The bottleneck has moved from silicon to substations

For two years, the Asia AI trade has been a story about chips: who gets allocation, who gets export licenses, who builds the fab. That story is being overtaken by a duller but more consequential one — whether the electrons can actually get to the racks. Nowhere is this clearer than in Johor, Malaysia, which has emerged as Southeast Asia's largest data-center cluster and is now the clearest real-world test of whether a national grid can absorb hyperscale AI demand. Wood Mackenzie's June 2026 analysis puts Johor's data-center load at roughly 3.8 gigawatts, more than double its 2024-2025 level and equivalent to 1.5 times the state's entire non-data-center electricity demand 1. Generation isn't the problem — Johor has 6.8 GW of installed gas and coal capacity. The constraint is transmission: a shortage of 132kV main-intake substations and a scarcity of nodal injection points, concentrated at flagship clusters like Sedenak Tech Park and Nusajaya Tech Park 1. For investors, this reframes the Asia AI-infrastructure trade: capacity announcements are no longer a reliable proxy for deliverable megawatts, and the real due-diligence question for any Southeast Asian data-center play is grid interconnection timing, not land banking.

Malaysia has already started rationing

Putrajaya has moved from planning documents to actual rationing. In February, Prime Minister Anwar Ibrahim confirmed a roughly two-year pause on new non-AI data-center approvals, explicitly to protect the power grid and water supply and to shield domestic consumers from tariff spikes 2. AI-oriented facilities get a faster lane; everything else — crypto mining, generic colocation, enterprise hosting — is now second priority for interconnection. This is a meaningful signal: a government is now actively picking winners among data-center subtypes based on grid headroom, not merely economic development criteria. Up to 30% of new data-center applications were reportedly rejected in 2025 3. Layer on top of that a tariff reset: Peninsular Malaysia's base electricity rate rises from 39.96 to 45.62 sen/kWh under the RP4 regulatory period running through December 2026, with a new voltage-based structure that specifically loads higher network and capacity charges onto high-voltage users like data centers — potentially adding $15-20 million a year in cost for a 100MW facility 4. Operators who modeled Malaysia as a low-cost power haven need to rerun those assumptions; the cost advantage is compressing just as the approval bottleneck tightens.

The fix is years away, and it runs through Sarawak and gas

Malaysia's structural answer is the Sarawak-Peninsular Malaysia interconnection — a planned 700km submarine cable moving up to 2,000MW of Bornean hydropower to Johor, part of the 13th Malaysia Plan (2026-2030) 5. But as of mid-2026 this is pre-construction: Tenaga Nasional is still assembling a development consortium, with a feasibility study only beginning this year and no final investment decision 5. Notably, 70% of that hydropower is earmarked for export to Singapore, not domestic use — meaning Johor's data centers won't be first in line even once the cable is built. The nearer-term supply additions are Malaysia's NewGen26 gas programme (6-8 GW) and a Southern Johor renewable corridor targeting up to 4 GWp of solar with storage 1, alongside roughly 2.1 GW of coal capacity scheduled for retirement in the mid-2030s that will need replacing even as demand climbs. The realistic read for investors: Johor's power crunch is a multi-year overhang, not a 2026-2027 problem that resolves itself. Capital tied to projects promising near-term energization in Johor should be discounted for interconnection risk specifically, separate from the usual construction and equipment risk.

Singapore and India show two different playbooks

The contrast with neighbors is instructive. Singapore, which capped data-center growth for years over grid and land constraints, reopened capacity in a more disciplined way: its DC-CFA2 tender released at least 200MW of new capacity in early 2026, conditional on at least 50% green power sourcing and PUE below 1.3, while a 700MW low-carbon data-center park is being planned for Jurong Island 6. That's a supply-constrained market rationing access by efficiency and decarbonization credentials rather than simply pausing approvals — a signal that premium, well-capitalized operators with strong ESG stories will keep winning allocation even as raw capacity stays tight. India is pursuing volume over rationing: Adani Group's $100 billion data-center pledge and Uttar Pradesh's new Data Centre Policy targeting 2GW of capacity are backed by power-purchase agreements that already source 30-35% of data-center electricity from renewables, with some operators going fully green 7 8. India's binding constraint is less approvals than the gap between renewable generation and transmission buildout — a lag that echoes Johor's substation shortage but at an earlier, more fixable stage of infrastructure maturity.

What this means for positioning

Three implications follow. First, the investable proxy for Asia's AI buildout is shifting from chip and cloud-capex headlines toward grid operators, transmission-equipment suppliers, and gas/renewable generation names with credible delivery timelines in Johor, Peninsular Malaysia, and adjacent Southeast Asian grids — Tenaga Nasional's consortium-building and PowerChina/Huawei-linked infrastructure work are worth tracking as bellwethers 1 9. Second, hyperscaler and neocloud capex guidance for Southeast Asia should be read skeptically where it assumes uninterrupted interconnection; Malaysia's rationing policy shows governments will intervene when grids are stressed, and similar moves are plausible elsewhere in the region as loads scale. Third, the widening gap between AI-tagged and non-AI data-center treatment in Malaysia previews a broader regulatory pattern — expect other Asian grid operators to formalize similar tiering, which will advantage large, well-capitalized AI-specific developers (NVIDIA-linked, hyperscaler-anchored) over generalist colocation players who lack the political and commercial leverage to jump the queue 9. The AI infrastructure trade in Asia is increasingly a power and permitting trade, not a compute trade.