Twelve arms, one core
Rack rows run outward along twelve arms from a central core. Each arm is its own power and cooling domain, so a fault stays in its arm. The wedges between arms are where the facade does its work.
Dossier 01 · Concept brief · October 2026
A data center that cools itself with its own walls.
Twelve arms of racks around a central flue. Warm-water liquid cooling to the chip. The building's facade is the heat rejection. No cooling towers, no evaporative pads, no chiller yard. A planning PUE of about 1.08 to 1.11 in any climate, and effectively zero water on site.
1.08
annual PUE, temperate site
1.11
annual PUE, Arizona-type heat
0.00 L/kWh
on-site water, every climate
0 to −0.27
Grid PUE with turbines in the core
Goal PUE < 1 · WUE = 0 · Grid PUE < 0
PUE as defined cannot fall below 1, because the building's own IT load is inside it; the 1.08 to 1.11 above is where the cooling design stands. So we add two honest numbers beside it. Energy reuse effectiveness (ERE) credits heat sold at 82 °C and can go below 1. Grid PUE is electricity bought from the grid, minus electricity sold to it, divided by IT energy: 0 when the building powers itself, negative when it sells. With the turbines in the core (section 05) the planning Grid PUE is 0 to about −0.27.
01 · The basic design
Shape the building like the problem.
Heat in a data center is made in the racks and has to leave through the skin. Starburst puts every rack a short walk from the skin and gives every arm its own cooling, power and containment.
Four flat floors, a tall flue
Four storeys of about seven metres stacked on the same plan. The core is an open flue: an updraft path for relief air and smoke, so the building never depends on a fan to stay safe.
Liquid to the chip, warm
Racks are at least 95 percent liquid-cooled, fed at 40 to 45 °C and returning at 52 to 57 °C. Warm water is the whole trick: the hotter the loop, the more of the year the outside air can take the heat without a compressor.
The facade is the heat rejection
Dry coils hang on the wedge facades between air-intake bays. Rack heat goes rack, to an in-row distribution unit, to the arm loop, to the wall. Nothing leaves the building but warm air.
The chiller moved indoors
On hot days a compressor is still needed. Instead of a chiller yard, the compressors sit in a walkable machinery band directly behind the facade coils, a few metres from the loop they serve. The same coil is a dry cooler on cool days and a condenser on hot ones, and the same machine lifts the heat to 82 °C when a neighbour wants it.
A kit, not a construction project
Rack rows, distribution units, wall modules, pump modules and power rooms repeat identically in every arm on every floor. Build one arm, learn it, build forty-seven more.
02 · Why it is better
What you get for building it this way.
A
Zero water on site
No towers, no pads, no humidifiers. The only water is what the staff drink. Communities fighting data centers over water have nothing to fight about here.
B
Best-in-class PUE without a cold climate
Planning PUE of about 1.08 at a temperate site and 1.11 in Arizona-type heat. The industry average is 1.56 and the best published fleet sits at 1.09. Starburst gets there because the loop runs hot, so the compressor only works on the hottest hours.
C
No chiller yard, no plume, less noise
The compressors are indoors. The fans are on the facade. There is no acre of chillers humming next to someone's house, and no visible plume because nothing evaporates.
D
Heat worth selling
The same indoor compressor lifts the loop to 82 °C district-heating water, several megawatts of thermal per arm floor. Export is reported honestly as an energy reuse factor beside PUE, not hidden in it.
E
Faults stay small
Every eleven racks is its own contained loop with its own isolation and leak response. Every arm has its own power and cooling. A problem is an arm's problem, never the building's.
F
Built from a repeating kit
Forty-eight identical arm floors. Factory-tested distribution units, pump modules and wall modules. Faster to build, faster to commission, and the second building is cheaper than the first.
03 · PUE and WUE
The numbers, and where they come from.
Planning arithmetic on one Compact arm floor: 5.8 MW of IT, of which 5.43 MW leaves as liquid and 0.37 MW as air. Overheads are pumps, coil fans, the indoor compressor when the weather calls for it, the air modules for the small air load, and electrical losses. Planning values, not measurements.
| Outdoor air | Overhead | Compressor | PUE |
|---|---|---|---|
| 20 °C, free cooling | 447 kW | 0 | 1.08 |
| 30 °C, early trim | 636 kW | 133 kW | 1.11 |
| 40 °C, design day | 945 kW | 427 kW | 1.16 |
| 45 °C, full lift | 1,364 kW | 836 kW | 1.24 |
| Site | Annual PUE |
|---|---|
| Temperate (northern US, north-west Europe) | about 1.08 |
| Hot and dry (Arizona) | about 1.11 |
| Hot and humid (Gulf coast) | about 1.10 |
| Industry average, 2024 (Uptime Institute) | 1.56 |
| Best published fleet average (Google) | 1.09 |
Water usage effectiveness, on site
0.00 to 0.02 L per IT kWh
Glycol make-up, domestic use and fire-system tests, in every climate. The industry runs near 1.8 to 1.9 L/kWh (Uptime figure, as cited); the best large fleet reports 0.80 (as compiled).
The honest caveat: water in the electricity
Off site, the water is whatever the grid uses. On a US-average grid, roughly 3.1 L is consumed per kWh generated (estimate, as cited), which is about 3.4 L per IT kWh at our PUE. Wind, solar or dry-cooled generation drives that toward zero. Anyone who says "zero water" without this sentence is selling something.
What moves the numbers: every megawatt of trim costs 130 to 155 kW, so compressor hours are the whole gap between 1.08 and 1.11. A double-conversion UPS on the compute would add two to four points. Heat export does not lower PUE; it is reported beside it.
05 · Power: the flue is a stack
The building makes its own electricity.
The central flue was always the relief and smoke path. It is also a stack. Gas turbines sit at grade in the core, and their exhaust rises through an insulated liner inside the flue. The hot liner warms the surrounding relief air and strengthens the natural draft the whole building leans on. No diesel generators, no waiting years for a grid connection, and the grid tie becomes a sales line rather than a lifeline.
| Item | Value |
|---|---|
| Building load at PUE 1.09 | 303 MW |
| Turbines, N+1 of 54 MW aeroderivative class | 7 × 54 = 378 MW |
| Grid PUE, six units running | −0.08 |
| Grid PUE, all seven running and selling | −0.27 |
| Fuel, simple cycle (41 percent) | about 2,800 MMBtu/h |
| Fuel, combined cycle with dry cooling (56 percent) | about 2,050 MMBtu/h |
| Exhaust up a 12 m liner, six units | about 780 kg/s at 500 °C, 15 m/s |
| Heat left in that exhaust above 120 °C | about 320 MW thermal |
| Source of the 303 MW | CO₂ per year | per IT MWh |
|---|---|---|
| Simple-cycle turbines on natural gas | about 1.30 Mt | 0.53 t |
| Combined cycle, dry cooled | about 0.95 Mt | 0.39 t |
| Buying from the average US grid (EIA, 0.81 lb/kWh) | about 0.98 Mt | 0.40 t |
Gas factor 52.91 kg CO₂ per MMBtu (EIA); turbine class figures from public aeroderivative data (53.8 MW, 42 percent, 498 °C exhaust). Simple cycle is dirtier than an average grid; combined cycle is about even; a clean grid beats both. Dry low-NOx combustors, no water injection, so WUE stays at zero.
Why it belongs here
Behind-the-meter generation has gone from niche to mainstream because grid connections take years (Cleanview). Starburst already has the stack, the core and the heat loop; the turbines are the missing piece, not an add-on.
Where the exhaust heat goes
Three hundred megawatts of 500 °C exhaust is too valuable to waste up the flue. Combined cycle turns it into another 100 MW of electricity; or it drives the trim and the 82 °C export so the compressors go quiet. That choice is open item 11.
What it costs
A gas pipeline sized for 2,000 to 2,800 MMBtu/h, an air permit for a 300 MW plant, a liner tall enough to disperse NOx, and the carbon column above. The site that wants this is the site with gas and without grid. Everywhere else, the flue stays a flue.
04 · Open items
Things we have not figured out yet.
This is the part of a brief nobody publishes. We are publishing it. If you have built, commissioned, broken or fixed something like this, we want your thoughts. Leave an email if you want a reply; leave it blank if you would rather not.
Submissions go to the inventor only. Nothing you write here is published. Anything you send becomes part of the project record; please do not send anything you are not free to share.
Footnote: Stephen DiMaria is running out of AI credits, but input is appreciated and will be run against the AI council once per month. Sprint 3 is scheduled for October 31st, 2026.