For two years, the binding constraint on scaling AI compute was not the GPU itself but the electrons feeding it. Data centers hit a wall where the grid could not deliver enough amperage to keep racks at full draw, and the answer was to build smaller sites or wait for utility upgrades. SpaceXAI is building a 1.2-gigawatt power plant on-site, which removes that ceiling entirely. The plant is the hinge: once it is online, the question shifts from power to interconnect and cooling.

The deployment plan is three tranches of 220,000 GB300 GPUs each, totaling 660,000 units this year. The first tranche lands at Colossus 2 by next week, per Musk's own statement, with the second and third arriving later in the year, and the site name suggests a second generation of the facility though the source does not specify what distinguishes it from the first. That brings the site past one million GPUs in operation and pushes the total fleet to 1.44 million, closing a target Musk set two years ago when the site was still in planning and the power infrastructure had not been committed. The three-tranche structure implies a phased ramp, each adding roughly 220,000 units of compute that the site must absorb incrementally.

What is not in the source is the interconnect topology, the per-GPU power draw, the cooling architecture, and the commissioning date of the 1.2-gigawatt plant. The GB300 is named, but its clock, memory configuration, and interconnect bus are not specified. No third-party measurement of the site's throughput has been published, and the 1.2-gigawatt figure is stated without a date, a location, or a description of the generation technology. The source is thin: a single claim from Musk, relayed through Tom's Hardware, with no independent verification of the tranche schedule or the plant's capacity.

The next data point to watch is whether the power plant is commissioned before the second 220,000-GPU tranche arrives, and the plant's generation technology is also unknown, whether gas, solar, or nuclear, which will affect the site's operational ceiling and its relationship to local grid operators. If the plant is not online by then, the second tranche will sit idle or run at reduced draw, and the 1.44-million figure becomes a paper number rather than an operational one. At 1.44 million GPUs, the network fabric must handle a bandwidth that no prior site has attempted, and the topology choice will determine whether the fleet runs as one coherent system or as a collection of smaller clusters. Watch for a second announcement or a site visit that confirms the plant's status.

The 1.2-gigawatt plant removes power as the binding constraint, shifting the bottleneck to interconnect and cooling at a scale no prior site has attempted. The three-tranche GB300 deployment implies a phased ramp that the site must absorb incrementally rather than all at once.