What’s changing in AI infrastructure across Asia/Pacific through 2030?
In Asia/Pacific, the AI race is no longer about who has the best model. It’s about who controls the stack underneath it. Three layers are being decided at once. Compute is shifting to always-on inference running on a mix of chips. Power and placement are becoming hard limits, with Asia/Pacific excluding Japan datacenter IT power capacity reaching 70.6GW by 2030. And trust is now part of the design, not an afterthought.
Why It Matters
Agentic AI isn’t a side project anymore. It’s running inside real workflows, around the clock. Can it run nonstop? Can you afford it? Is it sitting in the right country?
If you’re a CIO, CTO, or infrastructure leader, this shapes some big calls: where your inference runs and which platforms and partners you commit to.
Here’s the part that’s easy to miss. The choices you make in 2026 and 2027 will follow you well into the 2030s, because chips get replaced every few years and power contracts, sites and sovereignty setups stick around for a decade or more.
Market Dynamics
Compute: inference is taking over. Training got the headlines, but inference is where the spending is heading. In Asia/Pacific, 94.5% of organizations were using inference infrastructure in 2025, up from 40.2% just two years earlier. IDC expects AI hardware spending in Asia/Pacific excluding Japan to quadruple, from US$66.3 billion in 2025 to US$265.4 billion by 2030. And don’t expect one standard architecture across the region. India runs the widest mix of NVIDIA, AMD, and hyperscaler chips. Korea leans on NVIDIA and AMD. Greater China is the only market where AMD is ahead of NVIDIA, and where DeepSeek and Huawei Ascend have gone mainstream.
Power and placement: location beats price. More and more, governments decide where capacity goes. Singapore has set aside a 700MW site in Jurong, and South Korea is planning 3GW of AI clusters outside Seoul. Then there’s cooling. In Singapore’s heat, cooling can take up to 40% of a facility’s energy, and typical colocation PUE (1.5–1.7) is still above the government’s 1.3 target. Microsoft’s Indonesia region shows where this is heading: it’s built for in-country inference and regulated workloads, while heavy training stays in a few global hubs.
Trust: sovereignty is now built in. IDC’s Asia/Pacific Emerging AI Technologies Survey 2025 found that security, data residency, responsible AI and open standards now matter more than model performance when organizations pick a sovereign AI partner. Most aren’t all-or-nothing on residency: 27% apply strict rules to all data, 52% only to sensitive data and 20% stay flexible. In the same survey, more than 60% of enterprises say changing privacy, security and AI rules are already disrupting their IT operations.
The catch: there’s no one-size-fits-all. Rules, power grids and buyer priorities change from one border to the next, so placement and sovereignty have to be planned market by market.
Data Callout
- 70.6GW: datacenter IT power in Asia/Pacific excluding Japan by 2030
- US$66.3 billion to US$265.4 billion: AI hardware spending in Asia/Pacific excluding Japan from 2025 to 2030
- 79%: organizations that require data residency, at least for sensitive data
- More than 60%: enterprises report IT disruption from changing regulations
Analyst Insight
The mistake I see most often is treating compute, power and trust as three separate decisions, made one after the other. They’re connected. The chips you pick set your power density. That density decides which facilities can host you. And where that facility sits decides which trust controls you can offer.
Organizations that plan all three together will scale faster. And in Asia/Pacific, placement is the decision that’s hardest to undo.
Outlook
My view is that placement will drive the next wave of growth. Big training runs will stay in a few hubs, while inference moves in-country, close to users and to data that can’t leave.
What could speed things up? Government power allocations turning into usable capacity on time, open liquid cooling standards that make retrofits cheaper, and sovereign cloud regions that make it easier to get from pilot to production.
What could slow things down? Grid delays, not enough renewable energy, chip shortages, rising power bills and a different rulebook in every country.
Three things I’ll be watching next quarter:
- How much new capacity is built for high-density AI racks
- Whether models like DeepSeek catch on beyond India and Greater China
- Whether sovereign AI deals move from pilots to multiyear contracts
FAQs
If power is so tight, why is AI infrastructure investment still growing? Agentic AI turns inference into an always-on workload, so the demand isn’t going away. Tight power pushes that demand to places with power available, like Johor, Batam, India and Thailand, and into denser, liquid-cooled facilities.
Who benefits most? Service providers. Digital and cloud service providers are the fastest-growing datacenter segments through 2030. Colocation players with secured power, liquid cooling and sovereignty-ready capacity are best placed to win inference demand.
What could go wrong? Grid delays, chip shortages, rising energy costs and diverging data rules. Any one of them could push back new capacity or force country-by-country designs.
Learn more about IDC’s predictions for the next 5 years at IDC FutureScape 2027 in Singapore on 13 November 2026, where I’ll walk through IDC’s full view to 2030 in my session, “AI Infrastructure Race: The Compute, Power, and Trust Stack Reshaping Asia/Pacific.” Visit the IDC FutureScape 2027 event website for more information and register for the event today.
IDC FutureScape presents the top technology predictions for the coming 1–5 years in a particular focus area — written from a worldwide, regional (Asia/Pacific), or country-level view (China, Japan). Each report translates IDC’s market intelligence into actionable guidance, helping technology and business leaders shape their priorities, roadmaps, and investment decisions to achieve desired business outcomes. Explore how AI is rewriting the stack in the 25 IDC FutureScape 2027 reports that map every layer.
Data Sources: IDC Semiannual Datacenter Facilities Tracker and Worldwide AI and Generative AI Spending Guide (latest forecast) and other IDC Research, 2026
