Colossus 2 Is Adding GPUs Faster Than It Is Adding Power
Musk says xAI's Memphis cluster could pass 1.2 million Nvidia chips by year-end, a figure that implies more than double the output of the permanent power plant being built to run it.
Elon Musk has put a timetable on the biggest single AI cluster anyone has publicly described. Replying to a user on X on September 25, he laid out the chip counts for both Memphis supercomputers and then the schedule for the second one: "Colossus 2 is 110k GB200 and 440k GB300. Another 220k GB300 will be fully operational next week and another 220k in November. If we get lucky, yet another 220k GB300 by late December."
Add it up and Colossus 2 goes from 550,000 Nvidia Blackwell-class processors today to about 1.21 million by the end of the year, 1.1 million of them GB300s. Bloomberg framed it as a plan that "may more than double" the site's Nvidia chip count before 2027. For context, Musk put Colossus 1 at 150,000 H100s, 50,000 H200s and 30,000 GB200s.
The chip count is the headline. The more useful number is the power needed to run it, and on that Musk said nothing.
Why the steps are 220,000 chips
The size of each tranche is not arbitrary. Per AI Weekly, Musk said the 110,000-chip unit reflects "the number of fiber optic cables that can be plugged into a central switch." Each 220,000-GPU wave is two of those units.
That detail matters because it shows where the limits are. At this scale, getting Nvidia allocation is not the only constraint. The network fabric decides how many GPUs can be tied into one training job, and Colossus 2 is growing in blocks sized to what the switching layer can handle. The other constraint is electricity.
The power math
Nvidia's GB300 NVL72 racks each hold 72 GPUs. Integrator specs, including Lenovo's product guide, put a fully loaded rack at roughly 132 to 142 kilowatts in normal operation, with peaks near 155 kilowatts. At 135 kilowatts per rack, 1.1 million GB300s works out to about 15,300 racks and just over 2 gigawatts of IT load. That covers only the GB300 racks. It leaves out the 110,000 GB200s, networking, storage and the cooling plant that removes all that heat.
Now compare the supply side. Colossus 2 sits in the Whitehaven area of Memphis, and its dedicated generation is across the state line in Southaven, Mississippi. In March, the Mississippi Department of Environmental Quality approved an air permit for a permanent simple-cycle gas plant there. SpaceXAI has described it as "the 1.2 GW permanent power plant currently under construction," made up of 41 turbines, per Gizmodo.
In the meantime the site has run on mobile turbines, and that is where it has taken legal fire. Reuters reported in July that 59 natural-gas turbines had operated without federal clean-air permits. The NAACP and the Southern Environmental Law Center sued in April, and the Justice Department later asked a court to dismiss a Clean Air Act case, arguing the turbines serve national security. Under an agreement with Mississippi regulators reported by Gizmodo on August 1, SpaceX must remove 69 temporary mobile turbines by July 2027.
A 1.2-gigawatt plant cannot run roughly 2 gigawatts of GPU racks plus cooling. Either the rest comes from the grid through Memphis Light, Gas and Water and TVA, from more on-site generation not yet disclosed, or the December tranche depends on power that is not in place. Investing.com noted that Musk "did not address" whether the facility has enough power for the new chips. AI Weekly flagged the same gap: nobody has said whether permanent power will be running before the late-December batch arrives.
Compute as a product, not just a lab
The ramp also matters because the Memphis clusters are no longer only for Grok. SpaceX acquired xAI in an all-stock deal in February and folded it in as SpaceXAI in May. Since then, Colossus has become a commercial compute business.
Anthropic agreed in May to rent all of Colossus 1, more than 220,000 Nvidia GPUs and over 300 megawatts, Yahoo Finance reported. TechCrunch put that deal at $1.25 billion a month. In June, an SEC filing showed Google will pay SpaceX $920 million a month from October 2026 through June 2029 for about 110,000 Nvidia GPUs plus CPUs and memory. Google called it "a short-term, timely agreement to ensure we have bridge capacity" for Gemini Enterprise. The filing gives Google the right to terminate if SpaceX fails to deliver the committed GPUs by September 30, 2026, after a one-month grace period. The filing did not name the facility, and TechCrunch noted SpaceX had said Colossus 2 was reserved for xAI.
That context changes how to read "if we get lucky." Every 220,000-GPU block that comes online on schedule is either training capacity for Grok or capacity SpaceX can rent at hyperscaler prices. Every block that waits on power or networking is idle inventory worth billions of dollars.
What to watch
The first test comes this week: 220,000 GB300s "fully operational," which would take the site to roughly 770,000 processors. The November tranche will show whether the ramp holds. The December batch, the one Musk himself hedged, is the real stress test of the power plan.
Three things will say whether 1.21 million is real. First, any disclosure of added grid supply or new permits in Southaven or Shelby County beyond the 41-turbine plant. Second, the timing of the mobile turbine removals, which should speed up if permanent capacity is keeping pace and stall if it is not. Third, whether SpaceXAI adds more rental customers, which would suggest the capacity is running rather than waiting for power.
Why it matters
The frontier AI race has long been measured in GPUs because that is what companies announce. Colossus 2 shows why that measure is not enough. The chips are arriving in 220,000-unit blocks. The power is arriving in permits, turbines and interconnects that move on regulators' timelines, not Nvidia's. If Musk gets to 1.21 million by New Year's, it will be because SpaceXAI solved a power problem it has not yet explained. If it falls short, the reason will almost certainly be power, not chips.