Colossus capacity and the gas turbine problem
xAI expanded the Memphis supercluster to 230,000 GPUs in Colossus 1. The company now deploys Colossus 2 with 550,000 Nvidia GB200 and GB300 units to train Grok 5. This expansion moves the Memphis site closer to the 1 million GPU goal. The project requires massive electricity. While Colossus 1 draws 150 MW from the Memphis Light, Gas, and Water grid, it also uses 150 MW from Megapack Batteries to maintain power during outages or periods of increased demand. To manage the load, xAI uses 15 gas turbine generators. These 20-year-old turbines emit formaldehyde at levels that exceed the EPA 10-ton annual cap for a single source. This pollution makes the Memphis expansion a health liability for the 22,000 people living within five miles of the facility. The facility relies on natural gas turbine generators to meet its electricity needs. The company purchased the 785,000-square-foot Colossus 1 building and 217 acres of land for $185 million. For Colossus 2, xAI purchased 186.13 acres for $79.9 million. The company also requested a second 150 megawatts of power that remains under review. The Tennessee Valley Authority provides 60% carbon-free electricity to the region. Will the Memphis grid handle a million GPUs?
The merchant utility model
xAI is a SpaceX subsidiary after the February 2, 2026 merger. This relationship provides the capital for aggressive infrastructure growth. The company closed a $20 billion Series E funding round with Nvidia and Cisco, which values the startup at approximately $230 billion. The original Colossus build used 100,000 Nvidia H100 GPUs and took only 122 days to complete.
| Component | Specification |
|---|---|
| Colossus 1 GPU count | 230,000 |
| Colossus 2 GPU count | 550,000 |
| Colossus 1 power | 150 MW grid / 150 MW battery |
| Colossus 2 target | 550,000 Nvidia GB200/GB300 |
| Five-year GPU goal | 50,000,000 H100-equivalent |
The company does more than train Grok. It is a merchant utility. Anthropic pays $1.25 billion per month to lease 325,000 GPUs across Colossus 1 and 2. Google signed a deal for 110,000 GPUs at $920 million per month starting in October 2026. You should notice how xAI profits from its rivals. SpaceX also acquired Cursor in Q3 2026 for $60 billion. Reflection AI pays $150 million per month for access to GB300 chips. The company targets a deployment of 50 million units of H100-equivalent AI compute within five years. This includes Blackwell architecture chips like the GB200 and GB300. xAI also negotiates for billions in additional debt and equity to support this rapid scale.
The Stargate rivalry
OpenAI leads the Stargate project, a joint venture with SoftBank, Oracle, and MGX. The project targets $500 billion in infrastructure investment by 2029. In Abilene, Texas, four buildings operate with 450,000 Nvidia GB200 GPUs, while four more buildings should come online around November 2026. This buildout creates a direct confrontation with the Memphis supercluster. OpenAI and SoftBank have already invested $41 billion into this venture. The scale of these two projects creates a massive gap in how companies approach hardware. xAI builds in existing industrial shells to gain speed, whereas Stargate relies on massive multi-state data center campuses. The Stargate project involves technology partners including Arm, Microsoft, Nvidia, and Oracle. OpenAI also targets expansion into countries like the United Kingdom and Japan. This planned Stargate capacity reaches nearly 7 gigawatts across various international sites. The UAE Stargate project expects 200 MW to go live in 2026, supported by a $1.4 trillion investment from the UAE. The project targets a globally distributed AI network through its OpenAI for Countries initiative.
