Roughly $7 trillion in AI infrastructure spending is expected globally through 2030, according to converging estimates from Goldman Sachs, McKinsey, and Nvidia. The scale is no longer in question. Where it lands is.

Look at the announcements and a pattern jumps out. Microsoft has committed $17.5 billion to India through 2029. Blackstone and CPP Investments paid roughly $16 billion for AirTrunk, the APAC data center platform. Adani is pledging $100 billion to Indian AI infrastructure by 2035. Gulf commitments from Microsoft, AWS, and Google Cloud run into the tens of billions. Johor, in southern Malaysia, saw operational data center capacity grow 53% year-on-year in 2025; Melbourne grew 37%.

Now weigh that against population. Indonesia, Vietnam, Thailand, and the Philippines are not being ignored—Amazon has put roughly $5 billion each into Indonesia and Thailand, for instance, but relative to their populations and fast-growing user bases, the capital arriving is thin. The concentrations are forming instead in places with little in common: Johor has cheap power and land. Melbourne and the Gulf are expensive and distant from Asian demand. If cost or geography explained the pattern, similar markets would keep winning. They don’t.

Overlay a map of where AI capital concentrates onto a map of how the region’s address space is governed, and the same shape appears in both. APAC was the first region to exhaust its free IPv4 pool, in 2011—years ahead of ARIN and RIPE NCC—and it is the only RIR that interposes national registries, each incorporated under domestic law, between the regional registry and the network operator. That combination behaves like a tariff paid in delay and regulatory ambiguity rather than dollars, and it’s rarely discussed as an input to where capital lands.

The IPv4 layer

APNIC moved to a final /8 rationing policy in 2011: a fixed allocation regardless of need, currently capped at a /23 (512 addresses) per account under prop-127. This isn’t a supply problem—APNIC still held roughly 3.1 million addresses as of December 2025, and even proposals to loosen the cap don’t project exhaustion before the mid-2030s. It’s an allocation-size problem: a /23 runs a single network, not a hyperscale campus. Prop-168, which would have raised the ceiling to a /22, failed to reach consensus at APNIC 61 in February 2026. The community has consistently favored conserving the pool for new entrants over enlarging allocations for growing networks, pushing the latter toward the secondary market.

That’s where the region diverges further. Every RIR now caps new allocations—ARIN at a /22, RIPE NCC at a /24, but APNIC’s policy is silent on leasing. Transfers are explicitly provided for; leasing is neither authorized nor prohibited. LACNIC sits in the same undefined space, and unlike APNIC, it has an active proposal (LAC-2025-5) to formalize both mechanisms. APNIC has no equivalent under consideration. Compounding this, large blocks are essentially unavailable on the open market. IPv4.Global’s listings typically run from zero to a dozen small blocks at a time, so an operator hitting the cap has no meaningful way to buy its way past it.

Seven registries, one anomaly

Elsewhere, an operator seeking address space deals with one registry under one community-developed policy set. APNIC operates through seven National Internet Registries—CNNIC, IRINN/NIXI, IDNIC-APJII, VNNIC, JPNIC, KRNIC, and TWNIC—each applying its own membership criteria, pricing, and IPv6 rollout pace. Three are industry associations; four are government bodies. VNNIC is a unit of Vietnam’s Ministry of Information and Communications. CNNIC has answered to China’s Central Cyberspace Affairs Commission since 2014, the same body that oversees the country’s censorship apparatus.

In China, that commission also sits above data residency and cybersecurity rules (PIPL, the Cybersecurity Law) and generative AI licensing (the CAC’s Interim Measures, effective August 2023). The body that grants a data center’s IP addresses is institutionally the same body that grants its customers the right to run AI services at all. China resolved the resulting risk not by fixing it but by removing foreign capital from the picture—its June 2026 industrial plan directs roughly $295 billion in state spending to AI data centers, with an 80% domestic-supplier mandate. It isn’t paying this tariff; it opted out of the market that charges it.

Vietnam shares China’s institutional shape but, unlike China, wants foreign hyperscaler capital, so the overlap between IP allocation and content rules stays priced into every commitment. Indonesia, Thailand, and the Philippines lack that formal alignment but carry something subtler: regulatory cultures governed by unwritten norms, where the absence of explicit permission still reads as risk. That shows up starkly in the cross-border IPv4 leasing market, where deals routinely stall not on price or technical fit but because a prospective lessee cannot get anyone formally empowered to confirm the transaction is permitted.

Where the capital actually goes

The regional split bears this out. North America still absorbs the largest share of cumulative AI infrastructure investment. APAC’s share is large in absolute terms but small relative to its population—the region holds over half the world’s internet users. Within APAC, capital concentrates in India, Singapore, Malaysia, Japan, Korea, and Australia, while Indonesia (280 million people), Vietnam (100 million), the Philippines (115 million), and Thailand (70 million) collectively draw a fraction of their demographic weight.

Power and geopolitics explain part of this, but not all of it. Johor, a Malaysian state of roughly four million people, is outgrowing Indonesia’s build-out on every metric that matters, despite sitting in the same time zone and power-cost neighborhood. The difference is that Johor operates under a written, if restrictive, regulatory framework, while Indonesia’s is unwritten.

India offers the clearest counterexample within APNIC’s own footprint: it runs through IRINN, a government-backed body, and by the structural logic above should face the same friction Indonesia does. It largely doesn’t, because India has deliberately built written law, transparent allocation, and a codified data-protection regime (DPDP) over time. Korea is a partial exception in the other direction. Its registry sits inside a government security agency, yet gates no AI workloads to the addresses it issues, so capital still reads it as legible.

Conclusion

Power availability, geopolitics, and language explain part of where AI infrastructure lands in Asia. What they don’t explain is why markets with the population and demand to anchor the build-out keep underperforming it. Three foundational layers—IPv4 allocation policy, the national-registry structure unique to APNIC, and the legal-political overlap between address governance and content governance—account for the rest.

Capital reads a jurisdiction’s foundations in months. Building those foundations takes years. Every quarter a country leaves its rules unwritten is a quarter the build-out commits somewhere more legible instead, into infrastructure that will then run for a decade. The window to fix that before the money moves on is closing.

This article is part of a bigger opinion piece. Read the full story on IPXO’s blog.