On Earth, the power runs out first. A few hundred kilometres up, the Sun almost never sets.
Our orbit dawn-dusk sun-synchronous, ~650 km, 98° tilt · 12 rings of 8 MW satellites
The catch: orbit isn't free. A megawatt in orbit only costs the same as the cheapest site on Earth once satellites get lighter. That is our Design 2.
build and run for 5 years
for all our Earth sites together
to put anything in orbit
to the first word, chat · live agents
everyday prompters · 30 chats a day
agent developers · 12 live + 30 background agent tasks a day
knowledge workers · 20 chats, 1 live + 4 background tasks
An agent works on a task by itself: think, run a tool, read the result, repeat, about 80 times. Live agents have a person waiting; background agents get 4 hours.
Both serve 100% of the 4.45 billion users and survive losing any one site or orbital plane, even on the worst eclipse day of the year.
of the money is chips, in both designs
per user, for all 5 years
Design 1's Earth part, 6 sites
Design 1's orbit part, ground stations included
Rack prices and sizes are the organizers' (slide 12). Rack counts come from our search, which tried every chip type and group size for every job and kept the cheapest mix that answers in time.
Energy is what the sites draw over 5 years: computers at their simulated load, plus cooling (PUE 1.07–1.14, slide 14). Carbon is that energy × each grid's CO₂ per kWh, 237 g on average.
Set so orbit ties the Nordics at a $95/kg launch, as on slide 14. Design 2's lighter hardware costs 8% less.
| Planes | Crosses equator | Eclipses | Design 1 | Design 2 |
|---|---|---|---|---|
| 1, 3 … 11 | 06:00, dawn | June | 7 sats · 50 MW | 10 sats · 75 MW |
| 2, 4 … 12 | 18:00, dusk | December | 7 sats · 50 MW | 10 sats · 75 MW |
144 t · 20,600 m² of solar wings · 8,600 m² of radiators · chips at 85 °C
88 t · 18,900 m² of solar wings · 12,400 m² of radiators · chips at 58 °C
Inside every site and satellite: reader chips take in the whole question at once, writer chips write the answer word by word, a memory store holds agents' work while their tools run, and sandboxes run their code.
| Test | What it does | Design 1 | Design 2 |
|---|---|---|---|
| 1 · Match the organizers | reproduce the organizers' own numbers before trusting the model | 11 of 11 | |
| 2 · Every rule | re-check each rule from the raw design file | 39 of 39 | 33 of 33 |
| 3 · By hand | redo key numbers with plain arithmetic, no model code | 16 of 16 | – |
| 4 · Simulated days | whole days minute by minute: normal, outages, eclipses | 57 of 57 | 39 of 39 |
| 5 · Wrong estimates | each of 16 estimated inputs made 25% worse, then re-optimized | fits 9 of 16 | fits 10 of 16 |
organizers' slide 13 → ours
slide 13 → ours
slide 14 → ours
per kg, slide 14 → ours
| Simulated, minute by minute | Days | Live users served | Longest background wait |
|---|---|---|---|
| Design 1 · normal day · lose any one of 18 places | 1 + 18 | 100% every time | 55 · 85 of 228 min |
| Design 1 · worst eclipse days, alone and with each outage | 38 | 100% every time | 200 of 228 min |
| Design 2 · normal day · lose any one of 12 planes | 1 + 12 | 100% every time | 44 · 81 of 218 min |
| Design 2 · worst eclipse days, alone and with each outage | 26 | 100% every time | 81 of 218 min |
Background agents use the live-agent racks in quiet hours: $4.8B saved. The world is below average demand 13 hours a day.
Many small places instead of a few big ones. One outage takes out a small slice, so the backup is small.
Losing a big orbit section dropped 37% of live users. Spreading over 12 rings cost about $3.0B.
| Input | Ours | What the source says | Verdict | Source |
|---|
At today's electricity, carbon and building prices together, Design 1 would cost about $31.2B, over budget. Design 2 uses none of these inputs and stays at $29.44B. Rising Earth costs are one more reason for orbit.
python3 twin/prove.py
github.com/MadhanJ05/Orbital_DataCenter
madhanj05.github.io/Orbital_DataCenter/twin3d.html