Two hundred thousand GPUs, built in under a year — and the company’s own page gives two different counts. Colossus is the Memphis cluster that trains Grok. xAI: “Built in 122 days — outpacing every estimate — Colossus was the most powerful AI training system yet. Then we doubled it in 92 days to 200k GPUs.”SpaceXAI, Colossus, read at source 16 Sep 2026.
The published figures, and the one place they disagree with each other.
- 200,000 H100 GPUs is the headline figure on xAI’s own page.
- 180K is the figure in the stats block on that same page, with the same caption. It is not reconciled.SpaceXAI, Colossus, read at source 16 Sep 2026: the hero figure is “200,000”; the stats block below it renders 180K under the caption “NVIDIA H100 GPUs in a single interconnected cluster”.
- 122 days to build, 92 more to double it, against an estimate of 24 months.
- 0.5 EB of storage for training data and checkpoints.
- The roadmap says 1M GPUs. That is a plan, not a measurement.
The published numbers
| Figure | What xAI captions it as |
|---|---|
| 200,000 | H100 GPUs in a single interconnected cluster |
| 180K | NVIDIA H100 GPUs in a single interconnected cluster |
| 170 PB/s | aggregate memory bandwidth across the full system |
| 2.8 Tb/s | per-server network bandwidth for distributed training |
| 0.5 EB | total storage for training data and checkpoints |
| 276 | days from groundbreak |
SpaceXAI, Colossus, read at source 16 Sep 2026. The two GPU counts are both on that page; this site is reporting the discrepancy, not resolving it.
When a company’s own marketing page carries two sizes for the same cluster, neither number is a measurement. They are both claims, and they are twenty thousand GPUs apart.
How fast was it built?
The timeline on the page runs from May 2024 to February 2025, where it is labelled “Full scale” and “Doubled to 200K GPUs.” xAI frames the speed as the point: “We were told it would take 24 months to build. So we took the project into our own hands.”
Its own summary of the result is a superlative: “We doubled our compute at an unprecedented rate, with a roadmap to 1M GPUs. Progress in AI is driven by compute and no one has come close to building at this magnitude and speed.”
That is xAI’s claim about its competitors, published by xAI. This page has not checked it against anyone else’s build, and no rival publishes comparable figures in the same form.SpaceXAI, Colossus, read at source 16 Sep 2026: “no one has come close to building at this magnitude and speed”.
Why the size matters at all
Because the model announcements point back at it. Grok 4.5 was trained “across tens of thousands of NVIDIA GB300 GPUs” — a different, newer chip from the H100s counted above, which is a reminder that a cluster is not one generation of hardware forever.SpaceXAI, Introducing Grok 4.5, 16 Jul 2026, read at source 16 Sep 2026. What changed between 4.5 and 4.6 is downstream of this building.
It is also being rented out. A May 2026 xAI post says “SpaceXAI has signed an agreement with Anthropic to provide access to Colossus 1” — the reading behind that sentence, and the ownership change it sits inside, is on who owns Grok now.xAI, news index, read in a browser 16 Sep 2026; the site refuses automated fetches.
The renter gives a third number. Anthropic’s own announcement, dated 6 May 2026, says the deal covers all of Colossus 1 and puts it at over 220,000 NVIDIA GPUs — more than either count on xAI’s page, and without naming the chip. It does not say whether Colossus 1 is exactly the cluster xAI’s page describes, and this site is not reconciling the three figures.Anthropic, Higher usage limits for Claude and a compute deal with SpaceX, 6 May 2026, read at source 22 Sep 2026: “We’ve signed an agreement with SpaceX to use all of the compute capacity at their Colossus 1 data center. This gives us access to more than 300 megawatts of new capacity (over 220,000 NVIDIA GPUs) within the month.”
What this page could not verify
- Which GPU count is current. xAI’s two appear on the same page, undated; Anthropic’s over 220,000 is a third, from the customer rather than the operator.Anthropic, Higher usage limits for Claude and a compute deal with SpaceX, 6 May 2026, read at source 22 Sep 2026: “over 220,000 NVIDIA GPUs”.
- Power and water use. Not published on the page read here; the wider picture is in what AI actually uses.
- Anything a third party measured. Every figure here is the operator’s own, or its customer’s.
SpaceXAI, Colossus and Introducing Grok 4.5, 16 July 2026. Both read at source on 16 September 2026.
Cluster sizes change and marketing pages lag. Re-read before quoting.