Here is the full SpaceX Q2 $18.4B Ai mostly Capex Breakdown and full 2026 deployed compute breakdown: – Of the $18.4B total capex: $1.17B funded space operations (Starship, Starlink, production scale-up, launch infrastructure), while the majority, $15.8B, went into AI compute deployment. The $15.8B AI spend added 400 MW of nameplate capacity (1.0 GW → 1.4 GW), equivalent to ~220k GB300s (~130 kW per 72-GPU NVL72 rack) at $5.18M per deployed rack, a 25%+ discount to typical market deployment costs. – Total installed GPUs reached: 990k–1M (from ~770k+ in Q1: 220–230k Colossus 1 + 550k+ Colossus 2) note earlier Q1 filings did not fully reflect Colossus 1. – 2 GW+ year-end target per Elon: Reaching 2 GW of nameplate capacity requires 600 MW more deployed. Using Q2 averages ($3.95B per 100 MW for 55k GB300 GPUs), this implies $23.7B (23.7% of cash at hand) for 600MW in additional spend over the next two quarters which would add 330k GB300s & potentially including early Rubin systems, still at below market prices per deployed rack. – Projected year-end GPU count: 1.32M+ GPUs (using 2.3 GW): 220k+ in Colossus 1 550k+ in Colossus 2 550k in Macrohard's All of this totals ~3M+ H100-equivalents, making it the largest known AI compute deployment globally. – Compute Leasing: 450–465k GPUs already committed to Anthropic, Google, and Reflection AI, plus capacity for the new $20.8B+ unspecified cloud deals, estimated at 400k+ GPUs. Total to be leased: 860k+ GPUs, generating $48.6B ARR (with $27.8B confirmed annualized). – Remaining for SpaceXAI: 400k+ mostly GB300s (~1.2M H100-equivalents) for internal AI workloads and additional rentable capacity.