Musk's AI Computing Power Surges! SpaceXAI to Add 660,000 NVIDIA GB300 GPUs This Year, Nearing 1.44 Million Total

Elon Musk's artificial intelligence (AI) computing power expansion continues. Musk recently announced on X that SpaceX (SPCX-US) AI division SpaceXAI will add 660,000 NVIDIA (NVDA-US) GB300 GPUs this year, bringing total GPU holdings to nearly 1.44 million, surpassing the "million GPU" target he set two years ago.

According to data released by Musk, SpaceXAI's Colossus 1 currently has 150,000 H100s, 50,000 H200s, and 30,000 GB200s; Colossus 2 has 110,000 GB200s and 440,000 GB300s.

Musk said another 220,000 GB300s will be fully operational next week, another 220,000 are expected to come online in November, and if progress goes smoothly, another 220,000 GB300s will be added by the end of December.

At that point, SpaceXAI's total GB300 count will reach 1.1 million, and total GPU holdings will reach 1.44 million.

Colossus 2 Fully Adopts Blackwell Architecture

With the continuous expansion of computing power, SpaceXAI has built a large-scale AI computing cluster just three years after its founding. In comparison, upstream was founded about six years ago, and upstream about ten years ago.

Musk revealed that Colossus 1 mixes Hopper and Blackwell architectures, leading to poor training efficiency for the Grok model, so it has been rented to Anthropic for inference tasks.

Colossus 2, on the other hand, fully adopts the Blackwell architecture. Musk said this configuration avoids bottlenecks during the training process.

However, for SpaceXAI, continuously acquiring GPUs is not the biggest constraint in expanding data centers; power supply is another key issue.

SpaceXAI is currently building a 1.2 gigawatt (GW) self-owned power plant to supply electricity for the large-scale AI computing cluster. By building its own power generation facilities, the company can avoid the lengthy approval process required for grid connection.

But this approach has also sparked local controversy. SpaceXAI faces a lawsuit for using unlicensed gas turbine power generation equipment and has promised to gradually dismantle these temporary power generation facilities within one year while waiting for its own power plant to be completed and operational.

One Million GPUs Is Not the End

Large-scale deployment of one million GPUs is not the only case in the industry. Broadcom (AVGO-US) revealed in 2024 that three of its hyperscale customers plan to achieve similar scale by 2027.

But for Musk, one million GPUs is just the starting point for even larger expansion. He said SpaceXAI plans to increase data center capacity sevenfold by 2027 and has set a long-term goal of achieving 50 million H100-equivalent GPUs by 2030.

SpaceXAI's computing power layout is not limited to ground-based data centers. SpaceX also plans to launch an orbital data center system consisting of one million satellites, though NVIDIA CEO Jensen Huang said this is still "a dream" at present.

SpaceXAI's rapid expansion of computing power scale shows that AI infrastructure competition is no longer just about GPU procurement. From building self-owned power generation facilities, renting out existing computing power, to planning orbital data centers, Musk is simultaneously expanding his layout in energy and data center infrastructure.