What total AI/data-center compute capacity will SpaceX (including its subsidiary xAI, following their February 2026 merger) have physically operational and running, as of December 31, 2027?

Defining the Baseline and the Metric The resolution tracks operational IT power capacity (nameplate compute), which strictly excludes gross power-plant metrics, cooling overhead, and energized-but-unpopulated facilities 2 sources. As of mid-2026, SpaceX reported 1.4 GW of nameplate compute 2 sources, up from 0.4 GW a year earlier. Independent verification broadly aligns, with Epoch AI measuring Colossus 2 alone at 946 MW of IT capacity in June 2026 epoch.ai. Management guidance anticipates exceeding 2 GW by the end of 2026 2 sources, which serves as the most credible launching point for the 2027 buildout.

Discounting Aggressive Targets Musk has guided toward "closer to 10 GW" of compute by year-end 2027, corresponding to roughly 15–20 GW at the power-plant level 3 sources. While SpaceX and xAI genuinely execute faster than industry peers, these headline targets must be heavily discounted. Historically, announced milestones run 4 to 6 months and 20–40% ahead of verifiable, populated reality 2 sources. Achieving a 5x leap from a 2 GW end-2026 base within 12 months would far exceed anything the company has achieved from a mature baseline, especially when the metric demands hardware is physically powered on and usable for compute workloads.

Execution Constraints: Capex, Power, and Supply Three binding constraints will dictate the 2027 deployment pace. First, capital: adding 8 GW of IT capacity requires $300–$500B in capex 2 sources, vastly exceeding the current ~$15.8B/quarter run rate 2 sources. Second, power and permitting: the Memphis/Southaven cluster faces acute execution risks as 69 temporary unpermitted turbines (~1.7 GW) must be removed by July 2027 and replaced with a permanent plant, creating a real risk of capacity temporarily treading water 3 sources. Finally, while offtake agreements with Anthropic and Google de-risk financing 2 sources, memory and GPU supply bottlenecks remain a stated vulnerability 2 sources.

Orbital Compute is a Rounding Error Despite the strategic focus on the Starmind AI1 satellite and the million-satellite FCC filing spacenews.com, orbital capacity will be negligible by the December 2027 resolution. Each AI1 satellite draws only ~120–175 kW average IT power 3 sources. Volume production at the Bastrop Gigasat Factory is scheduled to begin "as soon as late 2027" 2 sources. Even an unusually successful launch cadence of a few hundred early units would add less than 0.05 GW. The operational total will be almost entirely determined by terrestrial facilities.

Synthesis of the Distribution A median of 5.6 GW reflects a massive but realistic near-tripling of the projected end-2026 base, balancing SpaceX's proven ability to parallelize construction newsletter.semianalysis.com against physical and financial gravity. The p90 of 10.17 GW respects the SemiAnalysis bull case where massive external financing and unconstrained hardware allocation allow the company to hit its 10 GW headline target using generous nameplate accounting. Conversely, the p10 of 2.833 GW accounts for severe downside risks, including a capex pullback, adverse turbine litigation, or severe interconnection delays that strand constructed facilities without power.