Hong Kong, Singapore, Tokyo: A Data Centre Comparison for Asian Site Selection
HONG KONG, Oct 25, 2026 - A regional headquarters needs a lease, a few racks, a legal entity, and people who can answer a regulator in English. A large training cluster needs a substation, a long-term power contract, water for cooling, and a construction workforce that will still be on site years after the ribbon cutting. Both get filed under "choosing a data centre market", and the same three cities appear on both shortlists. The criteria behind them barely overlap. Start with the workload, not the city That is why this comparison is organised around workloads instead of a league table. Hong Kong, Singapore and Tokyo each win on at least one axis and lose on at least one, and the axes do not carry equal weight for every tenant. A bank placing a matching engine cares about a 30 ms round trip to the next exchange. A hyperscaler cares whether a large load can be energised before the depreciation schedule eats the business case. Those two buyers should not read the same table the same way. The three cities are on the shortlist for different reasons, and it helps to say which. Hong Kong is a financial centre with a cable history. Singapore is a Southeast Asian gateway whose government rations power. Tokyo is a domestic market large enough to justify its own supply chain. A tenant that treats them as interchangeable will pick the cheapest lease and spend three years discovering why the choice was wrong. Six variables follow, each defined, sourced, and then read against the tenants it governs: cross-border latency and cable landings, industrial electricity price, rentable capacity and pipeline, licensing and ownership, technical talent, and the time needed to bring land and power online. The order runs from the questions that eliminate cities fastest to the ones that only break ties. Six variables, and why these six Each variable passes a simple test. It changes the answer for at least one plausible tenant, and it can be measured without asking a broker for a favour. Latency and cable landings decide whether a workload can be split across markets at all. Industrial electricity price is the largest controllable operating cost in a dense hall. Rentable capacity and pipeline decide whether a lease starts this year or in three. Licensing and ownership decide who may hold the asset. Talent decides whether the site can be staffed after the opening. Land and power timelines decide whether the project exists. Three candidates were dropped. Rack rents went first, because the three markets quote different products: Hong Kong space is commonly priced per square foot of raised floor, Singapore and Tokyo wholesale deals per kW, and neither basis converts cleanly without a density assumption the landlord will not warrant. Land cost per acre went next, because the figures I could reach sit behind subscription walls. Exchange-level latency went last, because venue latency depends on where a cage sits inside the building, which is a negotiation rather than a market statistic. What remains is narrow on purpose. It will not tell you which city has the better data centre industry. It will tell you which variable is likely to decide your project, and which city that variable points to. One methodological note before the numbers. Where two credible sources measure the same thing differently, I have kept both and said so instead of averaging them into a figure that belongs to neither. Singapore is the clearest case: the economic development board counts 1.4 GW of computing capacity across more than 70 facilities, while Cushman & Wakefield counts 1,068 MW across 51. The gap is scope, and scope matters more than the midpoint. Connectivity: the triangle is not equilateral The three cities sit on one trunk. The Southeast Asia–Japan Cable 2 runs 10,500 km and connects 11 landing stations, with its main trunk linking Singapore, Hong Kong and Japan; the system was designed for 144 Tbps and entered service on 16 July 2025. All three markets can therefore reach each other on capacity from the same build, which removes the old argument that one city is structurally cut off from another. The geometry still favours Hong Kong. Measured round-trip times between the three metros put Hong Kong 30.18 ms from Singapore and 62.17 ms from Tokyo, while Tokyo and Singapore sit 66.84 ms apart. Hong Kong is the hinge of the triangle and Tokyo is the far corner. For anything that must touch two of the three markets in sequence, that difference compounds with every request. Cable depth tells a different story. Hong Kong hosts 11 of the 14 intra-Asian submarine cables, according to Digital Realty. Singapore's advantage is breadth of networks rather than cable count: 742 distinct networks have a presence in the metro, against 583 in Hong Kong and 533 in Tokyo, on PeeringDB data. Tokyo's lower count is partly an artefact of how Japanese carriers peer and partly real. Read together, the connectivity picture splits three ways. Hong Kong is the lowest-latency place to sit if your traffic crosses the Pearl River Delta or shuttles between North and Southeast Asia. Singapore is the better place to buy transit, because more networks are already in the building. Tokyo is the right answer when the traffic is domestic Japanese or headed east across the Pacific, and a costly detour when it is not. Market Networks present RTT to Hong Kong RTT to Singapore RTT to Tokyo Hong Kong 583 — 30.18 ms 62.17 ms Singapore 742 30.18 ms — 66.84 ms Tokyo 533 62.17 ms 66.84 ms — Latency measured 25 October 2026 by WonderNetwork. Network counts from PeeringDB, October 2026. Power: the ranking flips when the tariff class changes Electricity is where a comparison most easily goes wrong, because "the industrial price" is not one number. Two credible series produce two different orderings of the same three cities, and both are defensible. Start with country-average business retail prices collected in March 2026. On that basis Hong Kong is cheapest at US$0.169 per kWh, Japan is second at US$0.179, and Singapore is dearest at US$0.212. A mid-sized enterprise buying a few megawatts would look at that table and put Hong Kong first. Now switch to large-load industrial tariffs, the product a data centre actually buys. Tokyo's special high-voltage average came to 19.40 yen per kWh across January to July 2026, or US$0.119. Hong Kong's all-customer average net tariff was HK$1.406 per kWh from January 2026, or US$0.179. Singapore's high-tension large tariff, weighted for a flat 24-hour load, came to SGD 0.2490 per kWh, or US$0.192. Japan moves from second to first, Hong Kong stays in the middle, and Singapore stays last by a wider margin. The reversal is not a data error. Hong Kong's published figure is an average across all customer classes, which flatters a large user. Japan's is the actual special high-voltage product. Singapore's includes a contracted capacity charge that a 24-hour load absorbs but a peaky load does not, so a tenant running batch training around the clock faces a different bill from a tenant running a trading floor for part of the day. Market structure explains part of the gap. Hong Kong sets tariffs under a scheme of control agreed with the government, which is why the published number is an average across customers rather than a menu. Singapore publishes a quarterly regulated tariff alongside a competitive wholesale market, so the same building can buy at two different prices depending on contract length. Japan's retail market is deregulated, and special high-voltage is a reporting category covering all retailers rather than a single offer. A tenant comparing headline numbers is comparing regulatory artefacts as much as fuel costs. There is a second power variable that does not appear in the tariff: how much of it is low-carbon. Japan generated 23.56% of its electricity from renewables in 2025, Singapore 5.49%, and Hong Kong 1.04%, on Ember data. For a tenant with a corporate renewable target, that gap is a procurement problem in Singapore and a reporting problem in Hong Kong. Demand is moving in one direction. CLP reported that its data centre electricity sales in Hong Kong grew 7.5% in 2025, driven by AI and digital services. Growth of that order against a fixed generation fleet is how a price advantage disappears. Capacity: what you can rent, and how fast Installed capacity is the least interesting number in this comparison, because none of the three markets is short of buildings. What matters is what is empty, what is under construction, and how quickly the empty space is absorbed. Hong Kong has 592 MW in service across 49 facilities and a vacancy rate of 15.6%, with 185 MW under construction. That is the softest market of the three by a wide margin, and the only one where a tenant with a near-term requirement can plausibly negotiate rather than queue. Vacancy and demand can coexist: in September 2026 Grand Ming agreed to sell two Fanling data centres for up to HK$2.445 billion, about US$312 million, with Amazon identified as the buyer. That is a platform purchase rather than a bet on empty space. Singapore has 1,068 MW across 51 facilities, a vacancy rate of 4.8%, and just 25 MW under construction. The pipeline equals about half a year of current take-up. New capacity arrives through a government allocation process rather than the open market: at least 300 MW was offered under the 2024 green data centre roadmap, and a second call for applications launched in December 2025 for at least 200 MW more, with additional power possible for proposals that bring new green energy pathways. Tokyo is the largest market at 1,397 MW across 123 facilities, with 4.4% vacancy and 220 MW under construction. It is also the fastest absorber: 223 MW of take-up in the first half of 2026 alone, against 30 MW in Hong Kong and 25 MW in Singapore. Pre-letting is normal. Colt's Inzai 4 was fully pre-let when it launched, with 4.8 MW energised, on the way to 20 MW at full build and 70 MW across the Inzai campus. The constraint that binds all three is the grid. Connection waits of more than three years are reported for Tokyo, Singapore and Hong Kong alike, which means the difference between the markets is less about how fast a utility can connect you and more about whether you can secure an allocation, a site, or a pre-let building that already holds the connection. Market Industrial power (US$/kWh) Operational capacity (MW) Vacancy Under construction (MW) Networks present Hong Kong 0.179 592 15.6% 185 583 Singapore 0.192 1,068 4.8% 25 742 Tokyo 0.119 1,397 4.4% 220 533 Capacity, vacancy and pipeline from Cushman & Wakefield APAC H1 2026, compiled by Datacentres.com. Power from CLP, SP Group and the Japanese energy surveillance commission, as compiled in the same source. Construction cost adds a second filter. Turner & Townsend's 2025 index put Singapore at US$14.53 per watt, up 5% year on year, second only to Tokyo and ahead of Silicon Valley; Kuala Lumpur came in at US$11.37 and Shanghai at US$6.12. Hong Kong does not appear in the coverage I could reach, which is itself informative: the market with the most empty space is also the one whose build economics are least publicly documented. Licensing: the gate is rarely the telecom licence All three markets allow a foreign investor to own 100% of a data centre operator. The differences sit in who needs a licence and who needs to notify. Hong Kong licenses facility-based public telecommunications through a unified carrier licence, a single instrument covering fixed and mobile services since August 2008. Running a colocation building does not require one; landing a cable or selling capacity on public circuits does. The investment regime around it is unusually open, with no distinction in law between foreign-controlled and locally controlled companies, and no requirement that directors reside in Hong Kong. Singapore requires a facilities-based operator licence to deploy network infrastructure and sell capacity, and the market has been fully liberalised since 2000, with foreign and domestic firms applying on the same terms. The State Department's investment climate report counted 76 facilities-based operators. A colocation landlord that does not run public transmission sits outside the licence, but any tenant wanting to resell cross-border capacity will need a partner that holds one. Japan is the outlier in ownership terms, though not where a data centre tenant would feel it. The foreign ownership caps that exist apply to NTT at 33.3% and to broadcasters below 20%, not to data centres. What does apply is a prior-notification regime: acquisitions of 10% or more of a company in a designated core industry require filing with the finance ministry and the sector regulator, a step that matters for anyone buying an operating platform rather than building one. Entity setup is quick by regional standards, at about one month for a branch and two for a subsidiary. Administrative speed is not the constraint in any of the three. Hong Kong registers a local company through a single electronic platform, typically on the same day, and a non-Hong Kong company in about two weeks. Japan measures entity setup in one to two months, which is fast for the region. The months that decide a project timeline are spent on land, power and construction approvals, not on the corporate paperwork. The practical gate in all three sits upstream of the licence. In Singapore it is the power allocation. In Hong Kong it is land, since almost every new hall involves a lease modification or a government site. In Tokyo it is the grid connection queue and, increasingly, local consent. Talent: a proxy with a wide error bar There is no comparable public count of liquid-cooling technicians by city, so the usable measure here is a proxy with its limits stated. The IMD World Talent Ranking places Hong Kong 4th, Singapore 7th and Japan 40th out of 69 economies in 2025, scoring investment in education, appeal to foreign talent, and the readiness of the existing pool. That ranking measures economies, not data centres, and it will not tell you whether a specific metro can staff a round-the-clock network operations centre in English. It does capture something a site-selection memo usually misses: the difference between a place where a shift team can be hired in weeks and a place where one has to be imported. For a regional headquarters with a small technical team, Hong Kong and Singapore are close enough that the decision will be settled on cost and lease terms. For a Japanese-language operation serving domestic customers, Tokyo's rank is irrelevant and its local labour market is the only one that counts. Reading the matrix Three readings follow from the tables, and they point in different directions. On latency, Hong Kong wins and Tokyo loses. On industrial power, Tokyo wins and Singapore loses. On available capacity, Hong Kong wins and Singapore loses. On network depth, Singapore wins. On market size and absorption, Tokyo wins. No city is first on more than two of the five, and that distribution is the finding: a comparison that produced a single winner would be hiding the trade the buyer actually has to make. Where the data runs out Four gaps are worth naming, because each one is a place where a confident-looking table would be inventing precision. Rack rents are not comparable across the three markets, for the reasons given earlier. Land cost per acre is published by the major brokers but sits behind subscription walls in every case I could reach. Permit timelines are estimated rather than observed, and the benchmarks behind them predate the current build cycle. Exchange-level latency, as opposed to metro-to-metro latency, is negotiated per cage and not published on a consistent basis. Two smaller gaps matter for specific tenants. Hong Kong's construction cost per watt is missing from the most recent index coverage I could obtain, so the market with the most vacant space cannot be priced on a like-for-like basis against Singapore and Tokyo. And the three markets define installed capacity differently enough that Singapore's total shifts by roughly a third depending on whether captive capacity is counted, which is why the table uses one compiler throughout instead of mixing sources. What would change these conclusions is specific and observable. If Hong Kong's 185 MW under construction is absorbed at anything like Tokyo's rate, its vacancy advantage disappears and the negotiating leverage moves to the landlord. If Singapore's second allocation clears and the awarded capacity is energised, its pipeline constraint eases, though the power price and the green conditions stay. If Tokyo's grid queue shortens, its cost advantage compounds. None of these is a forecast; each is a number to re-check before signing. If you need X, choose Y Latency-sensitive trading or settlement across North and Southeast Asia: Hong Kong. The 30.18 ms round trip to Singapore and 62.17 ms to Tokyo are the best pair in the set, and 11 of the 14 intra-Asian cables land there. A large, near-term block of space with negotiating leverage: Hong Kong. A 15.6% vacancy rate and 185 MW under construction are the only combination in the three markets that gives a tenant a credible alternative to the incumbent landlord. Deep peering and the widest choice of transit partners: Singapore. 742 networks is the largest ecosystem of the three, and it is why Singapore remains the default for cloud on-ramps even at the highest power price. The cheapest energy for a round-the-clock, power-dense load: Tokyo. At US$0.119 per kWh on the special high-voltage tariff, Japan is the only one of the three below twelve cents, and a 23.56% renewable share helps a corporate target as well as a bill. Scale without a government allocation: Tokyo, but only through a pre-let. 220 MW under construction against 223 MW absorbed in a single half-year means the buildings that exist are already spoken for, so the route is a negotiation with an operator rather than an application to a ministry. A regional headquarters with a small technical team and an English-speaking regulator: Singapore or Hong Kong. The talent proxy puts them 7th and 4th against Japan's 40th, and neither licensing regime imposes an ownership condition that would complicate a holding structure. A green-field campus where a renewable power purchase agreement is the binding constraint: Singapore, if you can win an allocation, and nowhere if you cannot. The 200 MW second call is the main route to new capacity, and it is scored on green energy pathways rather than price. Sources Submarine Networks, SJC2 system profile. https://www.submarinenetworks.com/en/systems/intra-asia/sjc2 Digital Realty, Hong Kong data centres. https://www.digitalrealty.co.uk/data-centers/china WonderNetwork, ping statistics between Hong Kong, Singapore and Tokyo, 25 October 2026. https://wondernetwork.com/pings/Singapore WonderNetwork, latency from Tokyo to Asian cities. https://wondernetwork.com/pings/Tokyo WonderNetwork, Tokyo to Singapore. https://wondernetwork.com/pings/Tokyo/Singapore Datacentres.com, Hong Kong market report. https://www.datacentres.com/markets/hong-kong Datacentres.com, Singapore market report. https://www.datacentres.com/markets/singapore Datacentres.com, Tokyo market report. https://www.datacentres.com/markets/tokyo GlobalPetrolPrices, Hong Kong electricity prices, March 2026. https://www.globalpetrolprices.com/Hong-Kong/electricity_prices/ GlobalPetrolPrices, Singapore electricity prices, March 2026. https://www.globalpetrolprices.com/Singapore/electricity_prices/ This article was prepared by the expert team at Global Infinity, drawing on insights and data provided through its MCP data services. About Global Infinity Global Infinity is an AI innovation company specializing in MCP data services. Website: https://globalinfinity.hk Email: support@globalinfinity.hk