SpaceX model explorer
An interactive reconstruction of Shaun Maguire's (Sequoia Capital) bottom-up SpaceX model, Q1'26 to Q4'28, which I have extended to 2033. It covers Starlink, then terrestrial compute (SpaceX Super Intelligence), then orbital compute. Solid lines are reported actuals (Q1 to Q3 2026). Dashed lines are projections. Figures in $ millions unless noted.
Total revenue by segment
Compute capacity (GW, end of quarter)
SI (AI) revenue lines
Connectivity: Starlink and Direct to Cell
Segment adj. EBITDA vs capex (total)
Change in cash
Quarterly values to Q4'28. For 2029–33 (shaded) each point is the average quarter (annual ÷ 4) so the scale stays the same. GW points are year-end. Shaded years are my extrapolation beyond Maguire's 2028 horizon.
Summary
FY2026 includes actuals for Q1 to Q3. The shaded columns (2029–33) are my extrapolation. "Implied EV" = EV/EBITDA multiple × Q4 EBITDA annualised (2026–28) or × annual EBITDA (2029+). This is an assumption I added, not something Maguire published. The values in brackets are Maguire's figures.
Maguire's 2020 Starlink model (Sequoia, Q4 2020)
In the article, Maguire published the internal Sequoia model he built in Q4 2020, during Starlink's first quarter of revenue. It is a capacity model: satellites launched each year, each cohort's download Gbps × "oversubscription" multiple, kept for a 5-year life, with retired satellites removed at the current year's Gbps × oversubscription, as in his sheet. That is followed by a "useful capacity" multiple, then divided by peak Mbps per customer to get the users that can be supported, then × ARPU. The dials below re-run his 2020 sheet. With defaults it reproduces his numbers, for example 52.1M supportable users and a $27.6B end-of-2026 run-rate.
Check against reality: his 2020 sheet forecast a $27.6B Starlink run-rate at end-2026. In SpaceX's reported Q3'26 results, Connectivity was $5.25B, about $21B annualised, and his 2026 model has Q4'26 Starlink at about $24B annualised. Source: his article (image 1). I use its logic, capacity per Starship launch with a 5-year life, to extend Connectivity past 2028.
How the model works
Maguire's sheet copies SpaceX's earnings layout: three segments (Space, Connectivity, SI), each with revenue, operating income, adjusted EBITDA and capex, plus an assumptions block covering launches and compute. With every dial at its default, this page reproduces his published numbers exactly. Each dial then changes the projected quarters (Q4'26 onwards) like this:
- Terrestrial GW: GWt = GWt−1 + (Maguire's ΔGWt × build-rate dial). His revenue is billed on the prior quarter's capacity: Terrestrial compute revt = GWt−1 × annual rev/GWt ÷ 4. For example, Q1'27 is 2.5 GW × $35B ÷ 4 = $21.9B. He uses $30B/GW in Q4'26, $35B in 2027 H1 and $40B from then on.
- SI capext = ΔGWt × capex/GWt. For example, Q4'26 is 0.8 GW × $45B = $36B. He steps capex/GW down from $45B to $30B. He says rev/GW and capex/GW are "coupled constants", so the coupling switch moves capex/GW with rev/GW.
- Orbital GWt = GWt−1 + orbital-compute Starship launches × MW per launch ÷ 1000. He has 10 to 25 MW per launch. Orbital revenue scales with average orbital GW in the quarter, calibrated to his figures.
- Starlink compounds his quarter-on-quarter growth rates (about 10 to 20% per quarter), multiplied by the growth dial. Direct to Cell, Grok + Cursor + Other, Space (launch) and Terafab are his figures times a scale dial.
- EBITDA = revenue × Maguire's implied segment margin for that quarter, plus the margin dial. SI margins run from about 60% to 75%, and Connectivity is about 65 to 70%. Operating income = EBITDA − his implied D&A, with SI D&A scaled by cumulative SI capex. Change in cash = his figure + ΔEBITDA − Δcapex.
2029–2033 (my extrapolation, not Maguire's): I extend his logic annually from his Q4'28 run-rates. Terrestrial GW additions start at his 2028 pace (10.8 GW a year) and grow at the chosen rate (default flat). Revenue and capex per GW start from his $40B and $30B and decline each year. Orbital compute ("star-mind") = orbital Starship launches (from his 2028 run-rate of 60 per quarter) × MW per launch (from 25), with revenue = average orbital GW × orbital rev/GW. Orbital capex/GW is my own assumption, because his sheet carries none. Starlink and Direct to Cell add revenue per Starship launch, following his article: about $1B a year per Starlink-equivalent launch. The DTC default of $0.5B per launch is calibrated so DTC overtakes Starlink around 2030, as he expects. Margins carry forward from his Q4'28 levels plus the shift dials. SI D&A = his 2028 ratio to cumulative SI capex. Cash = EBITDA − capex + his 2028 "other" leakage ratio.
Reconciliation note: in his sheet the SI total for 2028 is lower than the sum of its listed lines. Q4'28, for example, shows $201B against about $225B for the lines. The page keeps his published SI totals and spreads the gap in proportion to terrestrial compute.
His stated core assumptions: AI doesn't fully melt down, other compute suppliers miss their timelines while SpaceX delivers, and regulation slows SpaceX by only about 20%. He also expects SpaceX compute margins of 40 to 50% after depreciation. He says the widest error bars are on rev/GW, capex/GW and Terafab capex.
Sources
- Shaun Maguire, model post (30 Oct 2026): x.com/shaunmmaguire/status/2116261542052454704
- Article "People don't understand SpaceX": x.com/shaunmmaguire/status/2116260821080973679
- Q3'26 earnings table he copied: reply · core assumptions: reply · compute margins: reply
- SpaceX Super Intelligence: @SpaceXSI
Built by Simon Manning. All credit for the underlying models (2026–28 and the 2020 Starlink sheet) goes to Shaun Maguire. The years 2029–33 are my own extension.