Elon Musk wrote on X on 31 October that “There will be a total of around 1.2M Blackwells at Colossus 2 by end of year”. Earlier the same day he named four Colossus 2 buildings that are to hold about 860,000 GB300s between them, which leaves about 340,000 not accounted for by the buildings he named, by our arithmetic.
The posts count SpaceX’s Memphis-area AI campus building by building, and split it into training and inference halls.
Short explainer · Colossus 2 by the numbers
A vertical news Short summarising this story. Figures come from Elon Musk’s posts on X and our earlier SpaceX reporting; our arithmetic is labelled.
Business Pill · Neoclouds
A short explainer on neoclouds and who owns AI compute. It teaches the general idea only and says nothing about Colossus 2.
The key insight: As we read Musk’s posts, the four named buildings would hold about 860,000 GB300s, and his 1.2 million figure for the year-end leaves about 340,000 Blackwells outside them, by our arithmetic.
In a post at 17:45 UTC, Musk wrote that “Macroharder and Minihard both have 220k GB300s with 800 gig nics configured for training”, and that “Gun Range will have the same”.
He added that “Spam (used to house Hormel products) will have ~200k GB300s, but configured for inference, so much simpler networking and less than half the fiber”. He wrote that he had walked the whole Colossus 2 site twice, and that “Gun Range is tracking to be operational by end of November and Spam in December”.
Less than an hour later, replying to a post that added up the four buildings to roughly 860,000 GB300s, Musk gave the 1.2 million figure for the whole campus by the end of the year.
Colossus 2 chips by building as Elon Musk posted on X on 31 October 2026, and his estimate of about 1.2 million Blackwells by the end of the year. Sums, shares and the unnamed remainder are our arithmetic.
What Was Posted on X
Macroharder and Minihard both have 220k GB300s with 800 gig nics configured for training. Gun Range will have the same. Spam (used to house Hormel products) will have ~200k GB300s, but configured for inference, so much simpler networking and less than half the fiber.
I walked the whole Colossus 2 site around midnight and again this morning.
Good progress by the SpaceX team and contractors. Gun Range is tracking to be operational by end of November and Spam in December.
Elon Musk posted on 31 October that Macroharder and Minihard each have 220,000 GB300s configured for training, that Gun Range will have the same, and that Spam will have about 200,000 GB300s configured for inference. He wrote that Gun Range is tracking to be operational by the end of November and Spam in December.
@XFreeze There will be a total of around 1.2M Blackwells at Colossus 2 by end of year
Replying to a post that added up the four buildings to roughly 860,000 GB300s, Musk wrote the same day that there will be a total of around 1.2 million Blackwells at Colossus 2 by the end of the year. The piece compares the two figures.
SpaceX is making huge extensions to their Colossus 2 AI/SI compute campus:
The X account XFreeze posted on 31 October a building-by-building list for Colossus 2: Macrohard with 110,000 GB200s and 220,000 GB300s, Macroharder and Minihard with 220,000 GB300s each for training, Gun Range with 220,000 GB300s targeting the end of November, and Spam with about 200,000 GB300s for inference targeting December. Musk’s two 31 October posts do not name Macrohard.
What Happened Next (31 Oct, evening)
Musk’s own earlier numbers point to where the roughly 340,000 gap sits. In September, Musk wrote on X that Colossus 2 was running 110,000 Nvidia GB200 chips and 440,000 GB300s, and set out a schedule to about 1.21 million GB200 and GB300 processors by year-end, including 1.1 million GB300s, as Bloomberg reported on 25 September (summarised by AI Weekly). Set against the four buildings Musk named on 31 October (about 860,000 GB300s), that leaves the 110,000 GB200s plus about 240,000 GB300s (our arithmetic). The X account XFreeze places them in a fifth building it calls Macrohard, with 110,000 GB200s and 220,000 GB300s; Musk’s 31 October posts do not name that building. GB200 and GB300 are both Nvidia Blackwell chips.
The Numbers in His Posts
Two buildings, Macroharder and Minihard, are already full on Musk’s account: 440,000 GB300s. Gun Range and Spam add about 420,000 more, due by December. That is about 210,000 a month over November and December, by our arithmetic.
That is close to the pace SpaceX gave on its 20 October earnings call, when CNBC reported it was on track to bring 200,000 Grace Blackwell systems online every month for the rest of 2026.
Of the four named buildings, one is set up for inference: about 200,000 of 860,000 GB300s, or about 23%, by our arithmetic. The other 77% is configured for training.
The Business Lever: Owning the Compute
SpaceX is building its own AI capacity rather than renting it. In its third quarter, its AI segment took $16,444 million of $19,593 million in company capital spending, about 84%, as we reported from its filing.
As we read the first post, two of the four buildings are already in place, at 440,000 GB300s, and the other two add about 420,000 by December, about 210,000 a month by our arithmetic.
As we read it, that pace is close to the 200,000 Grace Blackwell systems a month that CNBC reported SpaceX gave on its 20 October call.
As we read the layout, three buildings are configured for training and one, Spam, for inference, about 23% of the named total by our arithmetic.
“There will be a total of around 1.2M Blackwells at Colossus 2 by end of year”
Three Implications
FOR NVIDIA Musk’s posts put about 860,000 GB300s in four named Colossus 2 buildings and about 1.2 million Blackwells across the campus by the year-end.
FOR SPACEX’S BALANCE SHEET Its AI segment took about 84% of third-quarter capital spending, $16,444 million of $19,593 million, per its filing as we reported.
FOR RIVALS RENTING COMPUTE Musk wrote that the inference building needs much simpler networking and less than half the fiber of the training halls.
The framework says “The structural bottleneck isn’t computing, it’s electrons.”
As we read them, Musk’s posts give dates building by building; they do not say what sets those dates. The framework is the owner’s argument and does not discuss Colossus 2.
The counts come from Musk’s posts on X, not from a SpaceX filing or an Nvidia statement. Neither post says whether “GB300s” and “Blackwells” count individual GPUs or larger systems, so part of the roughly 340,000 gap may reflect units rather than extra buildings. The posts do not say where those chips would sit or how many are already installed.
The Gun Range and Spam dates are targets Musk gave, and the 1.2 million figure is his estimate for the end of the year. Power, cost and the chips’ price are not in the posts, and we do not estimate them.
This Week In The AI Supercycle
Anthropic’s Reality Check, The Capital Squeeze, & The New World Order
This week’s issue: for two years the AI build-out was priced as a growthstory. This week it started being priced as a credit story.
Musk put Colossus 2 at about 1.2 million Blackwells by the end of 2026, with about 860,000 GB300s in four named buildings, 440,000 of them already in place and about 23% of the named total set up for inference, by our arithmetic.
A note on sourcing. Chip counts come from two posts on X by Elon Musk on 31 October 2026; the call pace, capex and debt figures come from our earlier reporting of CNBC, Bloomberg and SpaceX’s filing. Sums and shares are our arithmetic. We did not contact SpaceX or Nvidia.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
