On 17 January 2026 Elon Musk announced that Colossus 2, xAI’s cluster near Memphis on which Grok is trained, had come online as the world’s first gigawatt cluster, and he promised 1.5 GW by April. Independent data painted a different picture. In January, satellite images showed cooling for roughly 350 MW; by mid-June, by Epoch AI’s estimate, about 950 MW, and 1.5 GW is not expected until February 2027. In the meantime xAI has become part of SpaceX, the cluster has grown to hundreds of thousands of Blackwell accelerators, and the gas turbines that power it have ended up in federal court.
What has been built, and where
Colossus 2 consists of buildings on Tulane Road in Memphis. Nearby, across the state line in Southaven, Mississippi, stand the power plant that feeds it and a new building, Macrohardrr. The original Colossus is a separate site, in another part of Memphis. At the end of September Musk spoke of 110,000 GB200 accelerators and 440,000 GB300s, and promised to add another 660,000 by the end of the year. Epoch AI, an independent research group, puts the cost of the cluster at $35.8 billion.
A gigawatt is the load of around 800,000 American households. You cannot draw that much from the grid quickly: it takes substations, power lines and approvals. So xAI powers the cluster mainly with its own generation right next to it.
The truth about the gigawatt
In January Musk spoke of 1 GW of capacity already running. Epoch AI, working from the images, saw cooling for roughly 350 MW and did not expect a gigawatt before May. In its timeline Colossus 2 had 490 MW at the start of April and 946 MW in mid-June. In SpaceX’s pre-IPO filings, the 1 GW at the end of March is given for both campuses combined, and it was calculated from the rated power draw of the installed GPUs, not from actual consumption. The 1.5 GW promised for April has not been reached: Epoch expects it by February 2027.
What powers the cluster
In the summer of 2025 xAI bought the site of a former Duke Energy gas power station in Southaven. The site has gas turbines with a combined capacity of hundreds of megawatts, and peaks are smoothed out by Tesla Megapack batteries: in 2025 alone xAI bought $506 million worth of them.
The turbines became the project’s main problem. In March 2026 Mississippi’s environmental regulator approved a permanent plant of 41 turbines with 1.2 GW of capacity; environmentalists called it rubber-stamping. Lawyers from the SELC, an environmental law group, appealed the permit, and a hearing on the appeal was set for 9 November. In April the NAACP filed a federal lawsuit over air pollution, and in June the US Justice Department stepped in on the company’s side and asked for the case to be thrown out, citing national security. In October the court refused both the Justice Department and SpaceX itself: the judge ruled that the turbines need a permit, and the lawsuit goes on. The NAACP’s request to halt the turbines until the court rules has not yet been heard. By July there were 69 temporary turbines on the site, and they are due to be removed by July 2027.
What became of xAI
On 2 February 2026 SpaceX bought xAI, and the combined company was valued at $1.25 trillion; according to CNBC, it is the biggest merger in history. In June SpaceX held a $75 billion IPO, and in July xAI was renamed SpaceXAI; in October Musk announced another rename, to SpaceXSI (SI for “super intelligence”). According to SpaceX’s filings, the cluster is being used to train Grok 5. There is no release date, but on 10 November Musk wrote that Grok 5 might surpass Claude Opus 5.5 in two to three months. The latest released model is Grok 4.7 (September). According to Musk, Grok 4.8 has finished pre-training and is now in reinforcement learning; the company rewrote its reinforcement learning and inference stack in C/C++, and a limited rollout to Grok Bot users is promised for mid-November. In May the first Colossus was leased to Anthropic.
The bottom line
Colossus 2 is the clearest example of how the AI race is now decided not by models but by megawatts. xAI built a huge cluster faster than anyone, but Musk’s promises about capacity ran months ahead of reality, and the speed came from gas turbines that the company now has to answer for in court. Energy, not chips, is becoming the main constraint on the AI industry.
For an attempt to tackle the same problem in software, see Google’s TurboQuant algorithm.
