How this dataset is built, what the numbers mean, how we audit our own work, and where the limits sit. Last revised July 2026.
ArusGrid is an independent tracker of data center development across Southeast Asia, built to answer five questions about AI capacity. Who is building, who funds it, who buys it, how much power each site carries, and what that power costs. The name comes from arus, the Malay and Indonesian word for current. Malaysia, Singapore, Indonesia, Thailand, Vietnam and the Philippines receive full research depth. The other five countries in the region are monitored from a distance because no verifiable facility there has crossed our size threshold. Including them today would add rows without adding information.
A facility qualifies at 20 MW of planned or built IT load. Smaller facilities also qualify when a hyperscaler or AI tenant is confirmed. We exclude legacy enterprise colocation below 20 MW, corporate server rooms and telco exchanges, which means our country totals will read lower than surveys that count every telco room. Cancelled projects stay in the dataset because the graveyard carries signal, and they are excluded from every total on every page.
announced → land_permitting → financing_closed → under_construction → commissioning → operational → expanding, with terminal states cancelled and on_hold. A campus with both live and building phases is expanding.
A stage at or beyond under_construction requires a dated source saying construction actually happened. Land purchase plus regulatory approval does not qualify. A groundbreaking that was scheduled does not count until a source confirms it took place.
All megawatt figures aim to capture IT load, the power available to computing equipment. Press coverage routinely mixes IT load, total facility power and grid connection capacity for the same site, and one campus can appear at 180, 120 and 100 MW depending on which basis a writer picked up. Where sources conflict we take the conservative IT load figure and log the contradiction. Where only another basis exists we keep the number and tag its basis.
| Tier | Standard of evidence |
|---|---|
confirmed | A dated source shows liquid cooling, racks above 40kW, or GPUs actually deployed. |
likely | Strong signals support an inference. Examples include a recent hyperscaler self-build, a confirmed AI or hyperscaler tenant, a modern greenfield above 100MW, or a stated AI purpose. The signal is always recorded. |
unlikely | The design predates 2020, or the site is a small legacy colo focused on enterprise or interconnection work. |
unknown | No signal either way. Most operators never disclose cooling specs, so an honest unknown beats a guessed yes. |
Data center markets are catchments of power, fiber and land. Countries are the political rollup. Johor serves Singapore overflow hyperscale demand while Klang Valley serves domestic colocation. Batam sells latency to Singapore while Jakarta sells data residency to Indonesians. We track 16 markets across the six countries and match the market definitions used by the firms that publish vacancy, so their published figures stay comparable with ours.
Vacancy is the share of built, leasable colocation capacity sitting empty, with hyperscaler self-build excluded. Five of our 16 markets have a published figure. The rest show a blank and we never estimate one. A vacancy number also needs its cause before it means anything. Johor sits near 1 percent because demand absorbs everything built. Singapore sits near 5 percent because regulation caps supply. Greater Manila sits near 44 percent because capacity jumped 31 percent in six months and the new stock has yet to lease up. Three different stories hide behind what looks like one metric.
Facility level electricity prices are almost never public. Unless a facility has a cited public power purchase agreement, its power cost is the country industrial tariff and is labelled as such. The annual electricity bill shown on the map multiplies planned MW by 8,760 hours and the tariff. It assumes full IT load, excludes cooling overhead and is always labelled an estimate.
Every research batch is followed by an adversarial fact check. An independent pass reverifies sampled entries claim by claim against primary sources and benchmarks our coverage against third party trackers. Corrections go into the dataset and the audit files are kept as receipts.
| Batch | Sample | Claim error rate | Dominant failure found |
|---|---|---|---|
| Phase 1 (MY + SG) | 10 entries | ~5% | status inflation |
| Phase 2 (ID + TH) | 12 entries, ~105 claims | 9.5% (material 2.9%) | status inflation, stages and tenants recorded ahead of the evidence |
| Phase 3 (VN + PH) | 31 entries | 4 to 5% | status staleness, entries running behind the operators' own later statements |
Across all audits we found zero fabricated sources, zero banned aggregators cited and zero fabricated facilities. The two failure directions shaped two standing rules. A stage only moves up on a dated source, and a stage only freezes after a search for later operator statements. Coverage benchmarking cuts both ways as well. One audit showed we were undercounting Thailand and added 16 facilities we had missed.
Few trackers publish their own error rate. We do it because a reader deserves to know how wrong a dataset tends to be, and in which direction.
Spotted an error, or an announcement we missed? Corrections with a dated source are welcome. A contact channel for the project will be posted here.
The dataset and visualizations are licensed CC BY-NC 4.0. Share and adapt freely with attribution for noncommercial purposes. Commercial licensing is available separately.