Colossus is the record-breaking AI supercomputer that Elon Musk's xAI built in Memphis, Tennessee to train its Grok models. Its story is the whole gigawatt-factory era in miniature: an abandoned appliance plant turned into 100,000 GPUs in 122 days, then doubled — and now a second site, Colossus 2, that is the first AI data centre in the world to cross a full gigawatt of power.
In 2024 xAI took over the shell of a shuttered Electrolux appliance factory in southwest Memphis and, in just 122 days, filled it with 100,000 NVIDIA H100 GPUs — then doubled it to 200,000 in another 92 days. Standing up a cluster that size normally takes years; NVIDIA's own leadership called the pace unprecedented.
The result is Colossus, at the time the most powerful AI training supercomputer on Earth: not a research curiosity but a single machine built to do one job at industrial scale — train the next version of Grok, the large language model that powers xAI's chatbot on X.
Colossus is the clearest example yet of a gigawatt-scale factory whose product is intelligence rather than batteries or panels. It is rated not by what it makes but by the electricity it draws.
Rather than pour a new data centre from scratch, xAI moved into an existing industrial building with the floor space, height and grid connection already there. Reusing the Electrolux plant cut the slowest part — construction — down to a fit-out.
200,000 GPUs are only useful if they act as one machine. Colossus uses NVIDIA Spectrum-X Ethernet to wire every accelerator into a single training fabric, moving the enormous flood of data between chips without the network becoming the bottleneck.
Racks this dense are far too hot for air. Colossus runs a custom direct-to-chip liquid-cooling system — a design Musk said "nobody's done at scale" — pumping coolant straight across the GPUs so they can be packed tighter and run harder.
An AI cluster's power draw spikes violently as training loads shift. To keep from destabilising the local grid, Colossus leans on banks of Tesla Megapack batteries that absorb the surges and smooth the load — the same cells from the battery gigafactories.
Colossus 1 draws around 300 megawatts — already the scale of a small city. Feeding it that fast was the hard part: xAI brought on-site gas turbines to supply power quickly, a move that drew scrutiny from Memphis residents and regulators over local air quality, and highlighted the central tension of the AI build-out — the chips are ready long before the grid is.
The answer is to go bigger. Colossus 2, xAI's second Memphis site, is designed for a full 2 gigawatts and around 555,000 next-generation GPUs — roughly $18 billion of silicon — making it the first AI data centre in the world to reach gigawatt scale. Musk has said the goal is one million GPUs across the Colossus sites.
That trajectory — 300 MW to 2 GW in barely a year — is why power, not chips, is now the binding constraint on AI, and why data centres are increasingly sited next to their own generation. It is the same gigawatt logic driving the nuclear revival.
Everything about Colossus exists to serve one product: Grok, xAI's family of large language models. More compute means models can be trained on more data with more parameters and more reinforcement learning — the ingredients behind each jump in capability. A cluster this large lets xAI train frontier-scale models in-house and iterate quickly, rather than renting capacity from a cloud provider.
It is also a statement of strategy. By owning the metal — the building, the GPUs, the power and the cooling — xAI controls its own compute destiny in a market where access to gigawatts of GPUs has become the real barrier to entry. Colossus is to AI what a gigafactory is to batteries: the bet that whoever builds the most capacity, fastest, wins.
> Sources: xAI / Greater Memphis Chamber, NVIDIA newsroom (Spectrum-X), and public reporting on Colossus and Colossus 2. Figures are latest stated or reported values and change as the sites expand.