Compared to the McCalip calculator and the Economist¶
The Economist ("Data centres in space: less crazy than you think", 2 Mar 2026) cites Andrew McCalip's web calculator showing a 1 GW orbital data center at near-parity with Earth ($11–16.7bn vs $15.9bn) under optimistic assumptions, while this package reports space at ~19× Earth for the bundled text-inference design. The two are not in conflict. The difference is mostly the metric and the slider positions, not the physics. This page decomposes it with numbers from the model (orbital_1mw_inference.yaml, orbital_mccalip_optimistic.yaml, earth_hyperscale_baseline.yaml); regenerate them with orbitdc compare.
Two different metrics¶
McCalip compares total capex per nominal capacity (GPUs excluded, capacity assumed ~98% usable = the daylight fraction). This package's headline is LCOC = lifecycle cost / delivered PFLOP-days, after the full degradation waterfall (the bundled space design delivers ~29% of nominal, Earth ~47%). The denominator difference alone moves the ratio by ~2×.
To compare like with like, the package exposes capex_per_w_ex_gpu (capex per watt of IT power, accelerators excluded) in the compare summary and Evaluation.details. McCalip's $16.7bn for 1 GW is $16.7/W on this basis.
The capex-per-watt decomposition¶
| Earth | Space, default | Space, McCalip-optimistic | |
|---|---|---|---|
| capex $/W IT (ex-GPU) | ~$12 | ~$524 | ~$212 |
| — launch | $0 | ~$236 | ~$11 |
| — satellite (solar+radiator+bus+comms) | n/a | ~$288 | ~$200 |
| facility | ~$12 | n/a | n/a |
| network factor | 1.00 | 0.75 | 1.00 |
| delivered fraction | 47% | 29% | 38% |
| LCOC $/PFLOP-day | ~$66 | ~$1,237 | ~$428 |
| LCOC vs Earth | 1× | 19× | 6× |
Three things stand out:
- Our Earth number matches McCalip's. $12/W ex-GPU is ~$12bn for 1 GW, in line with his $15.9bn terrestrial estimate. The disagreement is entirely on the space side.
- We agree that Starship closes the launch gap. Moving to the speculative $200/kg launch case drops launch from ~$236/W to ~$11/W — a ~20× reduction, consistent with the article.
- The unreproduced assumption is satellite cost. Even with optimistic sliders, the satellite stays ~$200/W ex-GPU, against Starcloud's claimed "less than $5/W." That ~40× gap is dominated by the costed power system (the solar catalog is $50–60/W, and the array is oversized for eclipse recharge and degradation), plus comms and integration. The package does not reproduce a $5/W satellite from its component costs; whether Starcloud can is the open question.
Why the optimistic case is still 6×, not parity¶
McCalip's calculator stops at capex per capacity. This package then divides by delivered compute, applying the waterfall McCalip omits: sustained-vs-peak software (0.55), reliability-adjusted availability (~0.82), and — for ground-dependent workloads — the optical-downlink weather availability (0.75). The McCalip-optimistic scenario removes the downlink penalty (crosslink-only operation, network factor 1.0) and still delivers only ~38% of nominal, so its LCOC is ~6× Earth even though its capex/W is down ~2.5× from the default.
What this means¶
At equal inputs the two models agree: at McCalip's own pessimistic defaults ($500/kg, 37 W/kg, $22/W) his calculator gives $51bn, ~3× worse than Earth, which tests/test_references.py::test_mccalip_orbital_is_several_times_costlier reproduces from physics. The near-parity headline needs three things at once: a speculative launch price, a satellite cost roughly an order of magnitude below what the component catalog costs here, and a capex-per-capacity metric that does not charge for the delivered-compute waterfall. Each is defensible to argue; the package's job is to make all three visible rather than fold them into a single number.
Run it yourself:
orbitdc compare examples/scenarios/orbital_mccalip_optimistic.yaml examples/scenarios/earth_hyperscale_baseline.yaml