Pick a target compute scale and watch what it would actually take to build it in space — power, cooling, material, deployment time. Then compare engineered Dyson-scale construction against what one IMBH ergosphere does for free.
Best viewed on desktop or tablet.
Read the description instead →Compute (amber). Sets how many operations per second you want, what each op costs in energy, what fraction of time you're running (duty cycle), and whether the operations are reversible. Reversible computing (Bennett 1973, Fredkin & Toffoli 1982) avoids the Landauer thermodynamic floor of k_B T ln 2 per irreversible bit erasure, at the cost of needing more steps per logical operation in practice — the tool flags the toggle but doesn't auto-apply a serial-time penalty.
Cooling (teal). Sets the radiator temperature (passive Stefan-Boltzmann) and optionally an active cryocooling target below it. If T_compute < T_rad, the cryocooler has to pump heat against the gradient at a Carnot tax of (T_rad/T_compute − 1) / η_carnot Watts of input per Watt of heat moved.
Power (purple). Three sources. Solar PV: panel area scales as 1/d² with chosen efficiency. Fusion: assumes a reactor with chosen W/kg power density — engineering fiction beyond ~10³ W/kg. BZ-IMBH: gravitational power extraction from a spinning hole; engineered mass goes to zero because the substrate is the black hole itself.
MTH callout (red). Computes Lloyd's bound on the rest-mass energy of any IMBH and shows the ratio to your engineered compute target — concretely, how many of "you" the ergosphere is in compute terms.
Required ops/s = N × 3.17×10²¹ × (1/duty). Energy per op clamped to k_B T_rad ln 2 unless reversible. Compute power P_c = (ops/s) × (J/op). Cryocooler power = P_c × (T_rad/T_compute − 1) / η_carnot (zero if T_compute ≥ T_rad). Total power P_tot = P_c + P_cryo. Radiator area A_rad = P_tot / (0.9 σ T_rad⁴). Power-collector area: solar = P_tot / (S(d) η_pv); fusion = 0 (collector area is moot); BZ = 0. Mass densities: radiator 1.5 kg/m², solar panel 5 kg/m², fusion reactor by power density, BZ source zero. Build time = total mass / launch cadence.
Structural mass beyond radiators and collectors. Manufacturing energy. Communication latency across the array. Self-replication / von-Neumann manufacturing (which would change "build time" by orders of magnitude). Specific accretion-disk dynamics for BZ extraction (see Tool 1 for that calculation). Beam-down / power transmission losses if collector and compute aren't co-located. Treat absolute numbers as order-of-magnitude correct; treat scaling with N as exact.
The Macro Transcension Hypothesis argues that civilisations near or past Kardashev II would prefer to compress their computation inward toward a black hole rather than expand outward across stars, because the compute capacity per kilogram at the Lloyd bound on M·c² is many orders of magnitude higher than anything engineered substrates can achieve. This tool makes that argument concrete in commensurable units — surface area and material mass — rather than the non-commensurable units (parsecs vs ops/m³) that the v1.0 STEM Compression visualiser used and that the v1.1 spec replaces.
The transistor was invented at Bell Labs in December 1947. Cumulative transistors ever manufactured is somewhere between 10²² and 10²⁴ — the most-cited industry figure around 2018 was "13 sextillion" (≈ 10²²); another seven years of fab expansion plus the AI accelerator boom puts the 2026 figure closer to 10²³. For operations, you have to integrate exponentially growing throughput over time — and because compute doubles every ~1.5–2 years (Koomey's Law for efficiency, Moore's Law for density), the integral is dominated by the most recent few doublings. Current worldwide aggregate compute throughput is roughly 10²² ops/s (Hilbert & Lopez 2011 put 2007 at 6.4×10¹⁸ MIPS; ~17 doublings have happened since). Integrated over the past, that totals roughly 10²⁹–10³⁰ operations — a number that, if performed in a single year, would require ~3×10²¹ ops/s. That's the "1×" baseline this tool uses. Today's top supercomputer (El Capitan, late 2024, 1.74 exaFLOPS) does ~1.7×10¹⁸ FLOPS by itself; the Bitcoin network at ~7×10²⁰ hashes/s is roughly 7% of all human compute by raw throughput.
Lab record holders for photovoltaic conversion efficiency sit around 47.6% (NREL six-junction III-V concentrator cell, 2020). Production-grade space cells (Spectrolab XTJ Prime, AzurSpace 4G32C) deliver 32–34% AM0 efficiency in service. The Roll-Out Solar Arrays (iROSA) installed on the ISS in 2021–2023 deliver ~250 W/m² in LEO at ~33% efficiency and ~150–200 W/kg specific power — the highest-deployment-grade space arrays currently flying. Best commercial silicon is ~24% (SunPower IBC) at ~5 kg/m². At the tool's defaults (30% efficiency, 1 AU), you get ~410 W/m² usable — so a 1 GW orbital datacenter needs about 2.4 km² of panels, weighing ~12,000 tonnes at iROSA mass density.
Spacecraft thermal control today is dominated by white-painted aluminum honeycomb panels using Z93 paint (ε ≈ 0.9, α ≈ 0.15) or Optical Solar Reflector quartz tiles (ε ≈ 0.8, α ≈ 0.07). The ISS Photovoltaic Thermal Control radiators dump about 70 kW total across ~50 m² each, mass ~150 kg per panel — so roughly ~1.4 kW/m² at 270 K with ~3 kg/m² specific mass. The Stefan-Boltzmann ceiling at 200 K (deep-space passive default) is about 81 W/m²; at 300 K, ~410 W/m²; at 500 K, ~3,200 W/m². For active cryocooling, the best modern Stirling and pulse-tube coolers achieve 10–30% of Carnot efficiency at moderate temperature spans. The James Webb sunshield is the gold-standard passive cryocooler — its five-layer kapton-with-aluminum design drops detector temperatures from ~370 K (sun side) to ~40 K (cold side) with no moving parts.
Compute is useless if you can't save the outputs. As of 2025, the largest enterprise SSDs are Solidigm D5-P5336 (~60 TB) and Samsung PM1743 (30 TB) in 2.5" form factor, drawing ~10 W idle / ~25 W active. Energy per bit written is around 10⁻¹⁰ to 10⁻¹¹ J/bit in current QLC NAND — about 10¹⁰ times above the Landauer floor at 300 K (3×10⁻²¹ J/bit), so storage is nowhere near its thermodynamic limit. HAMR-based HDDs (Seagate Mozaic 30+, WD UltraSMR 32 TB+) match SSDs on capacity at ~5 W and one-tenth the throughput. Worldwide installed storage at end-2024 is around 10²³ bits (~10 ZB) and growing by roughly 5×10²² bits/year. For the MTH framing: storing the output of a 10⁶× Earth-equivalent compute run at modern density would itself require ~3×10⁵ exabytes of new SSD per year — comparable to multiple years of total worldwide flash production.
World electricity consumption averages about 3×10¹² W (~25,000 TWh/yr ≈ 10²⁰ J/yr). Total world primary energy is ~6×10²⁰ J/yr. Worldwide datacenters currently consume about 500 TWh/yr (~1.8×10¹⁸ J/yr, ~2% of electricity) and that share is rising rapidly with AI. The GPT-4 training run was estimated at ~50 GWh = 1.8×10¹⁴ J — about a single second of worldwide electricity output. At 1× in this tool (≈ 3×10²¹ ops/s × today's J/op), required power is ~3 GW continuous — roughly one large nuclear plant. At 10⁶×, you're at 3×10¹⁵ W ≈ 5,000 times current world electricity consumption — which is why the mass-comparison strip starts moving past asteroid-belt territory by that point. The MTH argument is precisely that this is where engineered substrate breaks down and the IMBH ergosphere starts looking like the only sensible option.
At the tool's defaults (N=10⁶, J/op=10⁻¹², T_rad=200 K, solar at 1 AU, 30% PV, 100% duty), you should see roughly: 3 PW of power, ~4×10⁷ km² radiators + ~8×10⁶ km² panels, ~10¹⁴ kg of structure (about ten-millionths of Vesta), and a build time of ~10 million years at today's launch cadence of ~10⁷ kg/yr. Slide the launch cadence up to Starship-era 10⁹ kg/yr and that drops to 10⁵ years — still longer than recorded human history. That gap between "scale required" and "deployment cadence achievable" is the practical engineering argument that complements the thermodynamic one — and the reason the ergosphere MTH comparison at the bottom of this tool isn't a rhetorical flourish.