Soupy Lab // Technical Disclosure v1.0 // April 2026
Subterranean Convective
Cooling Tower
A passive cooling and siting architecture for pulsed-workload AI compute infrastructure.
Author — John "Soupy" Beene · Soupy Lab LLC · Glen St. Mary, Florida
with cooperative AI collaboration — Claude (Anthropic)
§1 · Subterranean cooling tower
Architecture Overview
Abstract
A passive cooling architecture for AI compute infrastructure that exploits the naturally bursty thermal signature of significance-gated cognitive systems (SACE).
A half-buried vertical tower draws cold intake air through long underground trenches that equilibrate to stable ground temperature, passes that pre-conditioned air upward across the compute array, and exhausts waste heat through a chimney stack whose convective draft is pumped by the workload's own pulsed thermal profile.
Solar panels above the intake shade the inlet grates and provide both electrical power and an upper warm-air layer that assists exhaust evacuation.
The design inverts conventional AI datacenter siting: it performs best in hot, sunny climates with stable subsurface temperature.
Fig. 01 — Primary architecture / cutaway elevation
§2 · Pulsed convection + stack effect
Thermal Physics
SACE (pulsed) — pumps convection
CONCEPTUALConventional LLM (sustained) — just warms air
CONCEPTUALPulsed thermal pressure moves air through constrained channels; steady heat only warms it.
Stack Effect
ΔT across tower (°F) by climate zone
MODELEDHigher ΔT → stronger natural draft
Why this works
SACE architectures produce a thermal signature qualitatively different from conventional AI inference: ~200 ms cortex bursts separated by seconds of idle. At a rack, this becomes a 1–5 Hz pulse train with sharp leading edges.
Sharp thermal transients create pressure pulses, and pressure pulses move fluid through constrained channels far more efficiently than sustained warming — the same principle behind pulse-jet engines and some solar chimneys.
Combined with a tall stack (chimney effect ∝ √(H · ΔT)), deep earth-equilibrated intake, and a solar-heated upper layer, the architecture produces net airflow sufficient to displace a substantial portion of active cooling under favorable load profiles.
Exact values depend on site geometry, load mix, and local climate. All figures on this page are modeled from architectural principles, not deployment data.
Three operating modes — solar-preferred, safety-backed
Passive draft does the work in normal conditions. Optional induced-draft fans sit high in the exhaust path on bypass louvers, so they add no obstruction when idle.
- 01
PASSIVE
Passive draft
Chimney + geothermal intake + solar-warmed upper layer + workload pulses. No moving parts in the airflow path.
- 02
ASSIST
Solar-assisted induced draft
Low-power crown fans engage when draft margin drops below target. Solar-direct when sun is available.
- 03
PROTECT
Safety-backed override
Battery- or grid-backed forced ventilation for thermal safety events — heatwaves, dust storms, peak load after sunset.
Bypass louvers open in passive/assist; fan ring engages only when needed.
The fans are not the system; they are the seatbelt.
Fig. 02 — Thermal physics / flow dynamics
§3 · Top-down / radial trench layout
Site Plan
Fig. 03 — Top-down site plan, one tower
Notes
- • 6 radial trenches, 30 m each
- • Grates at grade, shaded
- • Trench depth 3–4 m (below frost line)
- • Tower 18 m above grade
- • Chamber 4 m below grade
- • ~1 hectare per tower
Legend
- Tower / shaft
- Subterranean trench
- Intake grate
- Solar panel cluster
- Exhaust (hot)
§4 · Four free inputs + one workload
Energy Stack
Four free energy inputs · one paid compute load
Layer 5
Compute Workload
The only paid input. SACE already filters 97% of invocations.
Layer 4
Solar Panel Array
Shade + generate + create upper warm layer.
Layer 3
Chimney Stack Effect
Tall exhaust shaft amplifies ΔT into natural draft.
Layer 2
Pulsed Convective Pumping
SACE workload burst profile → pressure waves → airflow.
Layer 1
Geothermal Cold Reservoir
Earth at 10–20 °C year-round. Infinite thermal mass.
Efficiency arithmetic
Conventional baseline
An 8-GPU rack dissipates ~3.2 kW continuously. Industry-standard cooling overhead adds roughly ~40 % of IT load — another ~1.3 kW per rack spent purely on heat rejection.
At datacenter scale, cooling represents a meaningful fraction of total operating cost.
SACE-served rack
At ~3 % cortex duty cycle, the rack averages a fraction of sustained load — and the profile is pulsed rather than steady. The pulse is the prerequisite for passive convective pumping, which can displace a portion of active cooling.
Combined with solar self-generation and climate-optimal siting, facility power draw falls below conventional AI baselines.
The architecture advantage is structural; specific numbers depend on deployment and are subject to load testing.
Illustrative baseline — values shown are modeled for conceptual comparison, not from measured deployment.
Fig. 04 — Energy stack diagram
§5 · Climate-optimal footprint
Siting & Conclusion
Where the architecture wants to live.
| Region | Fit | Notes |
|---|---|---|
| Arizona / SW US | A+ | High ΔT, dry, abundant solar, stable subsurface. |
| Saudi / Gulf | A | Highest ΔT; dust mitigation needed at intake. |
| N. Florida | B+ | Stable ground, humid; deeper trenches help. |
| Pacific NW | C | Cooler ambient compresses ΔT; smaller win. |
| N. Europe | C- | Low ΔT most of year; conventional cooling competitive. |
Conclusion
Conventional AI datacenter siting fights physics: it places sustained-load compute in cool climates, then spends a large fraction of total power moving heat out. A pulsed-workload architecture inverts the problem. The same thermal signature that makes SACE efficient at the model level also makes its infrastructure host efficient at the building level.
The cooling tower is not a separate invention; it is the natural deployment shape of a significance-gated cognitive engine.
Fig. 05 — Siting space, climate-optimal footprint
§7 · Compute as motion
Composable Architecture
The cooling tower is one expression of a more general principle: significance-gated cognition is naturally pulsed, and that pulse can carry the workload at every scale.
- 1BuildingPulsed convective cooling tower
Workload pulses pump the chimney's own draft.
- 2ChipHelical compute spine
Tiles fire as a wave climbing a spiral; geometry induces rotational airflow.
- 3WireWave-pipelined memory
Signal-propagation delay becomes storage. Data in flight is the working set.
Stop building machines that resist physics. Build ones that ride it.
Fig. 07 — Same pulse, three scales
Priority claim
The three elements above — pulsed convective cooling, helical traveling-wave compute, and circulating wave-pipelined memory — are disclosed here as a single composable architecture ("compute as motion") and reserved for further development. Engineering numbers are deliberately omitted; the claim is the structural relationship between scales, not specific tolerances.