Musk Walks Colossus 2 at Midnight and Reveals a Four-Building AI Plan

1. A midnight site check starts the Colossus 2 story

Elon Musk laid out the building-by-building makeup of Colossus 2 in an X reply early Sunday, Nov. 1. He said he walked the entire site twice, around midnight and again the next morning. He also said the SpaceX team and the contractors are making good progress.

The story is not a vague expansion pledge. It is a site plan that spells out which GPUs go in which building, what each will train, and where inference will run. But the start dates Musk gave are targets, not confirmation that the work is done.

2. From Macroharder to Spam: the role of each building

"Macroharder" reads like a parody of Microsoft's name. "Spam" is a building that once stored Hormel products. The names are playful, but the roles are clear.

In a separate post, Herbert Ong described Macrohard as roughly 550,000 GB200s, Macroharder as about 810,000 square feet and Minihard as about 312,000 square feet. He also mentioned a land swap for a fifth data center, a warehouse that could become a sixth facility, and 444 acres at Starbase in Brownsville. These are Ong's descriptions, not building-level figures Musk confirmed himself.

3. Why a model that learns and a model that answers need different networks

Training is closer to an AI reading its textbooks and learning how to solve problems. Inference is the already-trained AI answering a user's question. That difference reaches all the way down to how a data center is wired.

In training, many GPUs must trade results constantly. An 800 gig NIC is the high-speed network interface card that handles that traffic. Inference, depending on how the work is run, can reduce the communication load between GPUs. Musk said Spam's networking is far simpler and uses less than half the fiber of the training buildings. That is a design feature of Spam as he described it, not a rule for every inference system.

4. About 860,000 GB300s versus 1.2 million Blackwells: why the numbers differ

X Freeze added up the GB300s across the four buildings and got about 860,000. That total is X Freeze's arithmetic, not a figure Musk gave. Sawyer Merritt grouped the GPUs in the two new buildings, Gun Range and Spam, at about 420,000 and estimated their value at more than $35 billion. That cost estimate is not Musk's statement.

At 3:21 a.m. KST Sunday, Musk replied that around 1.2 million Blackwells would be at Colossus 2 in total by year-end. The 860,000 figure is the GB300 sum for four specific buildings. The 1.2 million figure refers to all Blackwells at Colossus 2. Because the scope and classification differ, the two numbers cannot be compared on the same basis, and the gap cannot be read as unconfirmed capacity. Unlike an earlier article on SpaceX's revenue outlook, this one focuses on building-level deployment and whether the sites are running.

5. The reaction on X, and Musk's joke

After Musk's building-by-building post, Merritt summarized the new facilities, and X Freeze totaled the GPUs across the four buildings. Musk then joked that he hoped Spam would not produce too much spam. Next came his year-end target of about 1.2 million Blackwells.

The engagement figures below were collected on the morning of Sunday, Nov. 1 (KST). They are not live numbers, and each post had a different exposure window.

6. Three timelines long-term investors can check

Publishing GPU placement and roles by building lets observers watch how AI compute buildouts are executed. But an X reply is not official company guidance or a filing. For investors who build positions gradually over the long term, it is more reasonable to look for a link between actual operations and business results than to get excited.

Installing GPUs, running them reliably, and turning that compute into revenue and cash flow are separate stages.

Insight Times Editorial Desk