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 the data center buildout, project by project. For every project it follows where the project stands in its construction lifecycle, who is building it, who funds it, who operates it, who buys the capacity, how much IT load it carries, the capital committed to it, and what its power costs. The buildout is the subject, and power is one dimension of it. The name comes from arus, the Malay and Indonesian word for current.
Coverage is Southeast Asia, the Gulf and India. Malaysia, Singapore, Indonesia, Thailand, Vietnam and the Philippines receive full research depth, as do the United Arab Emirates, Saudi Arabia and India. Qatar, Bahrain and Kuwait carry the few projects that clear our threshold. Oman carries none yet, because nothing verifiable there has crossed it, and the other five Southeast Asian countries are monitored from a distance for the same reason. Including them today would add rows without adding information. China entered in October 2026 and is being built hub by hub through the national computing hubs. Zhongwei in Ningxia, Horinger and Ulanqab in Inner Mongolia, Qingyang in Gansu and Gui'an in Guizhou are covered today. The other hubs follow. Many Chinese projects publish rack counts with no megawatt figure. We show those rack counts as published, label them as racks, and never convert them into megawatts. Rack sizes differ by project, so two rack counts are not always comparable.
India entered on 27 September 2026 with a first pass of 58 projects, and its figures carry the same caveat as the Gulf. Indian operators and state governments quote megawatts with no stated basis far more often than Southeast Asian ones, and sanctioned grid load, campus ceilings and memoranda of understanding are routinely reported as capacity. We record the committed or built figure, tag the basis where the source states one, and leave the ceiling in the project notes. Hyperscaler self-build in Hyderabad and Visakhapatnam publishes almost no capacity at all, so several of the largest Indian projects appear with no figure beside them. Electricity tariffs are set by each state, so no single country tariff is shown.
Gulf capacity figures understate the Gulf, and the reason is disclosure rather than research. Gulf operators rarely state whether a megawatt figure means IT load, grid capacity or total facility power, and many publish no figure at all. Where the basis is unstated we record the number untagged and say so on the project. Where no figure exists we leave it blank rather than invent one, so several real Gulf projects appear with no capacity beside them. The largest of them records only its committed first phase against a far larger announced programme. Treat Gulf planned capacity as a floor built from what operators have actually stated.
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.
Southeast Asian coverage is deeper than Gulf coverage, and the reasons are worth stating because they are not all the same kind of reason.
Part of it is simply time. Southeast Asia has been researched in successive passes with audits between them, so each project there rests on more citations than a Gulf one does. That part closes with work.
The count itself is a different matter. We list fewer Gulf projects than commercial directories do, but a sweep of the operators' own published estates found almost nothing above our size threshold that we were missing. Gulf directories, like Southeast Asian ones, are largely filled with small colocation halls that we exclude by design. Where we genuinely cannot see is stranger than that. The largest operator in the United Arab Emirates states that it runs thirty live data centres totalling 673 megawatts, and its website names two of them. Its own locations page does not exist. No amount of further searching resolves that, because the information has not been published.
The other part does not close. The layers of information that sit around a project, rather than inside it, largely do not exist in the Gulf. Six of our sixteen Southeast Asian markets have a vacancy rate published by a commercial research firm. None of the seven Gulf markets does, because no firm publishes a Gulf series at all. Every Southeast Asian country has an electricity tariff we can cite and defend. Neither Gulf country does: the two Emirates differ so widely that a single national figure would be misleading, and the Saudi regulator's own tariff pages could not be retrieved across repeated attempts through different tools. Where that is true we show a blank rather than a number, here as everywhere else.
What may surprise you is that individual Gulf projects are not more poorly documented than Southeast Asian ones. A larger share of them name a customer, and a larger share state whether a capacity figure means IT load. Gulf projects tend to be enormous and politically visible, so the ones that surface publicly are covered well. The Gulf tells you a great deal about a few buildings and very little about the market around them, and nothing at all about the ones its operators choose not to name. Southeast Asia is the reverse.
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. |
Announced pipelines mix real projects with paper. Sell-side coverage of the US market rates only about a third of its announced nameplate capacity as credible, and Southeast Asia is no different. Every active facility here therefore carries an evidence tier that grades how much verifiable support sits behind it. The tier grades a project's public paper trail. It says nothing about any company, and cancelled projects carry no tier at all.
| Tier | What it means |
|---|---|
verified | The facility is operational or expanding. Running capacity is its own evidence. |
corroborated | Construction or commissioning is evidenced by a dated source, or financing has closed with support, or several independent signals agree at an earlier stage. |
sourced | One or two hard signals sit behind the announcement. |
thin | The announcement stands alone, or the project is on hold. |
Tiers are computed by fixed rules rather than judgment calls, and every facility's detail panel shows the checklist behind its tier. Operational capacity is verified on its own. Reaching construction requires a dated source under our stage rules, so building sites are corroborated by definition. Everything earlier is scored on six signals. A named funder. A disclosed tenant, where a confirmed tenant counts double and a hyperscaler at its own self-build counts for nothing, since it is the operator. Committed, approved or borrowed capital. A dated and sized construction phase. Three or more cited sources. A cited source dated within the last 24 months. Projects on hold rate thin regardless of their other signals.
One limit is worth naming. Most citations recorded before mid 2026 carry no publication date, so the recency signal reads older entries as silent, and no rule can see private progress. Thin means the public record is empty. A real project with a quiet sponsor will rate thin until it produces evidence, and that is the tier working as designed.
A second limit matters more the further a project sits from Southeast Asia. The tier grades whether a project is real, not whether we know its size. A facility can be verified because it is demonstrably operating and still carry no capacity figure at all, because its operator has never published one. Read the tier as confidence that the project exists and stands where we say it stands, never as confidence that the megawatts beside it are complete.
Every active project carries this field since 27 September 2026. In Southeast Asia and India the constraint that decides whether a project happens is grid access and the price of power. In the Gulf and Vietnam a second constraint bites, which is whether a project can lawfully receive the US-controlled chips an AI facility is built around. The tier grades what the public record documents about that access. As with evidence tiers, the grade describes a project's paper trail and never a company, and it is computed by fixed rules from cited government documents.
The rule we apply is the one the US Bureau of Industry and Security enforces today, which is the October 2023 footprint. A licence is required for advanced AI chips going to destinations in Country Groups D:1, D:4 and D:5, which among our countries means Saudi Arabia, Qatar, Oman, Bahrain, Kuwait and Vietnam, and the United Arab Emirates except for its list of approved recipients. Malaysia, Singapore, Indonesia, Thailand, the Philippines and India carry no destination-based requirement. The January 2025 worldwide licensing rule is still printed in the regulations, is not enforced, and has neither been rescinded nor replaced. Separately, a licence is required everywhere for chips destined to an entity headquartered in China or Macau, or whose ultimate parent is, and we grade that from the parties named on a project. China itself sits in Country Groups D:1, D:3, D:4 and D:5. Its licence review is a presumption of denial, eased since January 2026 to case-by-case review for chips of the H200 class.
| Tier | What it means |
|---|---|
government-confirmed | A dated US government document names the project's operator, its parent, or a named funder as authorised to receive advanced computing items. |
framework-covered | The country has a country-level export framework and the project has an approved US operator or confirmed client on it. That is a structural reason to expect compliance. It is not a licence held by the operator. |
commercial signal only | A chip-vendor partnership or a US technology relationship has been announced, and no government document names anyone involved. |
no public signal | Nothing found. This describes the public record and nothing else. Applies only in countries where a licence is required. |
no US licence required | The destination country carries no US destination-based licence requirement as the rules are enforced today. Access is a commercial matter between the project and its vendors. |
licence required for a party | The operator, its group, a funder or a confirmed tenant is headquartered in China or Macau, or has an ultimate parent there, so US rules require a licence wherever the building stands. Reported tenants never trigger this. It describes the rule and says nothing about whether a licence was sought, granted or refused. |
Three limits are worth stating plainly. The lowest tier is not an accusation. Absence of a public document is not evidence that access was sought, refused or revoked, and most companies never publish their licensing position at all. The engine also does not treat corporate ownership as coverage, so a subsidiary whose parent is named in a government list still rates on its own record, because the document names the parent and not the building. And we publish no percentage. Exactly three entities across the region hold a named, dated authorisation, which is enough to sort projects into categories and nowhere near enough to support arithmetic, so there is no share-of-pipeline figure here of the kind we publish for AI readiness and evidence.
Two facts shape everything in the Gulf. The United Arab Emirates has a country-level framework with a named list of approved recipients. Saudi Arabia does not, and operates through individual authorisations instead. One country has a rule, the other has a licence, and reading a Gulf project without knowing which applies to it will mislead you. No entity in Qatar, Bahrain, Oman or Kuwait appears in any US authorisation we found, so every project there grades on commercial signals only. Malaysia adds a layer of its own. Since July 2025 its trade ministry requires a permit for US-origin AI chips leaving the country, which governs onward shipment rather than use in a Malaysian data centre.
Some countries inside a covered region hold no project. Where that is because we looked and found nothing at or above the threshold, we say so, with the date, on the map scoreboard and the directory, so that "checked and empty" is never confused with "never looked". Oman and Kuwait were researched on 27 September 2026 and hold nothing above 20MW; both stay in scope and are rechecked at every Gulf refresh. The five smaller Southeast Asian countries have not yet been researched and carry no such line.
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. Six 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 sharply in a single half and the new stock has yet to lease up. Three different stories hide behind what looks like one metric.
Figures carry different vintages and the gap matters. Greater Manila, Bangkok and Ho Chi Minh City are current to the first half of 2026, and Ho Chi Minh City received its first published vacancy figure in that edition after years of qualitative coverage only. Singapore and Greater Jakarta are still current to December 2025 because their newer figures sit behind a paywall. We show the vintage rather than rolling an old number forward.
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.
Each project panel can show four more layers. Project team lists the designers, advisers and contractors by role, from architect and engineer through main contractor and design and build contractor down to mechanical and electrical works, grid works and consultants, with the contract value in the currency it was announced in. Named bidders appear with their result where one is known. Land names the seller, landlord, industrial estate or government body behind the site. Power supply names the utility and the arrangement, such as a grid supply agreement, a dedicated substation or a renewable power deal, with its size in MVA or MW. Water and cooling names the water source and supplier and the cooling design.
Every row carries one of three labels.
When two to four projects fit, we pin the contract to none of them. It appears on each candidate under Possibly linked, with the other candidates named. When five or more fit, the contract tells a reader almost nothing about any one of them, so we hold it back until a source narrows it down. We never infer a link that would imply wrongdoing, such as a breach of confidentiality or sanctions exposure, and hold those at possibly linked until a source names the project.
A power or renewable deal that covers an operator's whole country portfolio is shown on the relevant project and says so, because it tells you where the power comes from without proving this site is the one using it. Contract values are converted to US dollars at an approximate recent exchange rate for scale only.
Coverage is thin and uneven. Of 272 active projects, 34 have a named builder, 71 have a land counterparty, 68 have a power supply fact and 62 have a water or cooling fact. Another 32 facts are shown as possibly linked and 18 are held back because too many projects fit them. Malaysia is by far the best covered because its listed contractors and land sellers must disclose data centre deals. Singapore and Indonesia builders are rarely named in public. A blank means nothing public was found, never that the project lacks a builder.
The Colo prices page shows what a kilowatt of capacity rents for, market by market, and keeps three prices apart because they measure different things. Wholesale asking rent is the landlord's quote for a block of capacity before negotiation, per kilowatt of IT load per month, and brokers publish it. Passing rent is what an owner actually collects across signed leases, and we show it only where the owner states the figure itself, in a prospectus, a circular or an earnings call. We never derive it from revenue and capacity, because a reader cannot audit the denominator. Retail rack price is a published list price for one rack with a stated power allowance, or a broker band for deployments under 250 kW, and it is the highest of the three per kilowatt.
Every figure carries its block size, because size moves the price more than the city does. Retail is below 250 kW, wholesale is 250 kW to 4 MW and hyperscale is 4 MW and above. Every figure also states whether electricity is inside it, or that the source did not say. Local currency is primary. Where we converted, we used the European Central Bank reference rate on the last business day of the figure's month and printed the rate. Market research inside a prospectus or a circular is commissioned by the seller or the issuer, and each row names who commissioned it. A figure from a rack under 3 kW is excluded because it mostly measures fixed overhead. An empty cell means nobody publishes the figure, never that the market is cheap. Every candidate figure that failed review is listed on the page with the reason, and a practitioner review on 30 September 2026 passed 21 of 96.
Every project that is not built or cancelled shows its build odds: seven gates a real project passes on the way to construction, each marked passed or open from the public record. The gates are land secured, power arranged, approvals granted, money committed, a tenant or own use, a builder appointed and construction started. A gate counts as passed when the record holds a dated, sourced fact for it, such as a land deal, a grid or substation agreement, an investment or building approval, a named funder or closed financing, a reported or confirmed user, or an awarded main contract. A project that has started construction passes the land and approval gates, and a built project passes all seven.
The count sets the band. Zero or one gate is a long shot, two or three is an open question, four or five is leaning yes and six or seven is strong. A project on hold shows as stalled whatever its count. Where a date matters, such as the end of an exclusivity period in a memorandum, the page shows it as the next checkpoint and says what ArusGrid will do if nothing has moved by then.
Build odds is a reading of the public record, not a forecast, a rating or advice. Private progress does not show, a passed gate can be undone, and projects that pass every gate still stop. The method has not yet been tested against outcomes. Readers can add their own call on each project page, and those votes are shown as they are, without weighting or checks.
Where no investment figure has been announced, a project page can show an ArusGrid estimate of building cost. It is the planned IT load in MW times a published construction cost benchmark for the country, in US dollars per watt of IT load, shown as a range of 15 percent either side. Projects presented as AI facilities get 10 percent more at the top of the range for liquid cooling. The benchmark is the Turner & Townsend Data Centre Construction Cost Index 2025, which covers the building, its fit-out and its mechanical and electrical plant, and excludes land, grid and site works, servers and fees. A country with no benchmark gets no estimate, and neither does a project whose capacity figure is not IT load. An estimate is labelled on every page and is never added to any total.
Cancelled and on-hold projects stay on record with the reason they stopped, from a fixed list: no tenant, no financing, no power, permits not in place, land or legal terms, a partner walked away, announcement only, no progress on record, the sponsor lacked the means, physical damage or conflict, or the market moved. Where no party gave a reason, the page says so. The graveyard also lists announced or permitting projects with no dated news for 18 months as gone quiet. Those are not declared dead. They are projects to check.
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 |
| Builders, land, power, water (all countries) | 42 facts, every inferred link included | 24% before correction | inferred links stretched too far and wrong labels on role or supplier. Three links were downgraded to possibly linked, one fact was removed and six were corrected before publishing |
| Builders, land, power, water, second pass | 29 facts, every inferred link plus 25 at random, and all 9 new projects | 38% of facts before correction | figures and roles read past what the source says, such as a site power capacity recorded as a utility supply. One link was downgraded to possibly linked, one fact was removed and nine were corrected. Eight of the nine new projects needed a fix, mostly a date, a ceiling recorded as planned MW or a funder left out |
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? Use Suggest a correction at the foot of any project panel on the map, or write to corrections@arusgrid.com. A link to a dated source gets your correction checked fastest. Every suggestion is checked against its source before anything changes, and accepted corrections are listed below with the date they were made. We never publish who sent a correction.
No corrections received yet.
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. The full dataset is on GitHub at github.com/arusgrid/arusgrid-data, updated daily from this site, with the schema and licence.
This site uses Umami, a cookieless analytics tool, to measure readership in aggregate. It sets no cookies and stores no name, email or IP address. To recognise a returning reader it keeps a random identifier in your browser's local storage, which clearing site data removes. It records how the site is used, including search terms typed on the map, the sections read and text copied, to learn what readers look for. The site also keeps its own anonymous visit log in Cloudflare for the people who keep this site. It holds the same random identifier, the pages read, the actions above including the first 80 characters of copied text, and the visitor's country, for up to 400 days. Notes sent through the feedback box on project and rules pages, and any email left with one, are stored for the people who keep this site to read and are never published.
Downloading the dataset as a CSV asks for an email address, confirmed by a one-time link. We keep the address and the date to see who uses the data, and may write when the dataset changes materially. To have it removed, write to hello@arusgrid.com.