Elon Musk has published his most detailed build schedule yet for Colossus 2, the AI supercomputer his company runs outside Memphis. He says the cluster could more than double its Nvidia chip count by December 31. If every shipment arrives on time, the site would pass 1.2 million Blackwell-generation processors, a scale no single AI cluster has reached.
In a post on X on Friday, September 25, Musk said Colossus 2 now runs 110,000 Nvidia GB200 chips and 440,000 GB300 chips, about 550,000 in total. He then laid out three more waves of 220,000 GB300s each. "Another 220k GB300 will be fully operational next week and another 220k in November," Musk wrote. The third wave comes with a caveat: "If we get lucky, yet another 220k GB300 by late December."
A Timetable Measured in Switch Ports
If all three waves land, Colossus 2 would reach about 1.21 million GB200 and GB300 processors, 1.1 million of them GB300s, according to Bloomberg and Benzinga. The same post gave an updated count for Colossus 1, the original Memphis cluster: 150,000 H100s, 50,000 H200s and 30,000 GB200s, or 230,000 chips. Added together, the Memphis-area footprint would come to roughly 1.44 million GPUs by year-end. That would clear the company's long-stated goal of 1 million GPUs in Memphis by the end of 2026 with room to spare.
The steps of 110,000 and 220,000 chips are deliberate. Musk said the 110,000-chip block reflects "the number of fiber optic cables that can be plugged into a central switch." In other words, the size of each expansion is set by the network as much as by chip supply. Every added block needs its own networking fabric, cooling loops and power. Nvidia's GB300 NVL72 rack, which links 72 Blackwell Ultra GPUs, can draw up to 142 kilowatts per rack, according to Nvidia specifications cited by Benzinga.
Several outlets now call the company SpaceXAI. SpaceX completed its acquisition of xAI in February 2026, and xAI became a wholly owned subsidiary. Grok and the Colossus clusters are now part of the SpaceX business. Colossus has also turned into a business that sells compute to other companies. Reuters reported in June that Reflection AI agreed to pay SpaceX $150 million per month for GB300 capacity at Colossus 2. Google signed a multiyear deal covering about 110,000 Nvidia GPUs, with a reported $920 million monthly price at full scale. Anthropic, meanwhile, has taken more than 300 megawatts at Colossus 1.
The Power Problem Behind the Chip Count
Power has been the hardest part of the Colossus buildout. Reuters reported in July that 59 natural-gas turbines had been installed to support Colossus 2 without federal clean-air permits. The NAACP sued xAI and its infrastructure subsidiary MZX Tech in April under the Clean Air Act over turbines at a site in Southaven, Mississippi, just across the state line from Memphis. The complaint cites 27 turbines totaling 495 megawatts. The company has said it is replacing the temporary generators with a 1.2-gigawatt permanent plant.
In June the Justice Department intervened on the company's side and asked the court to dismiss the case. It argued that Colossus 2 matters to national security. In a supporting filing, Defense Department Chief Digital and AI Officer Cameron Stanley said that if the data center lost power, "xAI would lose capacity to train and develop future, improved versions of Grok." He argued that inference capacity should be seen "not merely as commercial infrastructure, but as a long-term strategic tool," according to Utility Dive. Environmental groups called the filing an attempt to give the executive branch a "veto power" over citizen lawsuits. The NAACP has asked for an injunction to stop the turbines while the case continues.
Why It Matters
Musk's numbers matter as much for Nvidia's order book as for xAI's plans. Adding 660,000 GB300s in about three months would make xAI one of the largest single buyers of Blackwell Ultra systems this year. That supports the view that supply, not demand, is what limits Nvidia right now. Nvidia reported fiscal second-quarter revenue of $96.2 billion, up 106% year over year, and guided to $108 billion for the current quarter.
The figures also show how the frontier-AI race is now measured. Model releases still get the headlines, but the real contest is over how many chips a company can connect, cool and power at one site. Musk is betting that owning the whole stack, from turbines to switches to satellites, lets him outbuild OpenAI, Google and Anthropic. Two of those rivals now rent capacity from him. That turns Colossus from a cost center into a potential profit engine for the newly public SpaceX. The risk is that the fastest buildout in the industry depends on power arrangements that are still being fought in federal court.
What to Watch
The first test comes next week, when Musk says the fourth block of 220,000 GB300s goes fully live. The November wave will show whether the schedule holds, and Musk himself is hedging on December. Nvidia's next earnings report, due in late November, should show whether demand from xAI and others is turning into revenue at the pace investors expect. In Mississippi, a ruling on the NAACP's injunction request, or on the DOJ's motion to dismiss, could decide whether Colossus 2 keeps running on temporary turbines until the 1.2-gigawatt permanent plant is built.
“If we get lucky, yet another 220k GB300 by late December.”— Elon Musk, CEO, SpaceX (SpaceXAI)