Elon Musk is getting close to the goal he set almost two years ago: building an AI supercomputer with over a million GPUs. His company SpaceXAI is ramping up its Colossus cluster and expects to hit around 1.44 million GPUs in operation by the end of this year.
Colossus 1 vs Colossus 2
According to Musk, 220,000 Nvidia GB300 GPUs will be live by next week, with another 220,000 in November and a further 220,000 by late December “if we get lucky.”
Current breakdown:
- Colossus 2: 110,000 GB200 + 440,000 GB300 GPUs
- Colossus 1: 150,000 H100, 50,000 H200, 30,000 GB200 GPUs
If all goes to plan, GB300 units will reach roughly 1.1 million, and total GPUs across both sites will be about 1,440,000. That’s notable for a company that’s only three years old, competing with Anthropic (~6 years) and OpenAI (~10 years).
Architecture matters: Colossus 1 mixes Hopper and Blackwell GPUs and turned out to be inefficient for training the Grok model. Musk has instead rented it to Anthropic for inference workloads. Colossus 2 is built purely on Blackwell to avoid training bottlenecks.
Power is the real bottleneck
Getting GPUs from Jensen Huang’s Nvidia was only half the battle. The bigger constraint for hyperscale data centers today is power delivery. SpaceXAI tried to solve this by bringing its own generation capacity, using unpermitted gas turbines, which triggered local backlash and a lawsuit. The company has pledged to remove the turbines over a year while it brings a dedicated 1.2 GW power plant online.
Musk says the 1‑million‑GPU milestone is just a starting point. By 2027, he wants to 7× SpaceXAI’s data center capacity, and by 2030 he’s targeting 50 million H100‑equivalent GPUs. Longer term, SpaceX also talks about an Orbital Data Center System with one million satellites, even if Nvidia’s Huang currently calls that a “dream.”
Source: Tom's Hardware
