Clear · ΔT > 5 °C
- Komorebi hybrid92
- Dry cooler + adiabatic88
- Dry cooler + chiller95
- Radiative panels48
The case
The application that gets picked.
Cooling performance
The interconnect approval queue decides who wins compute capacity — not GPUs. Undifferentiated 50 MW interconnect filings wait. You are not buying solar. You are buying an existing grid connection and filing a load the operator can say yes to. Texas (Pecos) is the worked example. The same two clocks exist on every grid: power out, and permission to pull power in after dark.
- CFO
- Economics
- CTO
- Cooling & generation
- Counsel
- Permitting
$150M/year on 50 MW. +$42M on 64 MW. +$54M on 82 MW. Recip 22 MW.
50 MW holds on the hottest design day. 54 MW is the modeled meter draw at PUE 1.08, not a site cap — versus 62.5 MW at PUE 1.25.
Use the generation permit you already have. Dirty cap 54 MW unchanged. Recip 22 MW.
Why 50 MW holds when outdoor air is as hot as the cooling loop.
Pecos, Texas · grid-tied · existing PV interconnect · Phase 1
Comparison of three cooling solutions for 50 MW of AI compute
Two comparisons, same Pecos farm. First: an already-connected solar deed becomes a 50 MW IT host — a site with no interconnect and no 24/7 heat sink this year is 100% of nothing in IT offtake. Second: 8.5 MW less parasitic load versus PUE 1.25 at the same 50 MW IT (54 MW modeled meter draw at PUE 1.08, not a site cap). The farm stays on the grid. Reciprocating engines are not in this table; they are a later option.
| Metric | Dry cooler + chiller | Dry cooler + adiabatic | Komorebi Phase 1 | V% |
|---|---|---|---|---|
| PUE | 1.25 | ~1.18 | 1.08 | −14% |
| Overhead at the meter | 12.5 MW | ~9 MW | 4 MW | −68% |
| AI compute hosted | 50 MW | 50 MW | 50 MW | — |
| Draw from PV + grid | 62.5 MW | 59 MW | 54 MW | −14% |
| Annual cooling water | ≈ 0 | ~73 Mgal/y | ≈ 0 after fill | — |
| Plant noise at 10 m (est.) | ~80 dBA | ~70 dBA | ~55 dBA | −25 dB |
| Cooling electricity OPEX | $3.8M/y | $2.8M/y + water | $0.2M/y · pumps only | −95% |
| CAPEX per MW of AI compute | ~$1.5–2.0M / MW | ~$1.2M / MW | ~$0.72–0.78M / MW | ~−50% |
Overlay $6.4–7.5M. $36–39M to close the 50 MW hall ($0.72–0.78M/MW). Dry-cooler and adiabatic columns are mechanical yards. UPS/electrical of the hall is in the 4 MW overhead, not in this plant. OPEX at $35/MWh. AI at $3M per MW-year is a planning offtake, not a quote.
Eight-day Pecos screen on the radiator campus: 50 MW firm served in the window, PV held at 100 MWdc, 50 acres reserved for tables. A full-year radiator-campus run is not yet the published design target. Prior whole-field radial year is not this plant. A 10 MWdc pilot — about 5 MW of firm IT — is the step that validates hot-day performance, parasitic demand, and ground-temperature behavior at commercial scale.
Plant noise is a 10 m estimate at the cooling cluster, not a property-line model. Chiller yards ~75–85 dBA (compressors + condenser fans). Adiabatic dry coolers ~64–73 dBA per unit at 10 m in manufacturer data; ~70 for a yard. Komorebi is pumps under radiator rows — dry coolers as backup only — ~50–60 dBA outside a pump enclosure. −25 dB is about one-sixth as loud.
Pecos, Texas temperature envelope
Monthly air temperature. Left axis °C, right axis °F. Highs ~17 °C / 63 °F in January to ~37 °C / 98 °F in June–July.
- High
- Mean
- Low
Clear vs cloudy hours
Share of time. Clearer hours help dry and radiative duty; cloud still matters in winter. NSRDB 2024.
- Clearer
- Cloudier
More compute. Naturally.
Cooling performance
Pecos, Texas · challenging weather, firm rejection
In Pecos, Texas the hard hours are not the average hour. Clouds close a purely radiative sink. A dry cooler cannot go below ambient. A chiller can, but it spends the watts you wanted for compute. Phase 1 dumps 50 MWth on 40–50 acres of radiator tables, Ø800 under those rows only, and a heat-pump cold bank — so the loop still rejects when approach is zero and the sky is no longer a window.
Range across the year
Lower is better. Monthly cooling electrical demand to hold 50 MW of AI compute. Median marked.
Komorebi hybrid
3–5.3 MW, median 3.73.7MW
Dry cooler + adiabatic
7.3–11.4 MW, median 8.48.4MW
Dry cooler + chiller
10.2–15.6 MW, median 11.811.8MW
Radiative panels
2.9–5 MW, median 3.53.5MW
Bar is the inner half of months. Whisker is the full monthly range. Number is the median, MW.
Does the sink still work when weather turns?
Firm heat-rejection capability, 0–100. A chiller always rejects — and spends the watts you wanted for compute.
Cloudy · ΔT > 5 °C
- Komorebi hybrid86
- Dry cooler + adiabatic82
- Dry cooler + chiller93
- Radiative panels22
Clear · ΔT ≈ 0
- Komorebi hybrid78
- Dry cooler + adiabatic28
- Dry cooler + chiller90
- Radiative panels62
Cloudy · ΔT ≤ 0
- Komorebi hybrid70
- Dry cooler + adiabatic18
- Dry cooler + chiller88
- Radiative panels8
Tight band
Soil inventory and a heat pump buffer the Pecos, Texas summer peak.
Cheap on kWe
Radiative demand looks low until clouds remove the sink.
Pecos, Texas problem set
Clouds plus low or negative approach. Hours that decide available cooling — and therefore compute.
Heat-pump storage
Banked lift keeps rejection available at ΔT ≤ 0 when radiative falls off.
Left: electricity to hold the same cooling duty, in MW. Right: whether rejection remains available. Illustrative Pecos, Texas envelope — not a live plant meter.
Tracker height · vs fixed tilt
PV tracker fills the shoulders. Cooling tracker fills noon, stall, and night.
June 17, Pecos, apparent solar time (noon is 12:00). Same cassette. A SAT PV table buys morning and evening kWh (32% POA at 09:00 vs 25° south, 2% at noon). Inverse-track radiators do the opposite: they turn off the beam when the fixed rack is hottest. Stall-hour shed is 32% above a 25° fixed-tilt table; night zenith still 9% above a rack that never faces the sky. Civil clock in June is CDT — solar noon there is ~13:50.
Cooling · kWth per module · June 17 solar time
kWth
- Tracker · inverse / zenith
- Fixed tilt · 25° south
- 12–16 solar
PV analogue · POA W/m² · SAT vs 25° south
W/m²
- SAT tracker
- Fixed 25° south
- 12–16 solar
Band is 12–16 solar (post-noon heat). Axes are apparent solar time — noon is 12:00, not CDT. Air still lags ~2 h, so the fixed rack’s worst hour is ~14:30, not solar noon. The tracker inverse-tilts off the beam. Classic SAT: +32% at 09:00 solar, only +2% at noon. Cooling does not want that extra beam — it inverse-tracks, so the gain shows up where PV gain disappears.
| Hour | Tracker | Fixed tilt 25° | Tracker / fixed |
|---|---|---|---|
| Stall · 18,239 modules on 40 acres | 1.47 kWth · 26.7 MWth | 1.09 kWth · 19.9 MWth | 1.34× |
| June night (HP2 charge) | 2.78 kWth · 50.7 MWth | 2.53 kWth · 46.1 MWth | 1.10× |
| Wind at the fins (stall) | 3.13 m/s | 2.09 m/s | 1.50× |
| Sky view | 0.79 stall · 0.90 night | 0.75 all hours | zenith vs 25° |
| Air at the cassette | 44.6 °C stall | 47.1 °C stall (+2.5 K skin) | superadiabatic |
| Acres to close 50 MWth stall (plates only) | 75 acres | 101 acres | 1.3× land |
| Acres to close 50 MWth with HP2 cold storage | 40 acres | 43 acres | 1.1× land |
Physics mix · same energy balance
How the heat actually leaves.
Most of the heat leaves on the wind — air moving over the SPACECOOL lattice and the crown of the buried pipe. A smaller share radiates straight to the cold sky from those same surfaces. The rest soaks into the soil through the buried half of the pipe. Sunlight on the fins is extra heat to dump, not cooling, so it is not in the percentages.
- Convective 69% / 58.5 MWth yr-mean
- Radiative 16% / 13.9 MWth
- SPACECOOL ε=0.95
- Conductive 15% / 12.6 MWth
- buried Ø800
MWth · stacked monthly
18,239 radiator modules, 47,530 m Ø800 (radiator rows only), Pecos NSRDB monthly envelope. Shares are year-mean of each mode ÷ (convective + radiative + conductive). Convection is the main cooler. July is the stall trough — do not flatten it. Solar on the fins is omitted from the stack.
Height
Log wind: tracker 3.13 m/s vs fixed rack 2.09 m/s. Near-ground air is 2.5 K hotter on a stall afternoon. Convection is most of the kW.
Rotate
Inverse 40° turns the lattice off the beam (solar load 19 W vs 54 W) and opens a 1.12× venturi on the fins. Night goes to 0° — cold sky, also the 18 m/s stow.
What the fixed rack loses
25° never inverse-tracks and never goes to zenith. Axis at 0.6 m vs 1.38 m. Sky view 0.75. Soil 10 K above air radiates back.
Why Alion · AST 1.2
The tracker these modules actually hang on.
A cassette or a four-lens table is heavier than a dry 700 Wp module — glycol in the risers, glass in the optic. A torque-tube tracker puts that mass on one tube and one motor. Alion’s A-frame splits the load onto two feet and two north–south purlins. The box section is shorter. Wind pressure on the lattice does not have a long, empty span to work.
Distributed load
Two rail feet, not a single torque tube. Ø800 under cooler rows is ballast and header. The table does not cantilever a fluid-filled cassette off one pipe.
11:1 gearing
Inverse 40° at stall, zenith at night, 18 m/s stow. The reduction is what turns a heavier, wet module without a bigger drive. Same ±60° family as the PV tables.
Two N–S purlins
Headers nest in the purlin web. The east–west box is short. That is the wind story for a cassette with ten open fins — not a long torque-tube chord.
Screen, not CFD. Authority: juneDiurnal() + compareMounts(), finRatio = 3, 40 acres, 18,239 modules. Do not use TTP. Stall 1.47 vs 1.09 kWth (1.34×); night 2.78 vs 2.53 (1.10×); 75 vs 101 acres plates-only; 40 vs 43 acres with HP2 cold storage. Existing stall-kW file still uses the un-height-corrected cassette for the 50 MWth close — this table is the mount delta only. Ø800 half-buried pair is tracker ballast; a fixed-tilt rack has no drive tube and no dual-use foundation.
Land · same fenceline
Same 100 MWdc fenceline. Denser PV. Ten percent of tables become the cooler.
A 2015 tracker field uses about 5 acres per MWdc. This case re-modules that fenceline with 700 Wp bifacial (~3.13 ac/MWdc, GCR 0.35). Fifty acres (~10%) convert to radiator tables. Remaining PV stays generation. Leftover acres (~137) can take Phase 2 CSP — aperture in the ground, not a commissioned power block. TES and ORC stay Phase 2.
On 500 acres: 100 MWdc PV (~313 ac) + 50 ac radiators + leftover for Phase 2 CSP (~137 ac). CSP is not the Phase 1 cooler. Remaining PV stays generation.
The ground
An optional sink. Not the whole plant.
Phase 1 cooling is the radiator reservation. Pipe sits under those rows only. Ground coupling beyond that is site-specific — soils, year-round heat balance, and more than one summer of measurement. In a warm desert, an atmospheric sink may beat the ground. Dry coolers stay as fail-safe. The ground is an optional sink, not the platform.
Pecos, Texas · 100 MWdc existing PV · 50 MW firm AI
How the 50 MW is served across a day
Komorebi Phase 1 does not create the 50 MW. It stops cooling from consuming it. On an existing 100 MWdc farm that already has 2-hour batteries, parasitics fall from 12.5 MW of compressors to 4 MW of pumps — 8.5 MW continuous, 204 MWh/day. That is the headroom at the interconnect for 50 MW of 99.9% firm AI, and it is the midday surplus that charges heat-pump cold storage instead of feeding chillers. Phase 2 books TES and +14 MWe on the same radiator loop — 64 MW of IT, $42M/year incremental. No dry-cooler condenser. Phase 3 spends 430 °C exhaust into HT TES on extra halls: +18 MWe, 82 MW total, $54M/year incremental. Recip stays 22 MW. Dirty cap 54 MW unchanged. Not the 111/154 interconnect-off stack.
Peak PV 76 MW AC. Dry coolers stall — outdoor air is nearly as hot as the ambient cooling loop, so they cannot reject 50 MW of server heat. The MPC has already run the heat pump and topped up cold storage. Midday surplus after the 50 MW load goes into cold storage first, then the batteries.
Cooling stall (what operators call approach temperature): the gap between the hot cooling loop and outdoor air. When that gap shrinks toward zero, dry coolers stop moving heat. The heat-pump cold bank covers those hours.
Phase 1 — cooling first
This chart is the +50 MW. The dashed 50 MW line is filled all day on the existing 100 MWp interconnect: PV while the array is up, 2-hour BESS after sunset, grid only overnight. Midday surplus charges heat-pump cold storage so cooling does not consume the 50 MW. Komorebi draws 54 MW at the meter; the hatched band is the 8.5 MW a chiller would add (62.5 MW) — and on this stall day dry coolers would not hold the load.
PV to load 628 MWh · BESS discharge (after sunset) 100 MWh · BESS charge 102 MWh · Grid 472 MWh
After 17:00 the bank dumps into the hall — stall, surplus gone. After 20:00 outdoor ΔT returns and the field recharges it overnight. No night compressor on the grid.
- PV to load
- PV available
- Midday surplus → cold store + BESS
- Chiller tax (8.5 MW)
- BESS discharge (after sunset)
- BESS charge
- TES / ORC
- Generators · 54 MW recip.
- Grid
- CSP to TES
- Heat Pump Cold Storage
- Generator heat to TES
- 50 MW firm
+50MW
additional firm AI
Cooling: heat pump in service — dry coolers stall
Phase 2 — TES booked · +14 MWe
Phase 2 books TES here. Recovered server heat and CSP charge the tank. ORC is 14 MWe — 64 MW of IT, +$42M/year at $3M per MW-year. Same radiator loop. No dry cooler as the ORC sink. Dry coolers stay fail-safe only.
PV to load 556 MWh · BESS discharge (after sunset) 88 MWh · BESS charge 102 MWh · TES / ORC 275 MWh · Grid 281 MWh
- PV to load
- PV available
- Midday surplus → cold store + BESS
- Chiller tax (8.5 MW)
- BESS discharge (after sunset)
- BESS charge
- TES / ORC
- Generators · 54 MW recip.
- Grid
- CSP to TES
- Heat Pump Cold Storage
- Generator heat to TES
- 50 MW firm
+14MW
CSP/ORC uplift to the 50 MW · AI heat recovered
Cooling: heat pump in service — dry coolers stall
Phase 3 — extra halls · 82 MW
Phase 3 is extra halls from clean MWe, not a residual plant that holds 50 MW. 430 °C exhaust goes direct to HT TES. +18 MWe, 82 MW of IT, +$54M/year. Recip stays 22 MW. Dirty cap 54 MW unchanged. 111 MW / 154 MW is a different stack — the interconnect coming off — and is not this walk.
PV to load 556 MWh · BESS discharge (after sunset) 38 MWh · BESS charge 102 MWh · TES / ORC 431 MWh · Generators · 54 MW recip. 175 MWh · off-grid
Midday surplus PV charges the pack today. We discharge after sunset to displace genset MWh: stored PV is cheaper than fuel, hours, and permit burden on the recip fleet.
- PV to load
- PV available
- Midday surplus → cold store + BESS
- Chiller tax (8.5 MW)
- BESS discharge (after sunset)
- BESS charge
- TES / ORC
- Generators · 54 MW recip.
- Grid
- CSP to TES
- Heat Pump Cold Storage
- Generator heat to TES
- 50 MW firm
50MW
islanded off-grid · 22 MW recip shaft after TES/ORC · 54 MW fleet without TES/ORC — the permit problem
Cooling: heat pump in service — dry coolers stall
22 June Pecos, 100 MWp conversion farm, inverter clip 76.2 MW AC. PV to the 50 MW load; surplus first into the heat-pump cold bank, then batteries. Phase 2 TES/ORC is recovered server heat (~6 MW around the clock) plus stored CSP at 27.5% ORC. Phase 3 drops the grid. TES/ORC from CSP, recovered heat, and exhaust covers most of the night; recip shaft is the residual (~22 MW this day). Without TES/ORC you would install and run the full 54 MW fleet — 62.5 MW if the hall were chilled. Heat-pump cold storage is a 30 MWth bank; it is cooling, not electrical supply to the 50 MW. A planning day, not a forecast.
The farm arrives with 2–4 hours of BESS. Charts use the 2-hour floor: 50 MW / 100 MWh. Summer charges from midday surplus after the heat pump. Winter Phases 1–2 charge overnight from cheap grid. Discharge is after sunset only — never while the array is producing. Phase 3 is islanded: summer still charges from today's surplus; winter cannot, so that discharge is carry-in from a prior sunny day. The pack is never charged from the genset.
Dotted gold is unclipped PV. Gold fill above 50 MW is surplus: first into the heat-pump cold bank (COP 3.5, 30 MWth — the steel-blue band on the electrical chart), then the 100 MWh battery. The hatched red band at 54–62.5 MW is the chiller tax — 8.5 MW you would draw every hour, and do not. Heat-pump cold storage is cooling, not electrical supply to the 50 MW.
Compared with
Heat has to go somewhere.
Four substitutions a facilities team will offer. The stall day is the test — not PUE in isolation.
Warm-water liquid cooling and dry coolers
That path can reach a similar PUE on a campus that already has power. On a stall afternoon outdoor air matches the loop. Dry coolers stop. Adiabatic buys those hours with water the permit often will not give.
More batteries
Batteries store electricity. They do not reject heat. Every MWh spent running chillers is a MWh that never reaches GPUs.
Chillers on midday surplus
Noon PV can cover a 62.5 MW chiller hall. After sunset the array is dark. On stall hours dry coolers cannot dump 50 MWth.
CDUs and liquid-to-chip
They move heat out of the rack. They still need a place to dump 50 MWth when outdoor air equals the loop. That place is the 50-acre radiator reservation.
Why the overlay beats a new plant
- Lower capex per additional MW of AI than recips or simple-cycle GTs — and no new water right or chiller-yard air permit
- Reuses the interconnect and land you already have, or makes a campus you already permit hold more IT
- The array stays PV. A 50-acre radiator reservation is the 24/7 cooling plant — then AI compute on the interconnect you already have
- If overnight service is delayed — or the air permit cannot take 54 MW of engines — Phase 2 TES/ORC is already in, so the islanded plant is ~22–28 MW, not a 54 MW engine yard