Colossus data center buildouts: Can they ever be duplicated?

Gigantic achievements

SpaceX CEO Elon Musk's data center strategy involves available real estate, geographic constraints, facility construction and procurement complexities. Colossus 1, which occupies the former Electrolux manufacturing plant, was reportedly built as an AI model training cluster in a record 122 days, while Colossus 2, one of the largest behind-the-meter facilities in the U.S., was operational in six to eight months. The planned third data center is drawing most of its power from a recently purchased power plant site one mile across the state line in Southaven, Miss., where more than 1 gigawatt of power generation capacity is permitted by the state.

Estimates of operational AI capacity typically vary. Shareholder reports and press releases tend to conflate shells, planning assumptions and aspirations with operational capacity. SpaceX's data center buildout progress has been the subject of numerous conflicting reports, but independent satellite analysis from research institute Epoch AI and SemiAnalysis corroborated the pace and capacity of the Colossus buildouts. And rival AI labs like Google and Anthropic have seen enough progress to sign on as anchor tenants.

Yet media reports conflict on what Colossus 1 delivered during the 122-day buildout. In early September 2024, Musk posted that xAI brought a 100,000 H100 GPU training cluster online. Calling Colossus "the world's largest AI supercomputer," Nvidia reported that it took 19 days to install the first rack and commence the first training run. The local utility Memphis Light, Gas and Water indicated the site had 8 megawatts available, with plans to eventually expand to 150 megawatts of grid connectivity by 2025. Colossus 1 achieved 300 megawatts during the 122-day buildout, and Colossus 2 took six months to build a 200-megawatt capacity, while the Southaven power plant expanded from about 495 megawatts to 1.7 gigawatts, according to an August 2026 report by SemiAnalysis.

Compressing time to power

The power, GPUs, time to first training run and the ability to execute training runs at scale are all concerns for planning teams tasked with large-scale data center buildouts and operations. It's fair to say that the speed at which SpaceX and the broader constellation of Musk's enterprises operate is not ordinary. Their risk appetite might be well beyond the tolerance of even the most forward-thinking companies. Still, there are lessons other enterprises can take from SpaceX to help reduce time to power in data center buildouts.

Prioritize the interconnect, not the building

Colossus 1 was built in a refurbished Electrolux oven factory that closed in 2022 with a nominal 8 megawatts of interconnect capacity. The site was zoned for heavy industrial use and had a live substation, a 16-inch gas main, a 20-inch water main and a wastewater treatment plant next door. These amenities alone might normally take years to permit and build from scratch, making actual construction of the facility itself a secondary consideration.

Those types of unused sites exist elsewhere and have increased in value. "There are still a number of older facilities that could potentially replicate the Colossus model," said Tim Comerford, senior vice president at site selection firm Biggins Lacy Shapiro & Co. "The challenge is finding sites that have the right combination of existing infrastructure, available power and the ability to secure the necessary permits. The scarcity of sites that check all those boxes has certainly increased their value."

Coordinate grid and compute scheduling

The American Society of Civil Engineers 2025 report downgraded U.S. energy infrastructure from a C- to a D+, citing rising data center power demands as one factor. The SpaceX buildouts found ways to separate and work around permitting bottlenecks in grid connection, on-site power infrastructure, cooling and compute infrastructure.

Our grid in the U.S. is very old. Like our highways, our electric grid has a lot of potholes and congestion.

Treating each permit separately but in a coordinated fashion helped advance the Colossus aggregate schedule and gain operational experience more quickly, but it also came at a higher cost. On the other hand, treating each project as a single sequential problem rather than separate permitting issues can simplify scheduling and budgeting but increase overall build time.

"Our grid in the U.S. is very old," said Otto Lynch, vice president and head of power line systems at infrastructure software provider Bentley Systems. "Like our highways, our electric grid has a lot of 'potholes' and congestion. The only difference is that the public doesn't see them. Permitting is the main reason we don't have better grid interconnections. It takes much longer to permit power lines than to design, procure and build them, often more than 10 years. We need serious permitting reform if we are to even think about meeting the electricity demands society is rapidly moving to."

Disaggregate alternative power sources

SpaceX has demonstrated multiple strategies to skip the grid queue with behind-the-meter planning. Colossus 1 installed a large fleet of rented turbines until the grid interconnect was built. Colossus 2 was built about a mile from a purchased power plant site connected by privately owned wires. The third facility is being built on land adjacent to that power site. Private connectivity can open up opportunities to disaggregate other power sources, including solar and wind, especially if they have a large land footprint.

Rent power generators and pay the premium

SpaceX was willing to pay a premium for temporary energy capacity to accelerate the Colossus schedule, renting 14 truck-mounted generators with about 35 megawatts of capacity to augment its 8-megawatt connectivity. Larger rented gas turbines were installed later, and mobile cooling was rented on the same basis. These fleets can be deployed in weeks.

In contrast, new gas turbines can take more than two years to install, with permitting taking at least one year, according to Comerford's estimates. Renting turbines also provides more time to assess how much power is really needed down the road before investing in a permanent power generator.

Work around permitting bottlenecks

SpaceX designed a microgrid for Colossus to carry medium voltage via low-voltage transformers, bypassing the queue for larger power transformers that are more constrained. The objective was to identify components with the longest lead times and find other ways to achieve the same objective using a different architecture.

The data center reportedly imported about 2,200 metric tons of Chinese three-phase transformers in 2024, rated between 650 kilovolt-amps (kVA) and 10,000 kVA. But this kind of workaround could receive greater scrutiny in the wake of the August 2026 Presidential Executive Order tightening restrictions on foreign-made grid equipment that excludes local distribution equipment used across Colossus projects.

Harmonize power with load requirements

Gas generators of all types can take seconds to minutes to respond to changing demands, while AI training runs can change much faster, as thousands of GPUs might start and finish a run at the same time. Batteries that can adjust output in milliseconds can help buffer these different rates. Epoch AI estimates Colossus 1 and 2 at about 374 and 1,119 megawatt-hours of battery capacity, respectively. This level of backup carries important implications for large data centers and their energy needs in light of the proposed Federal Energy Regulatory Commission large load interconnect requirements by region.

Run data center construction in parallel

The Colossus 2 buildout used about 3,000 daily construction workers, considered below the number of workers needed at comparable large AI-native build sites, according to SemiAnalysis. SpaceX built Colossus with fewer work crews by running construction workflows in parallel, compressing commissioning schedules and staging pre-assembly off-site.

Some Colossus strategies are not transferable

Many of SpaceX's strategies will not easily map to other enterprises, particularly for data centers that are yet to be built.

The nearby Colossus power station was available only because an energy company overinvested during a previous economic boom that didn't go as planned. It was one of eight plants Duke Energy had sold off as distressed assets in 2004 for $475 million, with about 5.3 gigawatts of capacity among them. The turbines at the Southaven site were eventually removed. In July 2025, SpaceX bought the land for gas and electrical connectivity, along with securing the appropriate zoning and permitting rights.

SemiAnalysis estimated that SpaceX rented about two-thirds of Solaris Energy Infrastructure's mobile turbine fleet, and the Wall Street Journal reported SpaceX rented about one-fourth of the country's mobile cooling capacity. That amount of power use narrows the options for the next data center builder seeking a time-to-power advantage.

While bringing your own power can potentially provide a path around a lengthy interconnection queue, it isn't necessarily a shortcut.

Temporary exemptions for mobile power generation also gave SpaceX time to secure approvals for standard air permits with a longer lead time issued by environmental agencies. The Southaven site's units, for example, operated under temporary state approvals before Mississippi regulators granted permits for 41 permanent turbines in March 2026. Long-term operation of temporary power equipment might now conflict with national regulatory standards for classifying mobile generators for longer-duration loads.

Operating under temporary exemptions rather than standard air permits also exacerbated local opposition. "The community component cannot be overlooked, either," Comerford said. "New power generation projects can face significant local opposition, which can add additional uncertainty to the timeline. So, while bringing your own power can potentially provide a path around a lengthy interconnection queue, it isn't necessarily a shortcut. The economics and timeline have to be evaluated on a project-by-project basis."

The rapid pace of Colossus 1 was also achieved by taking advantage of a mix of GPUs that were available at the time even though they came with limited suitability for training. Colossus 2, planned more deliberately and over a longer period, was built from the ground up to avoid that problem. A faster mixed buildout makes more sense for inference loads, while a more consolidated approach is easier to manage for large training runs.

The time to operation let SpaceX carve out a new market for high-capacity training runs that could now have a limited window. SemiAnalysis estimated the annualized revenue per gigawatt at about $12 billion for neoclouds. Compare that to the Colossus tenant deals involving Anthropic at roughly $31 billion and Google at $48 billion. Both high-value contracts can be exited after a 90-day notice.

"I don't think there are too many long-term costs that are truly surprising to experienced data center developers or users," Comerford said. "They generally understand the replacement cycle for servers, which can be as short as 18 to 36 months, while most of the electrical infrastructure has a much longer useful life."

The name Colossus says as much about SpaceX's big ambitions for its data centers as it does about a cautionary tale found in D.F. Jones' trio of novels published several decades ago about a similarly named supercomputer going rogue -- hopefully, with a better fate for humankind this time.

George Lawton is a journalist based in London. Over the last 30 years, he has written more than 3,000 stories about computers, communications, knowledge management, business, health and other areas that interest him.