Colossus Has 770,000 GPUs and 1GW of Compute — Two Thirds of It Never Touches the Grid
When Elon Musk turned a shuttered Electrolux appliance factory in southwest Memphis into the world’s largest AI training cluster in 122 days during the summer of 2024, the focus was on speed. NVIDIA CEO Jensen Huang called the build “superhuman.” What got less attention was the harder problem being quietly solved in parallel: where the power would actually come from.
By early 2026, Colossus had grown into a three-building campus straddling the Tennessee-Mississippi state line. Total footprint: approximately 2.6 million square feet. GPU count: approximately 770,000. IT compute capacity: approximately 1GW, with Elon Musk projecting a doubling to 2GW.
The scale alone places Colossus among the most concentrated compute installations on the planet. But the power architecture is what separates it from every other gigawatt-class AI campus currently in operation.
Two of Three Buildings Run Off the Grid
Most frontier AI data centers are grid-dependent. Amazon’s New Carlisle, Microsoft’s Fairwater, Google’s Lake Mariner — all of them negotiate with utilities, queue for interconnection, and wait years for approvals. xAI did none of that for two of the three Colossus buildings.
Those two run entirely behind-the-meter, powered by a 1.2GW natural gas plant that xAI built for itself on the opposite side of the state line. Only the original Colossus 1 building draws firm power from the utility grid. The other two are structurally isolated from the regional transmission system.
Behind-the-meter generation at this scale is functionally a private utility. Building the generation asset:
- Eliminated the multi-year utility interconnection queue
- Insulated Colossus from grid volatility and rate renegotiation
- Preserved the ability to expand compute without fighting for new power allocation
- Sidestepped permitting delays that have pushed grid-tied data center timelines past 2027 in most US markets
What 770,000 GPUs Actually Means
The original Colossus 1 housed 100,000 NVIDIA H100s. Expanding to 770,000 across three buildings implies later-generation hardware fills much of the new space. At the footprint-to-GPU ratio reported, GB200 NVL72 racks are the most plausible candidate — consistent with the March 2026 expansion to GB200 on Colossus 2.
Anthropic now leases all 220,000 GPU slots at Colossus 1 from SpaceX following the xAI dissolution in May 2026. The remaining buildings operate under SpaceXAI’s own compute budget, primarily used to train Grok V9 — the 1.5-trillion-parameter foundation model xAI confirmed is underway.
At 1GW continuous IT load, Colossus draws roughly the equivalent of the annual electricity consumption of 890,000 average US homes. It generates that power on-site from its own plant.
The Pattern This Fits
The Colossus architecture foreshadows what frontier AI infrastructure is becoming: an energy business first, a compute business second. The clusters that will dominate the next half-decade are constrained not by chip supply but by generation capacity.
Microsoft signed a 2.67GW natural gas agreement with Chevron for its West Texas campus. Amazon is sourcing nuclear power for its Kuiper build. Google has contracted with nuclear startups and run TPU clusters on hydroelectric capacity in Oregon. All three are hedging grid dependency. xAI, characteristically, built a gas plant and skipped the hedge entirely.
At 2GW — Musk’s stated projection — Colossus would exceed the output of a mid-sized US nuclear plant. The AI compute race and the energy infrastructure race have converged into the same problem.
Key Numbers
| Metric | Value |
|---|---|
| Buildings | 3 (TN-MS border) |
| Total footprint | ~2.6M sq ft |
| GPU count | ~770,000 |
| IT compute capacity | ~1GW |
| Off-grid buildings | 2 of 3 |
| Gas plant capacity | 1.2GW |
| Projected capacity | 2GW |