—
Facilities with known coordinates (ATLAS, PeeringDB, OpenStreetMap). DayOne sites are approximate park locations. Johor's open record is thin, so the large campuses there show up through Epoch, not the facility map.
Epoch AI tracks four SIJORI campuses from satellite imagery and permits. DayOne's two Johor campuses lead; ByteDance-linked Bridge MY06 at Sedenak is the nearest rival.
Status from the F-1 and press. Power and H100-equivalents are Epoch estimates where the campus is tracked.
| Site | Market | Status | Power (MW, Epoch) | H100e (Epoch) | Note |
|---|
Eighteen metro markets. Facilities and operators come from the merged open facility table; IXPs and networks from PeeringDB; cables from TeleGeography; AI MW from Epoch; climate from NASA POWER 2015–2025; grid carbon from Ember (latest year). Click a market to see its operators. Sort by any column.
Facility count per operator group (top 12). Names are normalised across sources, so a few groups may still split.
Interconnection (networks present, from PeeringDB) against subsea cables landing in the market. Dot size is the facility count. Johor and Batam borrow Singapore's interconnection, so they sit low here despite large AI builds.
Wet bulb sets how hard evaporative and chiller plant work. SIJORI sits near 26 °C all year with no cool season. NASA POWER's grid puts Singapore and Johor in the same cell.
PUE = 1.10 + 0.012 × (mean wet bulb − 12 °C), floored at 1.10. This is a simple stated model for comparing markets, not a facility measurement.
EMA's Uniform Singapore Energy Price. The 2021–22 gas shock tripled prices, and 2026 is climbing again. DayOne passes power through to customers, but higher prices weigh on customers' all-in cost per MW.
Ember yearly data. Indonesia's coal-heavy grid makes Batam the hardest market for hyperscaler carbon targets. Malaysia is high too, which makes green-energy deals in Johor a selling point.
Each dot is one PDU or medium-voltage domain: mean versus 99th-percentile draw as a share of provisioned power, from Google's PowerData 2019 (BigQuery). Room under the 100% line at p99 is capacity that sits stranded unless it is oversubscribed.
Mean and 10th–90th percentile utilisation across all domains by hour of the trace day. The swing is small: cloud load is flat, so committed power is drawn almost all day.
Defaults come from the data above: DayOne's F-1 figures, Singapore's USEP, Google's fleet load factor and the climate PUE model. Change any input to test a market or a deal. FX and the server overhead are assumptions. Results are estimates for discussion, not a valuation.
Chip thermal design power (W) from Epoch's ML hardware database. Each generation pushes rack density up, which is why new AI halls in Johor are designed around liquid cooling.
Each market gets four factor scores from 0 to 1, scaled across the 18 markets. Set how much each factor matters, and the ranking updates. Scores compare markets with each other; they are not absolute grades.
Bar length is the weighted score; segments show what each factor contributes. Markets where DayOne already operates are outlined.
The raw factor scores behind the ranking, 0 (worst of the 18) to 1 (best), in the current rank order.
| Rank | Market | Connectivity | Climate PUE | Grid carbon | AI demand | Weighted score |
|---|
Data centers (c) Ringmast4r - Global-Data-Center-Map · PeeringDB · OpenStreetMap contributors (ODbL) · Epoch AI (CC-BY) · TeleGeography (CC BY-SA) · NASA POWER · Ember (CC BY 4.0) · EMA Singapore · Google PowerData 2019 (CC-BY) · DayOne F-1 (SEC, Oct 5, 2026).