SpaceX’s next two AI supercomputer buildings at Colossus 2 in Memphis have names now, and they’re very Elon: Gun Range and Spam. Elon Musk said Gun Range will hold 220,000 Nvidia GB300s set up for training and should be running by the end of November, while Spam will hold about 200,000 GB300s set up for inference, due in December.
Musk laid it out on Saturday, Oct. 31 at 1:45 PM ET / 10:45 AM PT (2:45 AM JST Sunday), in a reply on X to a walkthrough of the site. “Macroharder and Minihard both have 220k GB300s with 800 gig nics configured for training. Gun Range will have the same,” he wrote. “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 added that he “walked the whole Colossus 2 site around midnight and again this morning,” and that Gun Range “is tracking to be operational by end of November and Spam in December.” About half an hour later, he put a number on the whole campus: “There will be a total of around 1.2M Blackwells at Colossus 2 by end of year.”
Yes, the names are real
Gun Range isn’t a metaphor. “It was literally the local police department’s gun range that we bought,” Musk wrote, with a laughing emoji. Spam is a former warehouse for Hormel products, the company behind the canned meat. His follow-up: “I hope Spam doesn’t produce too much spam!” (An inference cluster’s whole job is producing text on demand, so the joke has an uncomfortable edge for anyone who reads their inbox.)
The two buildings join Macroharder and Minihard, which Musk says already hold 220,000 GB300s each. Add Gun Range and Spam, and the four named buildings come to about 860,000 GB300s, a tally the X account XFreeze posted and Musk replied to. Sawyer Merritt estimated that the two new clusters alone represent more than $35 billion of GPUs. That’s his estimate, not a SpaceX figure.
Why training and inference get different buildings
This is the genuinely new part. Until now, Colossus has mostly been described in raw chip counts. Musk is now splitting the campus by job.
Training a large model means thousands of chips swapping data constantly, so every GPU needs a fat, fast connection to every other one. That’s the “800 gig nics” line. Nvidia’s GB300 NVL72 platform pairs each Blackwell Ultra GPU with ConnectX-8 networking at 800 Gb/s. Inference, which is answering users’ questions with a model that’s already trained, doesn’t need nearly as much chip-to-chip traffic. Musk’s “less than half the fiber” remark is the practical result. Fewer cables, simpler switches, faster to build.
Of the roughly 860,000 GB300s in the four named buildings, about 23% would be inference, by FourWeekMBA’s arithmetic. That’s a sign of where demand is going. SpaceX rents compute to outside AI developers, and Musk has described a rough 50/50 split between internal use and rentals, TeslaNorth noted. Rented capacity and Grok both need serving hardware once the models are done.
What’s confirmed and what isn’t
Every number here comes from Musk’s posts, not a SpaceX filing or an Nvidia statement. The posts don’t say how many chips are already installed in Gun Range or Spam, how much power they need, or whether “GB300s” and “Blackwells” are counted the same way. The gap between the four buildings’ ~860,000 and the 1.2 million total isn’t explained either. It could include older Blackwell systems elsewhere on the campus.
The 1.2 million target itself isn’t new. Musk said in September he hoped to reach 1.21 million chips by late December, as we covered when he said SpaceX had closed $40 billion of GPU debt. He also said then that most of those debt-financed GPUs arrive in the first half of 2027, so the year-end figure and the debt-funded wave are separate steps. What Saturday added is the timeline by building, and the first clear split between training and serving.
The EVX Take
For scale, Japan’s flagship public AI supercomputer, AIST’s ABCI 3.0, runs 6,128 Nvidia H200 GPUs. Gun Range alone is meant to hold about 36 times that many GPUs, and of a newer generation, in a building that used to be a police shooting range. That gap is the real story for Japan. Domestic AI developers will keep renting overseas capacity like this rather than matching it at home. The upside lands with Japanese suppliers inside every Blackwell build, such as Advantest, whose testers check Nvidia’s AI chips and which has been a regular in our Stock Watch. If Gun Range and Spam hit their November and December dates, expect that demand to show up in Tokyo chip stocks before it shows up in any SpaceX filing.
The first test is close. End of November is four weeks away, and Musk has given himself a public deadline for Gun Range.
Sources: Elon Musk on X (clusters, 1.2M total, Gun Range name, Spam), Sawyer Merritt on X, TeslaNorth, FourWeekMBA, AI Industry Today, Nvidia (GB300 NVL72), ABCI 3.0 (AIST).
— The EVX News team
Featured image: An Nvidia DGX B200 system, built on Nvidia’s Blackwell GPUs, in a server rack. Not a photo of Colossus 2. Photo: Javils00 / Wikimedia Commons (CC0, cropped).
 goes live by end of November and Spam (~200,000 GB300s, inference) in December, toward 1.2M Blackwells at Colossus 2.){kind=link}