A.I.
Rycode

Jeffrey 3 min read

Hardware · Training campus

Memphis / Colossus

A hyperscale GPU campus in Memphis. Dated company figures for chips, power, and the cluster, including 170 PB/s and 2.8 Tb/s.

What it is

Colossus is published as an AI supercomputing cluster, a “gigafactory of compute” behind Grok. Company Memphis pages place it in the Memphis area, South Memphis and Southwest Memphis (former factory). Community updates also reference Southaven, Mississippi, for temporary turbines. Pages currently brand as SpaceXAI after the SpaceX acquisition of xAI. Product and news URLs remain under x.ai.

Colossus is the flagship site on Paul R. Lowry Road in Memphis. Colossus II is data centers in Memphis and in Southaven, Mississippi.

Facility footprint (leadership page): a vacant 1-million-square-foot factory in Southwest Memphis converted into a technology hub. Colossus 1 was later contracted as capacity for Anthropic (May 2026).

Dated GPU quotes

Date Source Quoted claim
23 Dec 2024 Series C 100,000 Nvidia Hopper GPUs operating in 122 days; plan to double to a combined 200,000 Hopper GPUs (Spectrum-X called out).
20 May 2025 Musk via CNBC 200,000 installed in Memphis.
Fetched 13 Sep 2026 x.ai/colossus Built in 122 days, doubled in 92 days to 200k GPUs. Hero: 200,000 H100 GPUs in one cluster. Same page also shows 180K NVIDIA H100 GPUs in one cluster. Roadmap to 1M GPUs. Timeline graphic: 276 days, May 2024 to Feb 2025.
9 Jul 2025 Grok 4 post Training for Grok 4 used Colossus, “our 200,000 GPU cluster.”
6 Jan 2026 Series E Colossus I and II ended 2025 above one million H100-equivalents.
Live Memphis copy x.ai/memphis Plan to equip “this facility” with 1 million GPUs by 2026.
6 May 2026 Anthropic partnership Colossus 1: over 220,000 Nvidia chips, a mix of H100, H200, and GB200.
6 May 2026 Anthropic post Anthropic uses all Colossus 1 capacity: more than 300 MW and over 220,000 NVIDIA GPUs “within the month.”
12 Jun 2026 SpaceX prospectus First Colossus cluster online in 122 days. First Colossus II cluster in 91 days. The two sites together about 1 gigawatt of compute power, with more capacity available.

Power

Figure Label Source
>300 MW for Colossus 1 under the Anthropic deal Company and partner announcement Anthropic 6 May 2026; SpaceXAI partnership post aligns on GPU count

Company power phrase

The 12 June 2026 prospectus says the two sites together provide about 1 gigawatt of compute power, with more capacity available.

Southaven

Southaven sits at 2875 Stanton Road. As of 30 July 2026, 69 temporary gas turbines were powering the site. The company said a 1.2 gigawatt plant of 41 permanent turbines is under construction. A state permit from 11 March 2026 is for those 41. The company said temporary turbines start coming off as early as August 2026 and are gone by July 2027.

Memphis grid

Undated company page: a finished substation rated 150 megawatts, which the company says cost $35 million, plus a second substation at $20 million.

Vera and Rubin

On 24 August 2026 Nvidia said it will put Vera processors and Vera Rubin chips into this build, including a planned Starmind satellite.

Capabilities

From the live Colossus page, fetched 13 Sep 2026: 170 PB/s aggregate memory bandwidth; 2.8 Tb/s per-server network; 0.5 EB storage. Series C called out NVIDIA Spectrum-X Ethernet for the 100k to 200k path.

Sources