Where should your AI actually run in Asia?

The questions I get about Asia have changed. A year ago they were about whether to be there at all. Now they arrive with a number attached — a gigawatt figure from a government announcement, a pipeline total from a market report — and what the caller wants to know is whether it means anything for the return. The excitement is real, and so is the unease underneath it. Enormous sums are being committed against capacity figures that nobody is quite sure how to read.

Part of the problem is that the published numbers answer a question nobody is asking. Take the four most quoted. India has a development pipeline of 8.33 GW against roughly 1.6 GW live.¹ Korea is targeting 8.4 GW of AI data center capacity by 2029 under a national program.² New South Wales has 11.4 GW in its pipeline, with national Australian capacity forecast to pass 5,000 MW in 2029/30.³ Malaysia roughly doubles to about 2,100 MW by the end of 2026, with Johor alone carrying a multi-gigawatt pipeline beyond that.^{4 5}

Notice that those four numbers are not measuring the same thing. Korea and Australia carry dates; the Indian and Johor pipeline figures do not, because a pipeline measures ultimate build-out and includes sites with no grid connection, no planning approval and no anchor tenant. Comparing them is the first mistake. The second, and the more expensive one, is assuming that capacity is the constraint at all.

It is not. What decides where a workload can run is a combination of what you are legally permitted to host, what the site can physically support, and what the power costs when you get there. On those criteria Asia is not one market. It divides into three, and the boundaries are sharper than most planning decks acknowledge.

Start with the arithmetic, because it sets the cost ceiling

Delivered power cost compounds twice in Asia: once through the tariff, and again through PUE, because tropical ambient temperatures push more of every megawatt into cooling rather than compute. Cost per megawatt-hour of IT load — the number that matters, rather than the headline tariff — varies more than fivefold across the region.⁶

A note on the Singapore and Johor PUE figures, because they sit 30km apart in identical climate and should not differ much on physics alone. They differ on regulation. Singapore caps PUE by policy — 1.3 for new builds under the Green Data Centre Roadmap, tightened to 1.25 in the latest capacity round — so its fleet is compelled to be efficient.^{7 8} Johor has no equivalent floor and a more mixed installed base, though its new hyperscale builds target similar numbers. The gap is a regulatory artifact, not a climate one.

On a 100 MW build at 85% utilization, the annual power bill in Johor runs about $45M above an Iowa equivalent, in Mumbai about $57M above, and in Seoul about $82M above. That is the cost ceiling: roughly how much margin the region takes off the top before a single commercial decision is made. Our modeling suggests a build of this kind can absorb something on the order of a quarter off its effective compute price before it stops earning its cost of capital — and in Asia, power alone consumes a large share of that cushion.⁶

The practical consequence is that these builds cannot rest on a price assumption. They have to rest on contracted demand. Which raises the question the cost table cannot answer: contracted for what? Johor runs a megawatt for about a third less than Seoul does, and yet the two are not remotely interchangeable in what they are permitted to host. Cost ranks the region. It does not sort it. That is where the second and third constraints come in, and where Asia stops behaving like one market.

Tier one: the markets that can host training

Korea, Japan and Australia can host a serious training run. Not because their power is cheap — Japan and Korea are the two most expensive in the table — but because three other things line up.

They sit in the allied tier of US export controls, so there is no licensing overhang on the hardware.⁹ Their security and legal posture can support the weight-protection requirements a frontier lab imposes, which is a real gate and one that rarely appears in diligence memos. And each has domestic demand for training rather than only for serving.

Korea is the clearest case. It has domestic model labs competing for national-champion status — Naver, LG, SK Telecom, NCSoft, Upstage — so the compute has customers at home. Nvidia has committed to deploying more than 250,000 GPUs across the country, with AI factories at Samsung, SK, Hyundai, LG and Doosan.¹⁰ And Korea makes the scarcest input in the entire stack: SK hynix and Samsung together hold roughly 90% of HBM supply. That combination does not exist anywhere else in Asia.

Japan has moved from talk to deployment. SoftBank broke ground on what could become a 1 GW campus at Tomakomai and is launching a GB200 NVL72-based GPU cloud from October 2026; Sakura Internet is taking its fleet from roughly 2,000 to 10,800 GPUs;¹¹ Microsoft has committed about $10bn through fiscal 2029, with compute sitting inside partner-operated racks at both.¹² Expensive power, but capability is not the constraint.

Australia is the one people underrate. Roughly 2,800 MW today heading past 5,000 MW by 2029/30,³ the cheapest delivered power of the three, allied status, and NextDC sovereign AI arrangement with OpenAI at Eastern Creek.¹³ Its limitation is the mirror image of Southeast Asia. It can train, but it has almost no one nearby to serve: Sydney to Singapore is around 90ms, to Tokyo around 110ms, and the domestic enterprise base is small relative to the markets its capacity would otherwise address.

Which raises the obvious question: if power in Seoul costs three times what it costs in Iowa, why not train offshore and bring the weights home? Because for training, power is not the dominant cost. On our model of a 100 MW build, hardware depreciation is roughly three quarters of the annual bill and electricity around fifteen percent — so relocating a Korean training run to the cheapest power in the world saves on the order of a tenth of total cost. Against that sit four things that do not travel: the training data, much of which is language, government, health and industrial data subject to its own residency rules; the weights, which are the strategic asset the whole sovereign program exists to own; the chip allocation, which Nvidia grants to national initiatives and which does not transfer to a third country; and the state funding, which is conditional on domestic deployment. A tenth of cost is a poor price for giving up all four.

Tier two: the markets that serve, and increasingly train at national scale

India and Malaysia both serve at scale and neither can host a frontier run. But they are in this tier for different reasons, and the difference is the single most important thing an underwriter should take from this essay.

India Digital Personal Data Protection Act moved into enforcement in 2026, with rules notified in November 2025. Entities designated Significant Data Fiduciaries — large or high-risk organizations named by the central government, expected to include major fintech, telecom, e-commerce and healthcare platforms — face a prohibition on transferring specified categories of personal data outside India.¹⁴ Layered on top sits the RBI long-standing requirement that payment data be stored in-country. For a bank or an insurer, the question is no longer whether to run inference in India. It is which vendor, and at what price.

Malaysia’s Personal Data Protection Act was amended in 2024 to delete the whitelist that had governed cross-border transfers; data may now leave Malaysia wherever the destination has substantially similar law or ensures adequate protection. There is no general localization requirement.¹⁵ So Malaysian demand is not created by statute. It is derived — from Singapore’s constraint 30km south, from cheaper land and power, and from proximity to the same users. That is real demand, and it is also the more fragile of the two: India’s is locked in by law, Malaysia’s is locked in by a price gap that policy could close.

A clarification is owed here, because I have put this to clients myself. The claim I have made — and still hold — is that these markets cannot train frontier models. What I did not say, and should have, is that they train perfectly well at a tier below that.

India is training. Sarvam AI has released 30B and 105B-parameter models trained on IndiaAI Mission compute,¹⁶ and Reliance is commissioning 120 MW at Jamnagar on Nvidia GB300 systems — the widely quoted “200,000 H100-equivalents” is a throughput comparison measured on an inference basis, not a count of H100s.¹⁷ What remains true is narrower and still decision-relevant: this is sovereign-scale training, not frontier training, and it is domestic weights rather than a foreign lab. The IndiaAI Mission had 34,333 GPUs empanelled as of May 2025 against a stated target of 100,000 by end-2026 — but empanelled, in the Mission’s own term, means contracted as available across seven separate providers.¹⁸ That is not the same thing as a contiguous cluster of 30,000 accelerators on one fabric, held for months, which is what a frontier run consumes. Fragmentation is the binding constraint in India, not the headline count. Krutrim retreat from frontier ambitions toward cloud services is the cautionary data point.

What Malaysia does share with India is the ceiling on training, and there the constraint is legal rather than physical. Since 14 July 2025, every export, transshipment or transit of US-origin high-performance AI chips has required a Strategic Trade Permit. A further thirty-day advance notification applies under the catch-all control in Section 12 of the Strategic Trade Act 2010, where the exporter knows or has reasonable grounds to suspect the item will be misused. MITI describes the measure as closing a regulatory gap while it reviews whether to add these chips to the Strategic Items List outright^{19 20} — which is to say the regime is an interim one, and more likely to tighten than relax. Washington watches Johor closely because it sits on the main suspected route into China. For an operator this is not a compliance cost line. Chip supply is allocated, not merely sold, and an operator found to have let restricted hardware reach a prohibited end user can lose access to future allocation entirely. A facility without chips has no business.

There is also a demand-side dependency nobody models. Johor absorbs Singapore overflow. Its demand is derived from Singapore constraint, not indigenous to Malaysia — which means the shock to underwrite is not a recession. It is Singapore loosening its capacity allocation and keeping that demand at home.

Tier three: Singapore is a control plane, not a capacity market

Singapore is where the map most misleads. At $229/IT-MWh it has the worst delivered-power economics in the region, and its capacity is not sold, it is allocated. The second Data Centre Call for Application, launched by EDB and IMDA in December 2025, offers at least 200 MW and requires a PUE of 1.25 or better at full IT load, at least 50% of power from eligible green pathways, and Green Mark for Data Centres 2024 Platinum certification. Capacity is awarded through this process rather than bought on the open market — IMDA calls the Call for Application “the primary mechanism for DC capacity allocation.”⁷

Read that as what it is: capacity rationed against a policy test rather than sold at a price. The asset in Singapore is the permit, not the land.

Which makes the workload question simple for Singapore. Only latency- and jurisdiction-critical work justifies the premium — regional control planes, financial services with Singapore-law requirements, peering and interconnection. Everything else migrates 30km north, which is the entire Johor trade.

What this means for your decision

If you are an enterprise, the tiers above describe what each market will permit. The prior question is how much of your estate — the portfolio of AI workloads you actually run — is location-bound at all. In my experience the honest answer is less than teams assume. A model that reads customer account records to answer a service query is bound, because the personal data it processes is itself regulated. A code-generation assistant working on your own source code usually is not: the input is corporate IP, which raises confidentiality and contractual questions, but not the residency statutes that force a jurisdiction.

Two constraints get conflated here and are worth separating. Residency is about where data physically sits — an Indian statute requiring records to remain on Indian soil. Jurisdiction is about whose law governs and who can compel disclosure — which depends on your provider corporate domicile, not the location of the disk. Data resident in India but held by a US-domiciled provider may still be reachable under US legal process. Residency pushes you into tier two. Jurisdiction pushes you toward a particular contracting entity and governing law — and toward a venue whose courts and regulator your own regulator will accept. That is a large part of why regulated firms tolerate Singapore pricing: MAS outsourcing guidance requires a financial institution to secure contractual rights of access and inspection for the regulator over its service providers, and those rights are far easier to evidence when the contract, the counterparty and the forum all sit in a jurisdiction the supervisor already trusts.

So sort by what forces a location, in order. Does law require this data to stay in a named country? Does the user experience require sub-50ms latency? Only then, what does it cost. Workloads bound by residency belong in tier two, near their users. Workloads bound by jurisdiction, or by latency to financial infrastructure, may justify tier three despite the price. Everything left over — usually the majority — can sit wherever power is cheapest, which will rarely be in Asia at all.

Two specifics worth carrying into a vendor conversation. Regional GPU-hour pricing runs roughly 15–40% off US rates, so confirm the quoted rate against local comparables rather than a global price list. And in India, ask how your provider buys its power. Industrial users on the state utility pay an added charge that funds cheaper power for farms and households, and a provider that buys directly from a generator, or owns its own, avoids much of it. That is one of the larger controllable lines in their cost base, and therefore one of the more negotiable lines in yours.²¹

If you are underwriting a build, the tiering tells you which demand you are actually buying. A tier-two facility anchored by a domestic regulated tenant is buying residency-driven demand that survives an AI downturn. The same facility anchored by a foreign lab taking burst capacity is buying the AI cycle at infrastructure pricing. Those are different assets at the same address, and the difference does not appear in the megawatt count.

The map, in the end, is not a map of capacity. It is a map of what each jurisdiction will let you do, what its grid will physically deliver, and who is near enough to serve. Gigawatts are the least informative number on it.

References

1. India development pipeline of 8.33 GW against roughly 1.6 GW operational, Knight Frank India, June 2026. Around a quarter of the national pipeline is assessed as fully deliverable. Pipeline figures measure ultimate build-out and carry no completion date. https://yourstory.com/2026/06/indias-data-centre-pipeline-knight-frank-india ↩
2. Korea national AI data center program: 8.4 GW targeted by 2029 in phase one, 18.4 GW by 2035. Government targets, not committed pipeline. https://www.lightreading.com/data-centers/south-korea-plans-massive-18-4gw-ai-data-center-buildout-by-2035 ↩
3. Australian capacity of approximately 2,800 MW in 2025/26 forecast to pass 5,000 MW in 2029/30; NSW pipeline of 11.4 GW across 44 facilities as at 31 March 2026. Australian Data Centre Forecast Report, April 2026. https://datacentres.org.au/wp-content/uploads/2026/04/2026_Australian-Data-Centre-Forecast-Report-Issue-1-1.pdf ↩
4. Malaysian capacity roughly doubling from 1,025 MW in 2025 to 2,100 MW by end-2026. https://technave.com/gadget/Malaysia-to-double-its-data-centre-capacity-by-the-end-of-2026-46066.html ↩
5. Johor data center demand and pipeline. Wood Mackenzie estimates Johor maximum demand at approximately 3.8 GW, close to one and a half times the state current electricity demand. https://www.woodmac.com/press-releases/jb-data-center-expansion/ ↩
6. Power and PUE figures are Aureak estimates compiled from published industrial tariffs and operator disclosures, not measured values. Economics are from the Aureak capacity model on a 100 MW build at 85% utilization. ↩
7. Infocomm Media Development Authority and Economic Development Board, Launch of second Data Centre Call for Application, 1 December 2025. At least 200 MW; PUE of 1.25 at 100% IT load or better; BCA-IMDA Green Mark for Data Centres 2024 Platinum; at least 50% of power from eligible green energy pathways; SS 715:2025 for IT equipment. The factsheet states that the DC-CFA is the primary mechanism for DC capacity allocation. https://www.imda.gov.sg/resources/press-releases-factsheets-and-speeches/factsheets/2025/launch-of-second-data-centre ↩
8. IMDA Green Data Centre Roadmap, launched 30 May 2024, including the refreshed Green Mark for Data Centres 2024 and the Singapore Standard on Energy Efficiency of Data Centre IT Equipment. https://www.imda.gov.sg/how-we-can-help/green-dc-roadmap ↩
9. US export controls and AI: allied-tier treatment, licensing requirements and the contested status of the AI Diffusion Rule. https://www.onelexpartners.com/news-and-insights/us-export-controls-and-ai-a-practitioners-guide ↩
10. Nvidia commitment to deploy more than 250,000 GPUs across Korea, with AI factories at Samsung, SK Group, Hyundai, LG and Doosan. https://www.datacenterdynamics.com/en/news/nvidia-to-deploy-more-than-250000-gpus-across-south-korea-with-samsung-sk-group-and-hyundai-all-announcing-ai-factories/ ↩
11. SoftBank Tomakomai campus and GB200 NVL72 GPU cloud launching October 2026; Sakura Internet expanding from roughly 2,000 to 10,800 GPUs. https://www.datacenterdynamics.com/en/news/softbank-corp-to-launch-ai-data-center-gpu-cloud-offering-in-japan/ ↩
12. Microsoft approximately $10bn Japan commitment through fiscal 2029, with GPU compute in partner-operated racks at SoftBank and Sakura Internet. https://www.aicerts.ai/news/microsofts-10b-japan-bet-elevates-sovereign-ai-infrastructure/ ↩
13. NextDC and OpenAI sovereign AI arrangement at Eastern Creek, and the wider Australian AI infrastructure position. https://introl.com/blog/australia-ai-infrastructure-openai-sovereign-compute-2026 ↩
14. DPDP Rules notified 14 November 2025 with enforcement from 2026. Significant Data Fiduciaries are large or high-risk entities designated by the central government, and face a prohibition on transferring specified categories of personal data outside India. https://mickai.co.uk/articles/india-dpdp-first-enforcement-localisation ↩
15. Personal Data Protection Commissioner (Malaysia), Public Consultation Paper 05/2024 on Cross Border Personal Data Transfer. The Personal Data Protection (Amendment) Act 2024 deletes the Section 129(1) whitelist; transfers out of Malaysia are permitted where the destination has substantially similar law or ensures adequate protection. Malaysia has no general data localization requirement. https://www.pdp.gov.my/ppdpv1/wp-content/uploads/2025/08/JPDP-FSB-241001-Cross-Border-PCP-ENG-TC.pdf ↩
16. Indian sovereign model program: Sarvam AI 30B and 105B models trained on IndiaAI Mission compute, and the scale of the Mission GPU allocation. https://explainx.ai/blog/india-sovereign-ai-status-indiaai-mission-2026 ↩
17. Reliance Intelligence commissioning 120 MW at Jamnagar on Nvidia GB300, with a stated path beyond 200,000 H100-equivalents measured on an inference basis. https://letsdatascience.com/news/reliance-intelligence-operationalises-120mw-ai-backbone-in-j-e028b01c ↩
18. Press Information Bureau, Ministry of Electronics and IT, 30 May 2025: India common compute capacity crosses 34,000 GPUs — 15,916 added to 18,417 already empanelled, across seven empanelled providers. https://www.pib.gov.in/PressReleasePage.aspx?PRID=2132817&reg=3&lang=2 ↩
19. Ministry of Investment, Trade and Industry (Malaysia), press statement of 14 July 2025: exports, transshipments and transits of high-performance AI chips of US origin are subject to a Strategic Trade Permit with immediate effect. The 30-day advance notification arises under the Section 12 catch-all control of the Strategic Trade Act 2010, which applies where the exporter knows or has reasonable grounds to suspect the item will be misused or used for a restricted activity. MITI describes the measure as closing a regulatory gap pending review of whether to add these chips to the Strategic Items List. https://www.miti.gov.my/miti/resources/Media%20Release/%5BFINAL%5D_MITI_Press_Stmt_Malaysia_Regulates_Trade_of_US_AI_Chips_2025-07-14.pdf ↩
20. Directive of the Strategic Trade Controller No. 1/2025 on advanced artificial intelligence chips, and the accompanying industry compliance guideline. https://www.miti.gov.my/miti/resources/STA%20Folder/PDF%20file/1_2025_Directive_Unlisted_Category_AI_Chips_as_of_14_July_2025.pdf ↩
21. Indian power procurement for data centers: discom cross-subsidy surcharges, open access, and the regulatory shift toward captive and private licence structures. https://www.thecore.in/business/india-data-centres-power-discoms-clouds-artificial-intelligence-energy-864716 ↩