SpaceX Colossus 2: 1.2 Million Blackwell GPUs by Year-End, Claimed vs Verified

SpaceX CEO Elon Musk says the Colossus 2 AI site will hold "a total of around 1.2M Blackwells" by the end of 2026, by adding two new clusters called Gun Range and Spam that together carry roughly 420,000 Nvidia GB300 chips. The headline number is plausible against the buildout Musk laid out in September, but today it rests on his own posts on X. No filing, Nvidia statement or independent count that we could find confirms it.

This is a claimed-versus-verified piece: what was said, by whom, what lines up with earlier reporting, and what nobody outside the company has checked.

TL;DR: what is claimed and what is verified

What Musk actually said

According to reporting on his X posts, Musk described the two new clusters this way: Gun Range is "tracking to be operational by end of November" and is configured for training like the site's other big clusters, while Spam arrives in December with "~200k GB300s, but configured for inference." He also said there will be "a total of around 1.2M Blackwells at Colossus 2 by end of year." The names are literal: Musk said Gun Range "was literally the local police department's gun range that we bought," and Spam is named for a facility that used to house Hormel products, per Tesla North's write-up.

The same write-up says the existing Macroharder and Minihard clusters each have about 220,000 GB300s with 800-gigabit networking, configured for training. Musk also said he walked the site around midnight and again that morning, which is color rather than evidence of chip counts.

We did not find an official SpaceX or Nvidia press release for this announcement. The primary source is Musk's own account, and the secondary coverage we read repeats his figures.

How this fits the September roadmap

The number did not come from nowhere. On September 25, 2026, Musk posted that Colossus 2 near Memphis ran 110,000 Nvidia GB200 chips and 440,000 GB300s, with another 220,000 GB300s "fully operational next week," another 220,000 in November, and, "If we get lucky, yet another 220k GB300 by late December." AI Weekly's summary and Invezz both walk through the arithmetic: 550,000 chips then, up to about 1.21 million if all three waves land.

Put the two sets of posts side by side and the story hangs together: the October names (Gun Range, Spam) appear to be the labels for the final waves in the September plan, and the year-end total is the same 1.2M. Two points do not line up cleanly, and we flag them rather than smooth them over:

• Per-cluster sizes are loose. Reports of the October posts imply a split near 220,000 and 200,000 chips, while the headline is "about 420,000." Sources differ on whether Gun Range is exactly 220,000.
• The "if we get lucky" caveat has quietly shrunk. In September the last 220,000 chips were explicitly conditional. In the October reporting the year-end total is stated flatly. Nothing we found says what changed.

What is verified, what is not

Reasonably supported (consistent across several outlets quoting Musk):

• The site and its rough scale in September: about 550,000 Blackwell-class chips.
• The planned shape of the buildout: waves of about 220,000 GB300s.
• That a training/inference split is planned for the new clusters.

Self-reported only, no independent confirmation found:

• The current installed count, and the "operational next week" claims from September.
• The 1.2M year-end total.
• The dates for Gun Range (end of November) and Spam (December).

Genuinely unknown:

• Whether power, cooling and networking will be ready for 420,000 more chips in two months.
• How many of the chips will train Grok and how many will serve outside customers.
• Whether the GB200 and GB300 supply Musk needs has actually been delivered, as opposed to ordered.

In September, Musk tied the step size to networking, not chip supply: he said the step size (110,000 chips, per AI Weekly's summary) reflects "the number of fiber optic cables that can be plugged into a central switch." That is a useful hint for reading these numbers, because it means the cadence of large blocks is set by switch design, not just by how many chips Nvidia ships.

Why training versus inference matters

Green chip in front of server towers, illustrating Colossus 2 GPUs split between training and inference

The two new clusters are not equivalent. Training clusters link many chips with dense, high-bandwidth networking so they behave like one very large machine. Inference clusters mostly run many independent copies of a model, so they need simpler networking and, per the Tesla North report, less than half the fiber.

That means the headline chip count overstates what the buildout does for frontier model training. If Spam's roughly 200,000 GB300s are inference, only Gun Range adds to the training pool this year. For anyone comparing labs by raw GPU count, the honest comparison separates the two uses. We made the same point about marketing numbers in our piece on China's 24 GW compute versus the US's 56 GW: a chip count is not a capability.

Who buys this capacity

Colossus is no longer just a Grok training farm. On SpaceX's third-quarter earnings call, Musk said the rough goal is about a 50/50 split between internal use and rentals, which can shift month to month, per the Tesla North report. Our earlier coverage of that call is in SpaceX Q3 2026: AI revenue and the 1.7 GW Cursor deal.

Anthropic is the headline tenant. Its IPO filing, as reported, says it agreed to spend up to $120.5 billion on SpaceX AI compute through 2029, roughly $3.1 billion a month once contracted capacity is delivered, which Anthropic expects in December. The deals can reportedly be cancelled with 90 days' notice after early 2027, so the total is a ceiling, not a locked-in cost. We dug into the revenue side of that relationship in the Bloomberg "Customer B" claim, and the relationship began with Anthropic's access to Colossus 1.

That makes the December timing matter outside SpaceX. If the new chips land on schedule, the contracted capacity Anthropic expects has somewhere to run. If Gun Range slips, the pressure shows up as rate limits and supply constraints for customers of the labs that rent here, not in a press release.

The money behind it

Whatever the exact chip count, 420,000 GB300s is expensive, and no cost for the added GPUs has been disclosed. SpaceX has separately locked in $40 billion of GPU debt, which we covered in Musk's $40B GPU debt claim. The two figures are not directly comparable, and neither is audited. If you want the filed numbers, the place to look is SpaceX's own SEC filings, not the social posts.

The part nobody has priced: power and neighbors

Power lines feeding a data center, the electricity question behind 1.2M Blackwell GPUs

Chips need electricity, and the Memphis-area buildout has drawn scrutiny for it. As AI Weekly noted in its September summary, Reuters reported in July that 59 natural-gas turbines had been installed on the site without federal clean-air permits, and xAI said it was replacing that temporary generation with a 1.2-gigawatt permanent plant. Neither the September nor the October posts say when that plant will be energized, or how 1.2M Blackwells will be powered in the meantime.

Local friction is already live. Our reporting on the xAI Southaven homeowner buyouts and noise lawsuit shows what neighbors of these sites are dealing with. More chips mean more load, more noise and more permit questions, which is why year-end targets at this scale deserve skepticism until power is confirmed.

How to check this yourself

You do not need to take Musk's word, or ours. These are the checks that would turn a claim into a fact:

1. SEC filings. SpaceX's quarterly and annual filings list capital expenditure and, usually, compute capacity. Compare the December quarter's disclosures against the 1.2M figure.
2. Nvidia's earnings commentary. Large GB300 shipments to a single customer tend to show up in Nvidia's remarks and in analyst notes.
3. Permits and utility filings. Power interconnection requests and air permits for the site show whether capacity is physically coming.
4. Customer behavior. If Anthropic or other renters announce new capacity or limits in December, that is indirect evidence the chips are live.
5. Dates. Gun Range is promised for the end of November. A missed date is the first public signal that the year-end total is slipping.

What this means for what you build or pay

If you build on Claude or Grok, the practical question is capacity, not chip counts. More compute online by December supports higher limits and fewer capacity-related slowdowns, but only if the schedule holds. Plan as if it might not: do not size a production workload around promised future throughput.

If you buy or rent GPUs, a supplier that is both a lab and a landlord has an incentive to announce big numbers. Ask for contracted delivery dates in writing.

For a broader look at how infrastructure claims have been checked this month, see our tracker-style coverage in the October 2026 AI data-center and cyberattack roundup and our explainer on what data-center moratoriums mean.

Bottom line

Musk's claim is specific, internally consistent with his September plan, and matches how large clusters are built in 220,000-chip blocks. It is also entirely self-reported, and the one hedge he gave in September ("if we get lucky") has dropped out of the October version. Until a filing, a customer, or an independent count says otherwise, read "1.2M Blackwells by year-end" as a target. Check again after the end-of-November date for Gun Range.

Related reading

• SpaceX Q3 2026 earnings: AI revenue and the Cursor deal
• Bloomberg's "Customer B" claim: Anthropic and SpaceX revenue
• Musk's $40B GPU debt claim
• Anthropic and SpaceX Colossus 1
• xAI Southaven buyouts and noise lawsuit
• China 24 GW vs US 56 GW compute

Figures are as reported on October 31, 2026 and may change; the primary source for the cluster plan is Musk's posts on X.