The estimated advantage is 9.1%, although the figures reflect theoretical computing capacity rather than verified GPU counts or actual training performance.
Colossus 2 vs. Stargate: Technical Specifications
| Specification | Colossus 2 | Stargate Abilene |
| Location | Memphis, Tennessee | Abilene, Texas |
| Compute capacity | 1.112M H100e | 1.019M H100e |
| 8-bit performance | 2.20 × 10²¹ OPS | 2.02 × 10²¹ OPS |
| IT power capacity | 946 MW | 824 MW |
| Estimated capital cost | $35.8B | $31.2B |
| GPU architecture | Nvidia Blackwell | Nvidia Blackwell |
| Status | Operational | Operational |
Source: Epoch AI estimates, November 2026.
Colossus 1 vs. Colossus 2
| Metric | Colossus 1 | Colossus 2 |
| Compute capacity | 275,796 H100e | 1,111,673 H100e |
| IT power | 340 MW | 946 MW |
| Estimated capital cost | $12.9B | $35.8B |
| GPU generations | Hopper + Blackwell | Blackwell |
Colossus 2 delivers approximately 4× the estimated compute capacity of Colossus 1, with 2.8× the IT power capacity.
Blackwell Expansion Plans
Epoch AI projects that Colossus 2 could reach the following specifications by Q1 2027:
- Compute: 1.824 million H100-equivalents (+64%).
- IT power: 1,531 MW (+62%).
- Estimated capital cost: $58 billion.
- Hardware: Additional Nvidia Blackwell-generation systems.
These figures are projections, not confirmed operational capacity.
Infrastructure Requirements
Colossus 2's expansion depends on three critical resources:
- Electricity: Natural gas generation and grid infrastructure supporting gigawatt-scale computing.
- Cooling: Industrial chillers and condenser systems for high-density GPU clusters.
- Hardware: Blackwell accelerator deliveries, networking equipment, and server deployment.
The facility's estimated 946 MW IT capacity does not imply continuous operation at that power level.
Key Takeaway
Colossus 2 leads Stargate Abilene by approximately 93,000 H100-equivalents in Epoch AI's estimates.
However, the 9.1% difference is within the uncertainty of the estimates. The ranking compares individual facilities, not SpaceXAI's and OpenAI's combined infrastructure.
