Memphis, Tennessee, United States
Colossus 2 is an operational AI data center in Memphis, TN, USA, with AI chips owned by SpaceXAI and used by Anthropic, Cursor, and SpaceXAI. It hosts about 1,112k H100-equivalent chips — supported by an estimated 946 MW of IT power — at an estimated capital cost of $35.8 B. We estimate that it uses Nvidia B200 and Nvidia B300 chips. The timeline of IT power and compute capacity is based on satellite imagery, our cooling equipment power model, company disclosures, and drone imagery. To power the site quickly, xAI installed natural gas turbines just across the state border in Mississippi, possibly to secure faster approval for their operation.
Satellite imagery
CURRENT COMPUTE
1,112k H100-eq
2.20e21 8-bit OP/s
CURRENT IT POWER
946 MW
PROJECTED IT POWER
946 MW
CURRENT COST
$35.8 B
PROJECTED COST
$35.8 B
Land clearing begins next to the existing building.
Battery facility connected to the building. 121 air-cooled chillers installed, which could support 200 MW. In the Duke faciltiy in nearby Mississippi, several natural gas turbines are in place but not fully connected up.
Battery facility looks complete (though more might be added). Turbines look connected up, minimal construction around them. We believe this state corresponds to the first compute cluster brought online at this site of "approximately 110,000 GB200 processors, approximately 210 megawatts of compute power", as detailed in the company S-1 filing, page 76. Based on those numbers, we believe "compute power" corresponds to GPU server power. Earlier tweets from Elon Musk also support this timing, and it's consistent with the estimated cooling capacity based on the satellite image.
Since October, an array of dry cooling air-cooled condensers has been added (the array of large fans to the left of the building), and appears almost complete. Two more similar arrays appear to be under construction. More air-cooled chillers, batteries, and gas turbines have also been installed.
The first array of dry cooling air-cooled condensers for Building 1 has been completed and is operational. A second array, with larger fans, is about halfway complete (located to the south of the first array). More chillers have also been added. For Building 2, an array of dry cooling air-cooled condensers of the same type as Building 1's first array is beginning construction. We believe this corresponds to the "second cluster of 110,000 GB300 processors and 220 megawatts of compute power" from the company S-1 filing, page 76, with compute power corresponding to GPU server power. This is also consistent with our estimates based on cooling equipment.
The site is estimated to have at least ~830 MW of GPU server power, with at least "at least 220,000 additional GB300 processors and over 400 additional megawatts of compute power", corresponding to the "next phase of expansion" found in the following company S-1 filing, page 76. The timing is based on the air-cooled condensers needing to be complete to handle this power capacity. The second array at MACROHARD and the first array at MACROHARDRR, although both projected to be completed prior, are not enough and more cooling capacity is needed. We estimate this expansion to be completed 90 days after April 6th. On one hand, SpaceXAI has progressively added more capacity in less time with their Colossus projects. On the other hand, 400 MW is more capacity than the previous expansions of 210 MW and 220 MW. We also believe that MACROHARDRR will be partly operational with this expansion, because its air-cooled condensor array is expected to be completed by then and fitting the entire expansion into only MACROHARD would result in extremely high levels of power density.
Read more about our methodology.