Singapore or Johor: Where Should Your AI Rack Live?
If you want to host a rack-scale AI system for a Singapore business, you quickly meet a geography problem. Singapore has the network, the customers and the rules you know. It does not have much spare power. Across the Causeway, Johor has been building data centres faster than almost anywhere on Earth.
The obvious conclusion is "put it in Johor". The honest picture is messier. Both locations are constrained, just by different things, and the right answer depends on your data, your timeline and how much you trust a pipeline figure.
The two markets in pictures
| What | πΈπ¬ Singapore | π²πΎ Johor |
|---|---|---|
| Live capacity | More than 1.1 GW | About 1.3 GW in Johor Bahru |
| New capacity coming | π΄ Tightly rationed: at least 200 MW in the latest allocation round |
π’ Pipeline of about 8,500 MW |
| Colocation vacancy today | 4.9% |
0.7% |
| Main constraint | Land and power | Grid capacity and water |
| Rules for new builds | Strict efficiency and green energy requirements | Tightening, with scrutiny of water use |
| Data stays in Singapore | β Yes | β No, it crosses a border |
Sources: TechNode Global on Knight Frank's findings, TechRepublic on Johor capacity, Real Estate Asia on Savills data. Figures come from different consultancies and dates, so treat them as indicative. The vacancy bars are scaled to 10%.
The number that changes the story
Look at the vacancy row. Johor has the enormous pipeline, yet its colocation vacancy rate is reported at just 0.7%, far tighter than Singapore's 4.9%. Planned capacity and capacity you can move into this quarter are different things. As one report put it, regional buyers need to separate announced capacity from infrastructure they can actually deploy.
So the common assumption, that Johor is where the space is, is only true on a multi-year view.
Singapore: scarce, strict and close
Singapore controls data centre growth through allocation rounds. The first, in 2023, awarded only 80 MW. The second made at least 200 MW available, with applicants required to achieve a power usage effectiveness of 1.25 or better and to source at least half their electricity from renewable energy.
The effect for a buyer is that high-density, liquid-cooled space exists but is limited and priced accordingly. What you get in return is proximity, a familiar regulatory environment, and data that never leaves the country.
Johor: abundant on paper, constrained in practice
Johor benefited enormously from Singapore's earlier moratorium on new data centres. Reports now describe a new phase, where the question is no longer how much investment the state can attract but how much its power, grid and water can sustain.
One report deserves particular attention if you are planning liquid-cooled AI. Nikkei Asia reported that an investment review committee under the Johor state government had begun rejecting new data centre applications that call for liquid cooling systems, citing water supply constraints. We have not seen the policy text and it may have changed, so verify the current position. But it is a reminder that "liquid cooling" and "water-stressed location" can be in tension, depending on how heat is ultimately rejected.
At the same time, purpose-built AI capacity is arriving. Singtel's Nxera and Telekom Malaysia announced an AI-ready campus in Iskandar Puteri, about 16 km from Singapore, designed around liquid cooling and GPU-heavy workloads.
Sources: Nikkei Asia via KrASIA, Reuters via Khaleej Times, JS-SEZ Monitor
The factors that are not about megawatts
| Factor | Why it matters | Who must answer |
|---|---|---|
| π Data residency | Some regulated data may need to stay in Singapore or meet transfer conditions | Your compliance team |
| π Export controls | Advanced AI hardware is subject to end-user and location documentation; installation country matters | Your supplier and legal adviser |
| π Physical access | Your engineers cross an international border for every visit | Your operations team |
| π Latency | Negligible for training; check for real-time inference | Your architects |
| π§° Vendor support | Warranty and on-site service terms can differ by country | Your hardware vendor |
| π± Tax and customs | Importing multi-million-dollar hardware into another jurisdiction | Your finance team |
The export-control row is the one most often overlooked. Where an advanced AI system is installed is part of its compliance paperwork. We cover why in Why Your AI Infrastructure Supplier Should Understand Export Compliance. This article is not legal advice; get the specifics from qualified advisers.
Three uncomfortable observations
1. Cheaper is not the same as available
Johor's appeal is lower land and energy cost. That advantage is real. It is also why space there filled up. A lower rate on a facility that cannot take your rack until 2028 is not a saving.
2. Liquid cooling is not free anywhere
Industry figures presented this year put the cost of liquid cooling for high-density chips at US$1.80 to US$2.40 per watt, raising construction costs by roughly 18 to 22%. Operators on both sides of the border pass that on. AI-ready space commands a premium wherever it is.
Source: The Nation, citing Datacloud Global Congress 2026
3. The same constraints follow the industry around
Singapore paused data centre growth because of power and water. Johor absorbed the overflow and is now facing its own version of the same pressures. Anyone planning a ten-year hosting strategy should assume that resource limits, not land, set the pace across the whole region.
A decision guide
| Your situation | Lean towards |
|---|---|
| π¦ Regulated data with residency requirements | πΈπ¬ Singapore |
| β‘ Real-time inference for Singapore users | πΈπ¬ Singapore, or test latency first |
| ποΈ One or two racks, need them soon | Whichever has a liquid-cooled position available now |
| ποΈ Large training cluster, cost-sensitive, non-sensitive data | π²πΎ Johor, with capacity confirmed in writing |
| π Multi-year growth plan | Split: sensitive and latency-critical in Singapore, bulk compute in Johor |
What to ask any operator, on either side
- Is a liquid-cooled position at my rack density available today, or is it planned?
- What is the committed power per position, and when is it energised?
- How is heat rejected, and what is the site's water source?
- What happens to my contract if power or water allocation is delayed?
For the technical side of that conversation, see Can Your Singapore Colocation Hall Take a 140 kW Rack?




