← Research · Digital Infrastructure

STRATEGY · APAC · JULY 2026

Print report

The Physical StackWhere the Bottlenecks Are

Dorrsum Advisory · July 2026

Executive Summary

Capital is no longer the binding constraint. The physical supply chain is.

19.4 GW

APAC development pipeline

13.8 GW operational · 3.7 GW under construction

92%

cite grid as hard ceiling

Power-first siting now drives valuation (Source: AlphaStruxure)

18–36 mo

grid interconnection lead time

vs. 18-month speed-to-market demand

Asia-Pacific's AI infrastructure buildout has moved from ambition to large-scale execution. Southeast Asia alone accounts for 31% of capacity currently under construction.

Capital is no longer the binding constraint — the physical supply chain is. This note maps six critical nodes, each classified by constraint status, to show where risk and opportunity sit in the stack.

Exhibit 1 — Supply chain stress summary

NodeStatusConstraint intensityAPAC dynamicTimeline to ease
PowerCritical
95
Grid interconnection 18–36 months; Singapore allocation; SEA nuclear 2035+3–5 years
SiliconTight
78
Rubin supply capped at industry-estimated 200–300K in 2026; TSMC CoWoS booked through 202712–18 months
CoolingInnovating
42
Asia supply chain advantage; STT GDC HVDC testbed; Rubin mandates liquidEasing now
Land / PermitsTightening
65
Malaysia tightened permits; Singapore quota; India greenfield runwayMarket-specific
ConnectivityStrategic
55
Hyperscale-used corridors; Singapore hub concentration; ICE II bypassPersistent
ConstructionStressed
72
MEP talent war; 7–10% cost CAGR; OSM baseline; IST commissioning risk2–3 years

Intensity — 0 (easing) to 100 (critical): Dorrsum's composite score of near-term supply-chain constraint severity for this node.

Source: Dorrsum analysis; node status per JLL 2026 Global Data Center Outlook; APAC pipeline per Cushman & Wakefield H2 2025 Update; Grid constraints per AlphaStruxure; Subsea capacity per TeleGeography.

1Critical Bottleneck

PowerThe New Asset Class Driver

So what

For investors, APAC data centre viability is now a pure-play bet on electrons — not capital or metro proximity.

See Exhibit 2.

Nearly 90% of industry respondents cite grid access as the hard ceiling for project delivery. The mismatch between 5-year transmission build-outs and 18-month speed-to-market demand is creating a structural premium for secured interconnection.

320 → 780 TWh

APAC DC Power demand by 2030

Power-first siting has officially displaced metro proximity as the primary valuation driver.

  • Singapore — allocation, not voltage

    The binding constraint is capacity allocation, connection queues, and the post-moratorium DC-CFA regime — not architectural voltage class.

  • Malaysia — centralized merit

    October 2025 framework moved power permitting to a merit system. Operators must prove grid-efficiency to earn utility allocation.

  • The renewable gap

    Only 32% of projected APAC DC demand can be met by renewables. Nuclear/SMR and firmed LNG bridge the baseload gap until grid expansion catches up.

Exhibit 2 — APAC operational DC capacity trajectory

GW
ActualProjected

Source: Wood Mackenzie SEA DC Power Demand; CBRE APAC Data Centre Boom 2026.

Investor implication

The data centre is no longer a real estate play

  • Grid scarcity premium: firm power and secured allocation rights — not voltage class — compress cap rates.
  • Efficiency as alpha: grid-efficient designs win Singapore DC-CFA / CFA2 capacity rounds.
  • Watch Japan nuclear restarts and Malaysia's merit-based permitting as regional capacity lead indicators.
2Tight

SiliconThe Heterogeneous Compute Era

So what

The GPU shortage is over — the complexity gap in memory, packaging, and custom silicon is the new bottleneck.

See Exhibits 3, 4.

Silicon supply is defined by integration of logic, memory, and specialized packaging. NVIDIA's Rubin transition opens a window for custom accelerators on inference workloads.

120 kW

per Rubin NVL72 rack

Shifts value to tokens per watt

  • HBM4 gating factor

    SK Hynix (~70% share) and Samsung's OpenAI-exclusive line signal the end of off-the-shelf memory for Tier 1 labs.

  • CoWoS fully booked

    TSMC at 120k wafers/mo remains committed to NVIDIA and Google through 2027.

  • Agentic silicon

    Google TPU v7, Microsoft Maia 200, and Meta MTIA v3 target inference and agent workflows — not training scale alone.

Exhibit 3 — Inference market capture (2028E)

NVIDIA GPUs~75%
Custom ASICs~25%

Source: Dorrsum analysis.

Exhibit 4 — Silicon — supply chain detail

ComponentKey players2026 context
HBM4 MemorySK Hynix, SamsungPrimary gating factor for Rubin; Samsung OpenAI-exclusive line (Apr 2026).
Advanced packagingTSMC (CoWoS-L), Amkor120k wafers/mo, fully booked through 2027.
ASIC designBroadcom, MarvellSilent winners designing custom silicon for Meta, Google, Amazon.

Source: Dorrsum analysis of public vendor disclosures (TSMC, SK Hynix, Samsung).

Investor implication

Silicon alpha shifts to the memory-logic-packaging stack

  • Diversification into Broadcom and SK Hynix is as critical as NVIDIA exposure.
  • Value metric is moving from TFLOPS per chip to tokens per watt.
  • Rubin supply capped at an industry-estimated 200–300K units in 2026 — plan for allocation risk.
3Rapidly Innovating

CoolingFrom Specialty to Standard

So what

Liquid cooling is no longer optional — it determines whether a facility is investable for Tier 1 AI workloads.

See Exhibit 5.

Air cooling for high-performance training clusters is a technical impossibility by H2 2026. Rubin's 45°C inlet water shift allows chiller-free operation in warm climates, improving PUE materially.

~30%

liquid cooling penetration in AI DCs

Mandatory for Rubin clusters

Any facility built for air-only operation faces immediate structural obsolescence.

  • NVIDIA liquid standard

    GTC 2026 certified cold plate suppliers: AVC, Cooler Master, Jentech, Delta Electronics.

  • China manufacturing corridor

    Envicool designated core CDU supplier for Google (~25% of 2026 procurement). Market >CNY 70B, 50%+ CAGR.

  • FutureGrid Accelerator (STT GDC)

    Southeast Asia's first live HVDC + liquid cooling testbed — ~30% energy loss reduction vs. conventional design.

Exhibit 5 — Cooling — supply chain detail

ComponentKey players2026 context
Cold Plates (Direct-to-Chip)AVC, Jentech, Cooler Master100% attach rate to Rubin clusters; high-margin components protected by strict NVIDIA certification moats.
Cooling Distribution Units (CDUs)Vertiv, Envicool, Delta ElectronicsCore bottleneck for facility retrofits; Envicool securing ~25% of Google procurement validates Asian hardware scale.
Quick Disconnects & ManifoldsStäubli, Parker Hannifin, DanfossCritical plumbing chokepoint; aerospace-grade leak-proof precision creates tight supply as adoption hits ~30%.

Source: Dorrsum analysis; NVIDIA GTC supplier certifications and public vendor disclosures.

Investor implication

Cooling determines investable grade

  • Operators with liquid retrofits (CoreWeave, Equinix) win high-margin training contracts.
  • Air-only facilities are relegated to commodity enterprise workloads.
  • Asia cold plate / CDU supply chain (AVC, Jentech, Envicool) is a higher-certainty picks-and-shovels play.
4Tightening

Land & PermittingThe Regulatory Variable

So what

Land is available — permits are not. Permitting timelines, not land cost, drive project IRR.

See Exhibit 6.

Singapore's post-moratorium quota regime favours incumbents. Overflow routes to Johor and Batam. Malaysia's late-2025 tightening adds friction for new entrants, especially Chinese-linked operators.

47%

India capacity in Mumbai

Chennai 20% — power-led siting

  • Singapore

    Capacity-based allocation with stringent PUE and sustainability requirements. High barriers, premium pricing.

  • Malaysia

    Tighter permits citing power/water constraints. GDS DayOne faces export-control-linked friction.

  • India & Japan

    Jamnagar (3 GW) and AdaniConneX Vizag (1 GW) diversify geography. Japan nuclear restarts unlock baseload-advantaged sites in Osaka.

Exhibit 6 — The power-led pipeline shift

Live CapacityPlanned Pipeline

Mumbai (Legacy Metro)

47%
15%

Chennai (Subsea Hub)

20%
20%

Greenfield (Jamnagar/Vizag)

33%
65%

Source: Dorrsum analysis of India data centre pipeline vs. live capacity.

Investor implication

Permitting is the IRR variable

  • Singapore favours incumbents; Malaysia creates uncertainty for new entrants.
  • India and Japan offer the most greenfield runway — with different risk profiles.
  • Site selection is increasingly power-availability-led, not metro-demand-led.
5Strategic

ConnectivityCommoditizing but Strategic

So what

Bandwidth is abundant — route concentration is the risk. The question is whose waters the cable crosses.

See Exhibit 7.

Over US$13 billion in subsea investment is hitting the water between 2025 and 2027. Priority shifted from shortest path to most trusted path — reshaping landing station valuations across APAC.

>70%

transcontinental capacity owned by hyperscalers

$13B+ subsea investment 2025–27

  • Meta Candle (570 Tbps)

    APAC's largest capacity system — Japan to Singapore via Taiwan, Philippines, Indonesia, Malaysia. H2 2026 construction phase underway.

  • ICE II bypass

    Telin's Singapore–Manado route avoids contested South China Sea corridor.

  • M2M traffic pattern

    Data prep in Vietnam/Philippines → training in Johor/Australia → inference at edge in Tier 1 cities.

Exhibit 7 — Connectivity — supply chain detail

Hub categoryLead marketsStrategic role (2026)
Primary hubSingaporeInterconnection king — network effect despite power/land premium.
Secondary surgeJakarta, MumbaiInference engines catching traffic before Singapore bottleneck.
Bypass corridorsBatam, JohorOverflow valves — massive clusters one millisecond from SG hub.

Source: Dorrsum analysis; subsea cable system disclosures.

Investor implication

Route concentration, not bandwidth, is the risk

  • Account for Singapore risk — most connected city, but single point of failure in grey-zone scenarios.
  • Hyperscalers are owners of the seabed, not just customers — watch landing station operators and Telin.
  • 800G / 1.6T optics (Ciena, Infinera) are critical for bursty asymmetric AI flows.
6Stressed

ConstructionCost Pressure

So what

Funding is abundant — execution capacity is not. MEP talent scarcity is the new construction bottleneck.

See Exhibits 8, 9.

Simultaneous AI factory build-outs across Tokyo, Mumbai, Johor, and Sydney created a bidding war for certified engineers. Projects fail at IST and commissioning — not for lack of concrete.

7–10%

APAC construction cost CAGR

Driven by MEP scarcity

  • MEP talent war

    120 kW+ racks need precision liquid plumbing and power distribution general crews cannot execute.

  • Modular baseline

    Prefab power skids (Schneider, Vertiv) cut on-site electrical labour ~40%. AirTrunk maintains <18-month cycles via repeatable design.

  • NVIDIA reference designs

    Schneider collaboration enables drop-in high-density clusters into existing facility shells.

Exhibit 8 — Construction cost pressure index (APAC DC, 2022 = 100)

Index
ActualProjected

Source: Cushman & Wakefield APAC DC Construction Cost Guide 2026.

Exhibit 9 — Construction — supply chain detail

StrategyKey adopters2026 impact
Prefab power skidsSchneider, Vertiv−40% on-site electrical labour
Modular coolingJohnson Controls (Silent-Aire)Plug-and-play liquid deployment
Standardized BIMAirTrunk, NEXTDCFabrication-ready designs before ground break

Source: Cushman & Wakefield APAC DC Construction Cost Guide 2026.

Investor implication

Construction alpha belongs to integrated operators

  • Favour NEXTDC and Leighton Asia with vertical integration or framework agreements.
  • Fixed-price contractors face margin erosion from MEP inflation.
  • Stick-built facilities risk 12–24 month slippages → delayed revenue in FY2027–28.

Glossary

CoWoS
Chip-on-Wafer-on-Substrate — TSMC's advanced chip-packaging process that bonds multiple silicon dies (logic and memory) onto one substrate. The bottleneck step for AI accelerators.
HBM4
High Bandwidth Memory, 4th generation — stacked memory chips placed next to a GPU/accelerator die for very high data throughput. The generation NVIDIA's Rubin chips require.
PUE
Power Usage Effectiveness — a data centre's total facility power divided by the power delivered to IT equipment. Closer to 1.0 means less power lost to cooling and overhead.
HVDC
High Voltage Direct Current — a power-distribution architecture that cuts conversion losses versus traditional AC, increasingly used in high-density AI data centres.
CDU
Coolant Distribution Unit — the hardware that circulates and manages liquid coolant between a facility's cooling loop and server racks in a liquid-cooled data centre.
MEP
Mechanical, Electrical, and Plumbing — the specialist trades (and the engineers who design and install them) responsible for a building's power, cooling, and utility systems. Scarcity of MEP talent is a primary driver of APAC data-centre construction cost inflation.
IST
Integrated Systems Testing — the commissioning phase where a data centre's power, cooling, and IT systems are tested together before go-live. A common source of project delay.
ICE II
Indonesia Cable Express II — Telin's submarine cable from Singapore to Manado that opens an eastward route avoiding the contested South China Sea corridor.
M2M
Machine-to-Machine — AI-era traffic where data moves between machines (data prep → training → inference) rather than user-to-cloud, favouring low-latency pathways for distributed model training.
OSM
Off-Site Manufacturing — prefabricating power and cooling modules in a factory rather than building them on-site, to compress construction schedules.
TFLOPS
Trillion Floating-Point Operations Per Second — a raw compute-speed metric for chips, increasingly less useful than efficiency metrics like tokens per watt for AI workloads.
TPU
Tensor Processing Unit — Google's custom AI accelerator chip, built in-house as an alternative to NVIDIA GPUs.
SMR
Small Modular Reactor — a smaller, factory-built nuclear reactor design pitched as a faster, more flexible way to add firm baseload power than a traditional large reactor.
ASIC
Application-Specific Integrated Circuit — a chip custom-designed for one workload (e.g. AI inference) rather than general-purpose, trading flexibility for efficiency.
BIM
Building Information Modeling — 3D digital modelling of a building's design, used to pre-fabricate and coordinate construction before ground is broken.
DC-CFA
Data Centre Colocation Framework Agreement — Singapore's regulatory regime allocating scarce data-centre power and land capacity to operators. CFA2 is its second round.

Sources & references

Running capital in this space?

Get in touch →
← Digital Infrastructure

The Worldview: Who Is Building the AI CloudIn progress

The Physical Stack: Where the Bottlenecks Are | Dorrsum