ELON, EXPLAINED
EN
SpaceX / Colossus 2All explainers
SpaceX · 4 min read · 2026 · Q4 2026

Colossus 2

SpaceXSI's Memphis supercomputer, which Elon Musk calls the first gigawatt AI training cluster.

Colossus 2 is xAI's second giant AI training supercomputer, built on Tulane Road in Shelby County, Tennessee, near Memphis. It is one tightly linked cluster of NVIDIA GB200 and GB300 graphics processing units, or GPUs, the specialized chips that do the heavy math of training a neural network, and Elon calls it the world's first gigawatt-scale AI training cluster. A typical data center stitches many small jobs together, but this one is built to work as one enormous brain solving one enormous problem, which is training xAI's Grok models.

What is Colossus 2?

It is the physical engine behind Grok. The machine runs hundreds of thousands of GPUs as one connected whole, fed by a power supply that mixes three sources on purpose: grid electricity, on-site gas turbines, and Tesla Megapacks, which are large battery units that store power and steady the grid. xAI broke ground in February 2025, and Elon Musk said on January 17, 2026 that Colossus 2 was operational, with an upgrade to 1.5 gigawatts (GW) planned for April. SpaceX's own filings report its total AI computing power: about 1.0 GW at the end of March 2026, 1.4 GW at the end of June, and 1.7 GW at the end of September.

Why Colossus 2 Exists

The logic is simple: bigger models need more computing power working together. Colossus 1, xAI's first cluster, started as 100,000 H100 GPUs stood up in 122 days and grew to about 230,000 GPUs, and that was the base for early training. To train a model the size of Grok 5, which Elon describes as roughly 6 trillion parameters, the adjustable numbers a model learns in training, xAI needed to jump up by another order of magnitude, meaning about 10 times bigger. Colossus 2 was built to hold what Elon called the first batch of 550,000 GB200s and GB300s.

How it Works

A training cluster only works if every chip can talk to every other chip fast enough to act as one machine, so the GPUs are wired together as one system rather than split into separate jobs. That tight linking is what lets one model the size of Grok 5 train across the whole site at once. The power setup is just as deliberate. The Megapacks act like shock absorbers, smoothing out the hard power swings a training cluster puts on a local grid, while gas turbines and grid electricity carry the steady load. The power infrastructure now reaches from the Memphis site toward Southaven, Mississippi, because the site's appetite outgrew one footprint.

SpaceXSI's computing power more than tripled in a year

SpaceX's reported computing power rose from 0.4 GW in June 2025 to 1.4 GW in June 2026 as Colossus 2 came online.

The Economics of Colossus 2

Colossus 2 is a multi-billion-dollar bet on owning frontier computing power instead of renting it from a cloud provider. Every extra GPU is both a cost and a bit more capability at the frontier, so scaling up is the whole strategic point.

xAI's largest GPU cluster grew more than fivefold in two years

xAI's biggest cluster went from 100,000 GPUs at Colossus 1 in September 2024 to about 550,000 at Colossus 2 two years later.

Colossus 2 is roughly 2.4x the size of Colossus 1

Colossus 2 holds roughly 2.4 times the GPUs of Colossus 1, and the plan points toward 1 million GPUs and beyond. The approach ties xAI's edge directly to two suppliers, NVIDIA for chips and Tesla's energy arm for Megapacks, and it turns raw computing power into the main advantage in the AI market.

Current Status of Colossus 2

Elon Musk said in late September 2026 that Colossus 2 held 110,000 GB200 and 440,000 GB300 chips, with another 220,000 GB300s due to be fully running the following week and another 220,000 in November. In early March 2026 xAI filed a permit, reported at more than $659 million, for a four-story building of about 312,000 square feet next to the site. The roof of the main building carries the word MACROHARD, painted in September 2025 in letters Elon said would be big enough to read from space. The main job for all of it is training Grok 5.

What Comes Next

If those deliveries arrive on schedule, Colossus 2 alone would hold close to 1 million GPUs by the end of 2026. Owning the cluster outright gives xAI an edge over rivals racing to build their own gigawatt clusters, with chip supply from NVIDIA and power deals setting the pace of growth.

The Bottom Line

Colossus 2 is xAI's bet that the company with the most connected chips and the steadiest power trains the best models. Grok 5, still in training there in October 2026, is the first model built to test that bet.

Share this explainerShare on X

Timeline

  • 2025 February: xAI breaks ground on the Tulane Road data center in Shelby County, Tennessee, deploying an initial 110,000 NVIDIA GB200 GPUs powered partly by Tesla Megapacks.
  • 2025 July 22: Elon Musk says Colossus 1 runs 230,000 GPUs (including 30,000 GB200s), and Colossus 2's first batch of 550,000 GB200s and GB300s starts coming online.
  • 2025 August 22: Elon Musk says Colossus 2 will be the world's first gigawatt-plus AI training supercomputer.
  • 2025 September 17: Elon Musk calls Colossus II the world's first gigawatt AI training cluster.
  • 2026 January 6: xAI says Grok 5 is in training.
  • 2026 January 17: Elon Musk says Colossus 2 is operational, with a planned upgrade to 1.5 GW in April.
  • 2026 March: xAI files a permit for a building of about 312,000 square feet next to the site, valued at more than $659 million.
  • 2026 June 30: SpaceX reports 1.4 GW of AI compute power across Colossus 1 and 2.
  • 2026 September 30: SpaceX reports 1.7 GW of AI compute power as the Colossus II buildout continues.
  • 2026 September 25: Elon Musk reports 550,000 GPUs at Colossus 2: 110,000 GB200 and 440,000 GB300.
Sources (12)
PremiumRead next · Free featureMACROHARD, ExplainedElon's plan to end Microsoft's $3 trillion empireIn September 2025 Elon posted that MACROHARD was being painted on the roof of the Colossus 2 supercomputer in Memphis, and this feature explains the SpaceXSI project behind the name: AI agents built to do the work of an entire software company.Free to read · 28 original chartsRead the feature
Keep exploring
↑ Top