TL;DR
SpaceXAI is racing toward 1.44 million GPUs at its Colossus facilities, backed by a self-built 1.2-gigawatt power plant, making it one of the densest AI compute concentrations on Earth despite being only three years old.
What happened
- 220,000 Nvidia GB300 GPUs were set to come online within days of the announcement, with two more tranches of 220,000 each arriving in November and December 2025.
- Combined with existing hardware, the additions bring SpaceXAI to 1.1 million GB300 GPUs and a total of 1,440,000 GPUs across both Colossus sites.
- Colossus 1 (Memphis) runs a mix of H100, H200, and GB200 chips and is being rented to Anthropic for inference because its mixed architecture creates training bottlenecks for Grok.
- Colossus 2 runs exclusively Blackwell GPUs, eliminating architectural bottlenecks for Grok training.
- SpaceXAI is building a 1.2-gigawatt dedicated power plant to replace unpermitted gas turbines that triggered a neighborhood lawsuit, with turbine removal phased over one year.
Why it matters
- Reaching 1.44 million GPUs in roughly three years compresses a timeline that rivals like OpenAI (10 years old) and Anthropic (6 years old) took far longer to approach.
- The power-plant-as-infrastructure model signals that hyperscale AI compute now requires sovereign energy, not just grid access, raising the capital bar for competitors.
- Musk has stated a sevenfold capacity expansion by 2027 and a target of 50 million H100-equivalent GPUs by 2030, numbers that would dwarf current hyperscaler deployments.
- Renting Colossus 1 to Anthropic for inference is a notable competitive irony: a direct rival is now dependent on Musk-controlled infrastructure.
- Broadcom disclosed in 2024 that three unnamed hyperscale customers are targeting the 1-million-GPU threshold by 2027, meaning the race to this milestone is crowded and accelerating.
What to watch next
- Whether the December 220,000-GPU tranche actually comes online on schedule, which Musk himself flagged as contingent on luck, will test supply chain and power readiness.
- The 1.2-GW power plant completion timeline is the critical path: delays would stall full utilization and could reignite regulatory friction with local authorities.
- Grok benchmark results over the next two quarters will reveal whether raw compute density at this scale translates into measurable model quality gains versus OpenAI and Anthropic.
Originally published on Present of AI, a daily source-linked AI news timeline. Read the full timeline or browse the open dataset.