vera Rubin platform, which Musk called the best architecture available. The existing Colossus clusters run on Nvidia's H100, GB200, and GB300 systems.
SpaceX's IPO filings put the current and planned Colossus buildout at roughly 100,000 H100, 110,000 GB200, and 110,000 GB300 processors, plus at least 220,000 more planned GB300 units. That's about 540,000 GPUs total. The new capacity target would push the math well past one million Rubin GPUs. If the entire expansion used Vera Rubin, the number could exceed two million. SpaceX hasn't said how much of the buildout will actually be Rubin.
Musk's AI company xAI merged into SpaceX in February 2026 through a deal funded mostly with stock. The transaction valued SpaceX at $1 trillion and xAI at $250 billion, for a combined $1.25 trillion. Long-term, Musk wants to run data centers in orbit.
SpaceXAI revenue triples but losses remain steep
SpaceX's AI segment, which includes xAI and X, posted $2.56 billion in revenue during Q2 2026 alongside an operating loss of $1.26 billion. In Q1, the segment had lost $2.47 billion on just $818 million in revenue. For the first half of 2026, operating losses in the segment totaled $3.73 billion.
Most of the revenue jump didn't come from Grok. It came from new cloud contracts that lease out Colossus compute capacity. A single unnamed AI customer accounted for roughly $1.52 billion of that revenue. That customer is likely Anthropic.
Competitors spread their compute across more providers. Anthropic uses AWS Trainium, Nvidia GPUs, and Google's TPUs to train and run Claude. The company has publicly announced up to five gigawatts of additional AWS infrastructure with Trainium and Graviton systems, plus about 3.5 gigawatts of extra TPU capacity. OpenAI has announced six gigawatts of AMD capacity, at least ten gigawatts of Nvidia capacity, and around two gigawatts of AWS Trainium capacity.
— Originally published at the-decoder.com
Want this in your inbox every morning?
Daily brief at your local 8am — bilingual EN/中文, free.
More from The Decoder
See more →
An AI model programmed nonstop for 19 days on a single MirrorCode task that cost $2,600 to run
Epoch AI's MirrorCode benchmark reveals Claude Opus 4.7 as the leader with a 56% solve rate, reconstructing a 16,000-line toolkit in 14 hours. Despite this, all models tested struggle with the most complex tasks, highlighting limitations in current AI capabilities. The single task consumed $2,600 over 19 days, raising questions about cost-effectiveness in AI development.
