Correction package — 9 August 2026
Working note on scope. This file contains revised text for three named sections of the published post, plus a new appendix. It is not the complete post — the surviving sections on CXMT, the DUV mass-production move, the ByteDance/Alibaba directive, the CDS convergence and the Chapter 6 field-note test are unchanged and should be retained as published. Splice the blocks below in at the marked headings.
Changelog
| # | Claim as published | Status | Revision |
|---|---|---|---|
| 1 | 50–70 Starship flights per GW | Optimistic corner | ~80 flights/GW (72–95 range) |
| 2 | $50–100bn annual market per GW | High by 5–10x | ~$10bn per GW-year |
| 3 | Terawatt fails on launch physics | True, but weaker than available | Terawatt fails on demand first |
| 4 | “This is the SpaceX investment thesis” | Inverted | Orbital is the option; terrestrial is the thesis |
| 5 | Both blocs at 1–30 GW orbital through the 2030s | Upper bound unsupportable | 1–10 GW |
Corrections 1, 2 and 5 are downward revisions to the corpus’s own numbers. Correction 3 is an upgrade — it replaces a contestable engineering argument with one that cannot be rebutted by improving the vehicle. Correction 4 is the one that matters for positioning.
SPLICE A — replaces the block beginning “What the gigawatt thesis looks like”
What the gigawatt thesis looks like — the corrected investment frame
Approximately 80 Starship flights deliver 1 GW of orbital compute, which generates roughly $10 billion in annual revenue at observed pricing.1
The flight count first. SpaceX’s regulatory filing figure of 100 kW of compute per tonne implies 10,000 tonnes per gigawatt, which at a 200-tonne reusable payload gives the 50–70 flights this corpus previously carried. But the disclosed AI1 specification is 70 kW per tonne, not 100 — and the realistic mid-period Starship payload is 150 tonnes, not 200. That combination gives 14,300 tonnes per gigawatt and 95 flights. The honest planning range is 72 to 95. Use 80.
The revenue figure is the material correction, and SpaceX’s own second-quarter 2026 disclosure settles it. The company ended the quarter with 1.4 GW of installed compute capacity generating $2.56 billion of AI segment revenue — an annualised $7.3 billion per gigawatt-year. Google’s Colossus lease at $920 million per month implies $11 billion per gigawatt. A bottom-up build from accelerator counts and cloud rates gives $9.4 billion. Three independent methods, all converging near $10 billion per gigawatt-year.2
One gigawatt of orbital compute is therefore a $10 billion annual revenue line, not a $50–100 billion one. Set against SpaceX’s $1.5 trillion market capitalisation, gigawatt-scale orbital compute is a rounding error — meaningful as a demonstration of the architecture, immaterial as a valuation driver this decade.
The terawatt frame fails twice over
The launch arithmetic is the familiar objection. One terawatt at 14,300 tonnes per gigawatt is roughly 80,000 Starship flights.3 Starship has flown thirteen times. A launch every day would take more than two centuries, and to compress it into a decade would require twenty-two launches a day, every day, with the entire fleet doing nothing else.
But the demand arithmetic kills it first, and more cleanly. McKinsey projects global AI data centre demand at 156 GW by 2030, with total data centre capacity — AI and conventional combined — reaching 219 GW. The IEA’s base case puts all data centre consumption at 945 TWh annually, about 108 GW continuous.
One terawatt is more than six times what the entire planet is forecast to want.
This is the stronger argument, and the corpus should lead with it. The launch objection can always be answered by a better rocket, a cheaper tonne, a higher power density. The demand objection cannot. No improvement to Starship creates a customer for 1,000 GW of compute in a world that wants 156.
SPLICE B — replaces the block beginning “The revised orbital compute investment thesis”
The revised orbital compute investment thesis
Not: SpaceX reaches terawatt scale this decade and reshapes global civilisation.
Also not: orbital compute is the SpaceX investment thesis. It is not, and the second-quarter 2026 numbers say so without ambiguity. SpaceX’s AI segment grew 247% to $2.56 billion with zero satellites in orbit. Every dollar came from terrestrial sources — hosting contracts with Google and Anthropic, Grok subscriptions, X advertising — running on the same grid, the same silicon and the same circular financing the BIS flagged in Source 41. The first two AI1 prototypes fly in early 2027. Commercial constellation deployment begins in 2028.4
Correct: SpaceX reaches gigawatt-scale orbital compute in the early 2030s at roughly 80 flights per gigawatt, adding a $10 billion annual revenue line to a business whose 2030 revenue will be determined by Starlink subscribers and ground-based data centres. China’s Three-Body constellation reaches comparable scale on a parallel timeline. Both blocs operate at 1–10 GW orbital through the 2030s.5
Terawatt scale is the correct long-run civilisational frame. Gigawatt scale is the correct decadal frame. Neither is the reason to own the equity.
Where the constraint actually sits
The binding constraint was never launch cadence. It is that AI compute demand, while enormous, is finite: roughly a $1.5 trillion market in 2030 — 156 GW at $10 billion per gigawatt — contested by six trillion-dollar balance sheets.
In that contest, the reliable economics sit with the toll collectors rather than the contestants.
| Return on invested capital | Capex per $1 of new revenue | |
|---|---|---|
| Nvidia | +105% | $0.06 |
| Vertiv | +38% | $0.18 |
| Microsoft | +26% | $2.31 |
| SpaceX | –15% | $5.66 |
The firms closest to the money creation absorb the most of it per unit of output, and deploy it into assets with four-to-six-year depreciation lives. The firms selling into that spending — silicon, thermal, power — earn 38 to 105 per cent on capital while doing so. This is the Cantillon structure of Chapter 6 read at the level of the income statement, and it does not favour the contestant.
There is a second-order point that cuts the same way. This corpus’s compute-deflation thesis — the mask over the debasement — works against the seller as well as the buyer. Median energy per text query has fallen from roughly 2.9 watt-hours in 2024 to 0.24–0.3 watt-hours, close to a tenfold efficiency gain in two years. Gigawatts deployed can rise while revenue per gigawatt falls. The $10 billion figure is a 2026 anchor and the corpus’s own thesis implies it declines.
SPLICE C — replaces the final clause of the one-line synthesis
Structural note. The synthesis as published is a single sentence of roughly four hundred words. It now carries four corrections and one reversal, which is more than the device can hold. The recommended break point is after “but the clock is wrong by orders of magnitude” — everything after that is a separate finding and reads better as one.
…but the clock is wrong by orders of magnitude.
The correction has three parts. Terawatt-scale orbital compute is not a decadal investment thesis and not merely for want of flights: one terawatt exceeds total forecast world AI demand in 2030 by more than six times, a constraint no improvement to the vehicle can relieve. Gigawatt scale by the early 2030s is the correct frame, at approximately eighty flights per gigawatt rather than fifty to seventy, generating roughly ten billion dollars annually rather than fifty to one hundred — a figure now anchored to SpaceX’s own disclosure of 1.4 GW producing $2.56 billion in a quarter. And orbital compute is not the SpaceX investment thesis at all: the AI segment grew 247 per cent with no satellites in orbit, which means the near-term variable is terrestrial capacity, grid interconnection and fab allocation, not launch cadence.
Both blocs are nonetheless executing Hamiltonian stealth QE toward the achievable gigawatt target — China through IPO-permission equity inflation and SOE mandates in the correct Hamiltonian sequence, state first and private capital within it; the United States through CHIPS Act subsidies and SpaceX national security contracts in the inverted sequence, private capital ahead of the CLARITY Act framework that would codify it.
The corrected corpus instruction: position for gigawatt-scale orbital compute in the early 2030s as an option rather than a thesis, recognising that the near-term revenue of the orbital-compute complex is terrestrial and grid-bound; reduce memory chip and mature-node ASML exposure to the Chinese competition confirmed today; increase energy, thermal and silicon infrastructure exposure as the true binding variable for both blocs — the toll layer, not the contestants; maintain hard outside money held directly rather than through levered listed proxies, whose funding models depend on the same credit conditions the hedge is meant to escape;((A listed bitcoin treasury vehicle carrying perpetual preferred at 10–12 per cent, funded from capital markets rather than operations, is not an exit from the monetary system this corpus describes. It is a leveraged participant in it that happens to hold the exit asset.)) and build ERC-8004 before the orbital geometry window closes — because in orbit the watcher is unreachable at any price, and the agent that carries its own verified identity is not a governance preference but the only architecture that works when the radio goes silent.
APPENDIX — the arithmetic, so it can be checked
Physical assumptions
| Parameter | Value | Basis |
|---|---|---|
| AI1 compute power | 150 kW peak / 120 kW average | SpaceX disclosure |
| AI1 power density | 70 kW per tonne | SpaceX disclosure |
| AI1 mass (derived) | ≈ 2.14 t | 150 kW ÷ 70 kW/t — derived, not disclosed |
| Mass per gigawatt | 14,300 t | 1,000,000 kW ÷ 70 kW/t |
| Starship payload, mid-period | 150 t reusable | V3 target range 100–150 t |
| Satellites per flight | ≈ 70 | 150 t ÷ 2.14 t — see caveat |
| Radiator area per GW | 0.71 km² at 1,400 W/m² | unchanged from Chapter 6 |
Caveat on the derived mass. SpaceX has published AI1’s power and its power density but not its mass. The 2.14-tonne figure divides one by the other. If the 70 kW/tonne specification refers to average rather than peak power, AI1 masses 1.71 tonnes and 88 fit per flight rather than 70. Treat both figures as mid-range estimates.
Flights per gigawatt
| Payload per flight | @10,000 t/GW (filing) | @14,300 t/GW (disclosed) | @17,900 t/GW (avg power) |
|---|---|---|---|
| 100 t | 100 | 143 | 179 |
| 150 t | 67 | 95 | 119 |
| 200 t | 50 | 72 | 90 |
| 250 t (expendable) | 40 | 57 | 72 |
The previously published 50–70 sits in the top-right corner of the optimistic case. Planning figure: 80.
Revenue per gigawatt-year — three methods
| Method | Implied $/GW-year |
|---|---|
| SpaceX Q2 2026: 1.4 GW producing $2.56bn | $7.3bn |
| Google Colossus lease at $920m/month | $11.0bn |
| Bottom-up: 1 GW ÷ 1.4 kW ≈ 715,000 accelerators × $1.50/hr × 8,760h | $9.4bn |
| Planning figure | $10bn |
Unit economics per Starship flight
| Line | Value |
|---|---|
| Satellites per flight | ~70 |
| Power deployed | 10.5 MW |
| Annual revenue per flight | ~$105m |
| Lifetime revenue (5-year satellite life) | ~$525m |
| Hardware cost (70 × ~$8m) | ~$560m |
| Launch cost | $30–100m |
| Simple payback | ~6 years, against a ~5-year asset life |
At disclosed specifications and 2026 pricing, orbital compute does not yet clear its cost of capital. Three things could change that: payload mass doubling, silicon cost per megawatt halving, or satellite life extending. SpaceX is pursuing all three, and the Nvidia Rubin partnership addresses the second.
Demand ceiling
| Forecast | 2030 global demand |
|---|---|
| McKinsey — AI workloads | 44 GW (2025) → 156 GW |
| McKinsey — all data centres | 82 GW → 219 GW |
| IEA — all data centres | 945 TWh ≈ 108 GW continuous |
| Implied 2030 AI compute market | ~$1.5 trillion (156 GW × $10bn) |
| One terawatt as a share of that | 641% |
Provenance
Revision v2, 9 August 2026. Corrections 1, 2 and 5 revise figures first published in Source 45. Correction 3 supersedes the launch-physics argument with a demand-side argument of the same conclusion and greater robustness. Correction 4 reverses the positioning implication of the section as published.
Financial figures are drawn from SpaceX’s Q2 2026 report (4 August 2026), McKinsey’s global data centre capacity analysis, and the IEA Energy and AI base case. The AI1 mass figure and all per-flight economics are derived rather than disclosed and are flagged as such above. The $10 billion per gigawatt-year anchor reflects 2026 pricing and, by this corpus’s own compute-deflation thesis, should be expected to decline.
Nothing in this post is investment advice.
- Revised 9 August 2026. As first published, this section stated 50–70 flights and a $50–100 billion annual addressable market. Both figures were wrong in the same direction and the revenue figure was wrong by roughly an order of magnitude. The corrected arithmetic is set out in the appendix. [↩]
- The published $50–100 billion figure applied retail per-GPU-hour list pricing to bulk-leased fleet capacity. Wholesale capacity does not clear at retail rates; the gap is the error. [↩]
- The published figure of 50,000 flights used the more generous 10,000 tonnes-per-gigawatt assumption. At the disclosed AI1 density the true figure is higher. [↩]
- This is the correction that most changes positioning. A reader of the post as published would have concluded that Starship cadence is the variable to track. It is not. Through 2028 the variable is terrestrial compute capacity, which is constrained by grid interconnection and Nvidia’s fab allocation, and has nothing to do with launch. [↩]
- Revised down from the published 1–30 GW. The upper bound implied one bloc’s orbital segment alone absorbing roughly a fifth of forecast world AI demand — not supportable before the 2040s. [↩]