Orbital Data Center Model
The Starbench Orbital Data Center Model provides first-principles unit economics for compute in orbit: a coupled power-thermal-mass-launch solver producing WACC-levelized $/PFLOP-hr against the ground-datacenter baseline — plus the tracked field and two in-browser simulators running the same physics. Launch costs ride capacity-basis anchors from the Launchonomics stack — list price (or the operator's own cost) over the payload capacity that price buys, price and cost never pooled — ground power comes from EIA tariff data, and delivered compute is derated to dense FLOPS at realistic training utilization.
The headline
The model's current answer, stated up front (values as of the 2026-09-01 refresh — the solar-basis rework that flipped the parity call; every number below ships with a P10/P50/P90 fan inside):
- Orbit is ~1.4× ground for a merchant buyer today. A 1 MW B200-class platform prices at $11.28 /PFLOP-hr [9.72 – 12.44] against $7.81 for the marginal ground data center — at the Falcon 9 capacity-basis launch anchor ($4,229/kg: list price over the reusable capacity it buys), on the mass-silicon PV basis. A vertically integrated operator paying its own launch cost prices at $7.80 — level with ground.
- Launch price decides the rental race; the panel still guards the owners'. Published silicon price lists collapsed the space-PV term, so against the ground lease breakeven launch is a real ~$1,050/kg and P(parity ≤ 2040) is 0.22 merchant / 0.65 vertically integrated (central crossing ~2032). Against a hyperscaler's own cost ($2.56/PFLOP-hr), even $0/kg launch leaves orbit at $4.42 — not reachable at any launch price on that quadrant. The prior ≈0 call stays frozen in the ledger and will be graded as made.
- Who pays for launch decides the sign. Merchant buyers ride the F9 list price ($4,229/kg, flat for sixteen years); a vertically integrated AI-sat operator pays marginal cost ($1,038/kg today, $802/kg on the mature-Starship cost cell) — price and cost are never pooled, and every cost table carries the split. At fleet scale the communications ceiling still binds before economics does.
- The calls are public. Nine frozen forecast calls — cost parity by 2040, LCOC in 2032, on-orbit PFLOPs milestones, GW on orbit at 2030/2035/2040 — graded in public, misses included.
Orbit vs ground, at three scales
A sample of the comparison table subscribers get live (sortable, with the CSV behind it):
| Scale | Orbit $/PFLOP-hr — P50 [P10 – P90] | Ground | Premium | Launch-price breakeven |
|---|---|---|---|---|
| 10 kW | 25.30 [23.74 – 26.46] | $7.81 | 3.2× | not reached |
| 1 MW | 11.28 [9.72 – 12.44] | $7.81 | 1.4× | ~$1,052/kg |
| 1 GW | 11.27 [9.71 – 12.43] | $7.81 | 1.4× | ~$1,061/kg |
Merchant-buyer rows against the market-lease ground basis (F9 capacity price vs what a renter pays). Every cost table also carries the other quadrants — the vertically-integrated cut ($7.80 at the operator's own launch cost, level with the lease) and the modeled_own_cost ground basis ($2.56/PFLOP-hr bottoms-up owner TCO, the SemiAnalysis zone), against which orbit is 3.0–4.4× and unreachable even at free launch. Cost compares to cost, price to price: orbit can beat renting; it does not touch owning.
A sample simulator run
The read-only baseline from Build a satellite — 1 MW, NVIDIA B200, LEO sun-synchronous:
mass to orbit 15,438 kg · shared two-sided panel 7,407 m² · binding constraint solar · capex $0.16B · design life 5 yr · one launch · LCOC $11.28 /PFLOP-hr → 1.4× ground
Inside, every one of those figures is an input — GPU, orbit, solar cell, coating, radiator temperature, launch price down to $0/kg — with mass, yield, cost, and the verdict recomputed live on a JS port of the solver, parity-checked against the Python model.
Who uses this
- Investment diligence — pressure-test an orbital-compute pitch against physics-anchored unit economics and the funding register of every player in the field
- Corporate strategy — a defensible answer to "should we care?", with the levers that would change it stated and quantified
- Supplier positioning — where the mass, power, and thermal budgets actually sit, and which lanes (thermal above all) are empty
- Policy and program screening — flying-to-filed discipline on announced constellations, China programs scored separately, claims dated and sourced
What's inside
- The solver: power, thermal, mass, and launch cost solved together, with P10/P50/P90 fans on all headline numbers and breakeven contours down to the free-launch limit
- The forecast ledger: nine frozen calls — cost parity by 2040, LCOC in 2032, on-orbit PFLOPs milestones, GW on orbit at 2030/2035/2040 — graded in public, misses included
- The tracked field: space hyperscalers and pure-plays with per-card funding, valuation, and FCC-filed fleet size, plus a subcategorized picks-and-shovels register; per-operator satellite spec cards; per-player demo manifests; a China program scoreboard tracked flying-to-filed
- Model compilations: Nanjangud, Turyshev, ABI, Google, and the rest reconciled against the Starbench stack, disagreements stated rather than smoothed
- Two simulators: build a satellite and build a constellation in the browser, on a JS port of the solver parity-checked against the Python model
What subscribers receive
- The $/PFLOP-hr scenario model with breakeven contours, refreshed quarterly
- The mass-budget teardown and sensitivity analysis
- Ten raw CSV downloads — unit economics, LCOC anchors, orbit vs ground, yield waterfall, workload fit, specific-mass league, plausibility grades, fleet half-life, manifest ceiling, cost trajectory
- All physics inputs tied to resolvable source ids
Two ways to read Starbench
- Substack subscription — the newsletter plus the State of Launch board. It does not include this model.
- Institutional seat — the full model: ten gated sections, both simulators, the ten CSV downloads, quarterly refresh. Request a demo and we reply within one business day, typically with a time-limited demo seat followed by a scope and pricing conversation.