A gigawatt behind the count
Musk's 25 September post puts up to 660,000 more GB300s on the Colossus calendar by late December, and a path toward about 1.44 million GPUs if the last wave lands. The chips are the number people will repeat. The schedule that decides whether they train anything is a 1.2-gigawatt gas plant still replacing temporary turbines that Southaven has to live with into July 2027.
Situation
01Elon Musk used X on 25 September to publish a Colossus inventory and a year-end schedule. Colossus 2, he said, is already 110,000 Nvidia GB200s and 440,000 GB300s. Three more blocks of 220,000 GB300s would follow: one fully operational next week, one in November, and one by late December "if we get lucky." Added to Colossus 1's older mix of H100s, H200s, and GB200s, the lucky path is about 1.44 million GPUs. Tom's Hardware's read of the same post is the line that belongs on the desk. Getting the chips was the old ask. SpaceXAI is building a 1.2-gigawatt natural gas plant to bring the systems fully online, while temporary turbines in Southaven, Mississippi, across the state line from the Memphis Colossus halls, stay on a removal clock that runs to July 2027.
Facts
02- 01On 25 September 2026 Musk posted on X that Colossus 1 is 150,000 H100s, 50,000 H200s, and 30,000 GB200s, and that Colossus 2 is 110,000 GB200s and 440,000 GB300s. He said another 220,000 GB300s would be fully operational next week, another 220,000 in November, and another 220,000 by late December "if we get lucky."
- 02If that December block lands, Tom's Hardware puts SpaceXAI at 1,440,000 GPUs in operation, including 1.1 million GB300s. AI Industry Today reaches the same combined figure: about 1.21 million accelerators at Colossus 2 plus about 230,000 at Colossus 1.
- 03The 440,000 GB300 count is Musk's stated figure. Epoch AI, as of 23 September, still estimated Colossus 2 at about 110,000 GB200s and 330,000 GB300s on roughly 946 megawatts of IT power. A 2026 SpaceX securities filing described the first Colossus 2 cluster as about 110,000 GB200s and 210 megawatts online in 91 days, a second cluster of 110,000 GB300s and about 220 megawatts in 64 days, and a next phase of at least 220,000 more GB300s and more than 400 megawatts.
- 04Tom's Hardware reports that obtaining the GPUs is no longer the hardest part of the build-out, and that SpaceXAI is building a 1.2-gigawatt power plant to bring the systems fully online after controversy and a lawsuit over unpermitted gas turbines.
- 05SpaceXAI agreed with the Mississippi Department of Environmental Quality to remove all 69 temporary mobile turbines at its Southaven power site, with removals beginning as early as August 2026 and the last units allowed to remain until July 2027. The replacement is a permanent 1.2-gigawatt plant of 41 turbines under a Clean Air Act permit granted in March 2026. The Southaven site sits across the state line from Colossus 2 in Memphis.
- 06July 2027 is an extension. Mississippi regulators had set 31 December 2026 as the deactivation deadline. The agreed order cites supply-chain delays on some of the 41 permanent turbines and lets 13 specific temporary turbines run past the original one-year limit.
- 07By July, Mississippi regulators counted 69 temporary turbines at Southaven, about 1,775 megawatts, more than double the number the company had acknowledged at the start of the year. The permitted permanent plant is about 1,200 megawatts. At least 27 of those permanent turbines were under construction as of July. The NAACP has sued over the unpermitted fleet, and residents have a noise case in court.
Assessment
03We assess, with moderate confidence, that the 25 September post is a chip schedule published into a power deficit. Musk's own qualifier, "if we get lucky," applies to the third GPU wave. It should also apply to watts. A prior filing tied one 220,000-GB300 phase to more than 400 megawatts of compute. Three phases on that ratio are a gigawatt-class addition on top of a Colossus 2 load that independent tracking already put near 946 megawatts, before the latest GB300 claim. The permitted plant is 1.2 gigawatts, and it is the replacement for a temporary fleet sized near 1,775 megawatts, not a spare socket for 660,000 new chips.
The count itself is not yet an audit. Epoch's 23 September snapshot sat about 110,000 GB300s short of the figure Musk posted two days later. "Fully operational" is his verb. It is not a utility interconnection record, and it is not proof the racks are in a training job. Treat 440,000, and each 220,000 block after it, as a stated plan until a tracker, a filing, or a commissioning note catches up.
Operators who still model frontier capacity as a Nvidia allocation problem are a quarter late. The allocation was the 2024-2025 fight. The 2026 fight at this campus is whether generation, permits, and a neighborhood clock can energize racks that have already been promised. Bringing your own power is the strategy. The cost is already on the docket: a Clean Air Act suit, a July 2027 removal order born from a missed December deadline, and a permanent plant that is still catching the temporary machines it is supposed to retire.
Implications
04- GPU purchase orders are no longer a sufficient leading indicator for Colossus capacity. Energization dates are.
- The 1.2-gigawatt on-site plant is the product strategy for this campus, and it arrives with a permit fight, a supply-chain slip, and a residential noise case.
- Anyone leasing Colossus capacity inherits the Southaven clock. A delayed turbine retirement is a serving risk.
- On July figures, the temporary fleet already outscales the permitted permanent plant. More GPUs without more permitted megawatts means idle silicon or a longer temporary-turbine fight.
- Rival capacity maps should discount the December block, and discount any block that does not come with a named megawatt.
Recommended action
05If a capacity plan, investment memo, or competitive map treats Colossus GPU totals as available training or inference, split the number this week into three buckets: stated and tracker-confirmed, stated and not yet confirmed, and conditional on "if we get lucky." Put a power line next to each bucket. Do not brief 1.44 million as online. Brief a 550,000-class Blackwell cluster whose next 220,000-GPU step is a 25 September claim that the block would be fully operational the following week, gated by a 1.2-gigawatt plant that is still replacing turbines allowed to run until July 2027.
What to watch
06- Independent counts, Epoch or a filing, that confirm or cut the 440,000 GB300 figure and the "next week" 220,000 block.
- A commissioning date for the 41 permitted turbines, set against the July 2027 temporary-removal deadline.
- The NAACP suit, MDEQ enforcement, and any further slip of the 13 extended turbines.
- Language from Musk or SpaceXAI that separates installed, powered, and in a workload.
- Any new GPU block posted without a megawatt figure beside it.
Sources
07PRIORITY / SPACEXAI / Colossus 2: 550k Blackwell stated; up to 660k more GB300s by late Dec if lucky; path toward ~1.44M. 1.2 GW plant, 41 turbines. Temp fleet out by Jul 2027. / TTB 0927