Colossus: How xAI Built an AI Factory in 122 Days, and What Speed Cost
Most large data centers take 18 to 24 months to build. xAI turned an empty Electrolux factory in South Memphis into a cluster of about 100,000 H100 GPUs in 122 days, then built a second campus, Colossus 2, in phases of 91 and 64 days. This first case study looks at how the speed was achieved, why gas turbines became the story, and why three honest ways of counting its size give three different answers.
This series visits the world's largest AI factories one at a time. The other "Build the Machine" series explain the layers: power, the factory, bring-up and the rest. Here we look at how one organisation put those layers together, and which trade-offs it chose. Each case study uses the same template: a case file, a build timeline, the bet, power, community and lessons. Part 10 will put all the timelines on one chart.
1. Case file
The S-1 says the next Colossus 2 phase will add at least 220,000 more GB300s and over 400 MW once fully operational. In September 2026 Musk posted that Colossus 1 holds 150k H100, 50k H200 and 30k GB200, and Colossus 2 holds 110k GB200 and 440k GB300, with more GB300 waves aimed at about 1.44 million GPUs by December. Those are his statements, not filing data.
2. Build timeline
Colossus, 2024 โ 2026
Each lane is one strand of the story. Solid dots are confirmed by a filing or a regulator, hollow dots come from trackers or the press, and โ ๏ธ marks claims we could not confirm. Click a dot for its source. Dashed items are planned.
3. The bet: speed beats everything
Every case study in this series has one bet, a design choice that makes the site different. For Colossus it is speed. Typical hyperscale builds start with land, then grid studies, permits, a purpose-built shell and a power contract, and the queue for a large grid connection can run for years. xAI skipped as much of that as it could:
- Existing building. The S-1 says the first cluster came online by repurposing an existing factory shell, the former Electrolux plant, instead of building from scratch.
- Power before the grid was ready. The site started with only a small grid connection, so mobile gas turbines covered the gap (section 5).
- Liquid cooling and batteries brought in as kits. Racks, chillers and battery storage were installed in parallel with the building work, not in sequence. Secondary sources say Tesla Megapacks buffer the load swings of training โ ๏ธ.
- One customer. xAI built for its own models, so there was no tenant to negotiate with and no multi-tenant design to satisfy.
4. The build race
Days from start to compute
xAI's reported build phases, compared with the 18โ24 months a large data center usually takes. The "typical" bar is a rule of thumb, not a measured average.
Sources: SpaceX S-1 (122, 91 and 64 days); "92 days" for doubling Colossus 1 to 200,000 GPUs is xAI's own claim reported secondhand โ ๏ธ.
The shrinking numbers matter as much as the first one. The second Colossus 2 phase took 64 days for about the same number of GPUs as the 91-day first phase, which suggests the bottleneck moved from the building to the supply of chips and power. Speed of this kind is a repeatable process, not a one-off sprint, and that is the real competitive asset.
5. Power: the real story
Chips can be bought; power cannot be bought quickly. When Colossus 1 switched on, the grid connection was tiny. DCD reported that the facility launched with an 8 MW grid connection plus gas turbines. Memphis Light, Gas and Water (MLGW) said in May 2025 that xAI was receiving 150 MW and had asked for another 150 MW. Turbines filled the gap.
Where the power comes from
Approximate megawatts by source at each stage. Toggle the site. Bars show reported capacity, not actual output on any given day.
For Colossus 2, xAI moved generation across the state line to Southaven, Mississippi, about six miles from the Tulane Road data center. Reports in 2026 described dozens of gas turbines there. The Mississippi Department of Environmental Quality (MDEQ) let xAI run turbines classed as "temporary-mobile" without an air permit for up to a year, then granted an air permit in March 2026. Local TV reported that permit covered 27 turbines; one secondary source describes 41 turbines rated at 1.2 GW โ ๏ธ. Tom's Hardware reported that the company is building a 1.2 GW power plant to bring its systems fully online.
Turbines: permitted vs counted
Counts are snapshots from different organisations on different dates, so they are not strictly comparable.
Gas turbines are fast to bring in but inefficient and polluting compared with a large combined-cycle plant. They emit nitrogen oxides and formaldehyde close to where people live, which is why the turbine count became the central fight. Related reading: private power for AI covers why labs are building their own generation.
6. Three ways to count a gigawatt
Claims that Colossus 2 is "the first gigawatt AI data center" depend on what you count. There are at least three honest measures, and they give different answers.
How big is Colossus 2?
Pick a way of counting.
7. Site map
Memphis and Southaven
A simplified diagram, not to exact scale. Hover or tap a site.
8. Community and law
Colossus 1 sits in South Memphis near Boxtown, a predominantly Black neighbourhood that already lives with heavy industry. In April 2025 SELC said aerial imagery showed "a whopping 35 gas turbines" on site, while xAI had applied for permits for 15 to run temporarily. In July 2025 the Shelby County Health Department approved a permit for 15 turbines, generating up to 247.2 MW, valid until 1 January 2027. Environmental groups argued that the extra turbines had run without permits; xAI's position, as reported, was that the units were temporary.
9. Lessons for the series
We will score every site on the same four questions: how fast it was built, where the power comes from, how big it really is, and how it treated its neighbours. Next we look at OpenAI's Stargate campus in Abilene, Texas, which bet on financing and partners rather than on doing everything itself.
10. Reading the numbers
- From the S-1
- 122 days and ~100,000 H100s / ~130 MW for the first cluster; 110,000 GB200 / ~210 MW in 91 days and 110,000 GB300 / 220 MW in 64 days for Colossus 2; at least 220,000 more GB300 and over 400 MW planned.
- From regulators
- 15 turbines / 247.2 MW permitted at Colossus 1 until 1 Jan 2027 (Shelby County, via DCD and Fox13); MDEQ air permit March 2026 (count reported as 27, โ ๏ธ one source says 41 at 1.2 GW).
- Estimates
- ~946 MW current IT power for Colossus 2 is Epoch AI's estimate. ~350 MW of cooling seen in January satellite imagery was reported by DCD โ ๏ธ. 1.7 GW nameplate on 30 Sep 2026 comes from a secondary reading of SpaceX's disclosures; we did not find it in the S-1 text โ ๏ธ.
- Statements
- GPU totals and the 1.44 million target are Musk's posts on X. "92 days" to double Colossus 1 and the July vs September 2024 go-live date vary between sources โ ๏ธ.
- Illustrations
- The power bars round figures and the site map is a diagram, not a survey.
- Primary
- SpaceX Form S-1 (SEC EDGAR)
- Epoch AI: Frontier data centers
- Elon Musk on X: Colossus GPU counts (Sep 2026)
- Power and permits
- DCD: xAI granted permits for 15 gas turbines
- CNBC: SELC says xAI using more turbines than permitted
- Inside Climate News: turbines without permits, residents claim
- DCD: turbines at Southaven for Colossus 2
- Tom's Hardware: 1.44 million GPUs and a 1.2 GW plant
- Global Energy Monitor: Colossus 2 power station
- Community and law
- WLBT: NAACP sues xAI over Southaven turbines
- Mississippi Free Press: motion to shut turbines down
- SELC: xAI built an illegal power plant
- Earthjustice: a community takes Colossus to court
- Scientific American: Colossus backlash
- Wikipedia: Colossus (supercomputer)