SpaceXAI to Add 660,000 GPUs This Year, Build 1.2GW Power Plant to Run Them

Elon Musk's SpaceXAI will add another 660,000 AI GPUs this year, pushing its total operational fleet to roughly 1.44 million accelerators, according to details circulating about the company's buildout. The firm is also constructing a 1.2-gigawatt power plant to bring those systems fully online, a pairing that underscores how power, not silicon, is becoming the binding constraint on large-scale AI training.

The numbers put SpaceXAI in the top tier of private AI compute operators. For comparison, Meta has said it expects to hold roughly 600,000 H100-equivalent GPUs by the end of 2024, and xAI's Colossus cluster in Memphis reached 100,000 H100s when it came online last year before expanding. A fleet approaching 1.44 million chips would place SpaceXAI's capacity in the same conversation as the largest hyperscaler deployments, though the mix of GPU generations and whether all units run at full utilization remain open questions.

The 1.2-gigawatt power plant is the more unusual piece. Data-center operators typically secure power through utility contracts and on-site gas turbines rather than building generation outright, but the scale of the cluster makes grid interconnection timelines a bottleneck. A 1.2GW plant is comparable to a large natural-gas combined-cycle facility and would rank among the biggest single-site power projects tied to AI compute. Musk has previously leaned on gas turbines and battery storage for xAI's Memphis site, which drew scrutiny from environmental regulators over unpermitted turbines.

Musk's AI ambitions now span two corporate entities, xAI and the SpaceXAI name referenced in the buildout, plus the social platform X that feeds data into the models. The GPU expansion signals that training runs, not inference alone, are driving the capex. At current market prices, 660,000 additional accelerators represent a multi-billion-dollar outlay before power infrastructure, cooling, and networking are counted, and the power plant adds hundreds of millions more.

What to watch is whether the power comes online in step with the chips. GPUs that arrive before generation capacity sits idle or runs throttled, and Musk's track record on timelines has been optimistic. The other question is where the electricity comes from and what regulators say about it, given the scrutiny that followed xAI's Memphis turbines. If SpaceXAI hits even close to 1.44 million operational GPUs with dedicated generation behind them, it changes the compute math for every lab competing for training capacity.