SpaceXAI is approaching a milestone that seemed impossible just years ago: operating 1.44 million graphics processing units (GPUs) across its Colossus supercomputer clusters. But as the company adds 660,000 new GPUs this year alone, the challenge isn't acquiring cutting-edge chips anymore. It's generating enough power to run them.

How Much Power Does a Million-GPU Data Center Actually Need?

The scale of SpaceXAI's ambition is staggering. The company announced that 220,000 Nvidia GB300 GPUs will be operational by late September 2026, with another 220,000 coming online in November and potentially another 220,000 by late December, contingent on favorable circumstances. These additions build on existing infrastructure: Colossus 1 already runs 150,000 H100 GPUs, 50,000 H200 GPUs, and 30,000 GB200 GPUs, while Colossus 2 operates 110,000 GB200 GPUs and 440,000 GB300 GPUs.

To power this sprawling operation, SpaceXAI is constructing a dedicated 1.2-gigawatt power plant, a facility that would rank among the largest power generators in many U.S. states. For context, a gigawatt can supply electricity to roughly 750,000 homes. This isn't a minor infrastructure project; it's a fundamental requirement for the company's AI ambitions.

Why Is Power the New Bottleneck in AI Data Centers?

For years, the limiting factor in building massive AI systems was GPU availability. Elon Musk famously lobbied Nvidia's leadership for access to the latest chips. But the industry has shifted. Getting GPUs is still competitive, yet securing reliable, abundant electricity has become equally critical. SpaceXAI's decision to build its own power plant reflects a broader reality: hyperscale AI infrastructure requires energy independence.

The company's approach has already generated friction. SpaceXAI faced a lawsuit from its surrounding community over unpermitted gas turbines that were temporarily powering the facility. The company has committed to removing these turbines over the course of one year as its permanent 1.2-gigawatt power plant comes online. This dispute underscores the tension between rapid AI expansion and local infrastructure constraints.

What Makes SpaceXAI's GPU Strategy Different?

SpaceXAI's infrastructure is organized into two distinct clusters, each optimized for different workloads. Colossus 1 combines older Hopper and Blackwell generation GPUs, making it less efficient for training Grok, the company's AI model. Rather than leave the hardware idle, SpaceXAI rented Colossus 1 to Anthropic, a competing AI company, for inference tasks, where the mixed GPU generations perform adequately. Colossus 2, by contrast, uses exclusively Blackwell GPUs, eliminating bottlenecks and ensuring consistent performance for training operations.

This pragmatic approach reveals how the AI industry is maturing. Companies are no longer hoarding every available GPU; they're optimizing utilization and monetizing excess capacity. SpaceXAI's rental arrangement with Anthropic demonstrates that even competitors can find mutually beneficial arrangements when infrastructure is at stake.

Steps to Understanding Data Center Power Requirements

• GPU Power Consumption: Modern high-end GPUs like the Nvidia GB300 consume significant electricity during training, with a single GPU drawing hundreds of watts under full load, multiplied across hundreds of thousands of units.
• Cooling and Infrastructure Overhead: Beyond the GPUs themselves, data centers require substantial power for cooling systems, power distribution, and networking equipment, often adding 30 to 50 percent to total energy demands.
• Power Plant Capacity Planning: A 1.2-gigawatt facility must account for peak demand, redundancy, and future expansion, requiring careful engineering to avoid bottlenecks as GPU counts grow.

What Are SpaceXAI's Longer-Term Ambitions?

The 1.44 million GPU milestone is just the beginning. Musk stated that SpaceXAI plans to expand its data center capacity sevenfold by 2027, and he's targeting 50 million H100-equivalent GPUs by 2030. These projections suggest that power infrastructure will remain a critical constraint for years to come. The company is even exploring unconventional approaches, including plans to launch an Orbital Data Center System using a million satellites, though Nvidia CEO Jensen Huang characterized this concept as a "dream" for the present.

SpaceXAI isn't alone in pursuing million-GPU clusters. Broadcom reported in 2024 that three hyperscale customers are targeting this milestone by 2027, though the company declined to identify them. This suggests that power infrastructure challenges will become industry-wide concerns as AI training scales globally.

The shift from GPU scarcity to power scarcity marks a turning point in AI infrastructure development. Companies that can secure reliable, affordable electricity will have a decisive advantage in the race to build the largest and most capable AI systems. For SpaceXAI, the 1.2-gigawatt power plant isn't just a supporting detail; it's the foundation upon which its AI ambitions rest.