APAC AI spending is outgrowing the budgets set in 2024

In September 2024, IDC forecast that APAC AI spending would reach US$110 billion by 2028. In September 2026, the same spending guide put the number at US$555 billion for 2030 (IDC, 2026). Both are forecasts. But many IT budgets in the region were built on the 2024 view, and the bill now arriving contains items those budgets did not price in.

The cost driver is not the AI licence alone. It is power and floor space in a few hubs, regional price premiums at the hyperscalers, data that has to stay in-country, and contracts denominated in US dollars. Each is manageable alone. Together they break cost cases built on US reference prices.

APAC AI spending is growing faster than the forecasts behind 2024 budgets

The 2024 IDC release projected a 24% compound annual growth rate for AI and generative AI in Asia Pacific between 2023 and 2028 (IDC, 2024). The 2026 guide puts regional spend at US$121.5 billion in 2025 and expects 35.5% annual growth to 2030. The largest single use case is AI infrastructure provisioning, at about 32% of total spending.

Gartner points the same way. It forecasts APAC IT spending of US$1.78 trillion in 2026, up 11.8%, with data centre systems rising 63.1% to US$232 billion (iTnews Asia reporting Gartner, April 2026). Enterprises pay for that capex later, through cloud rates and colocation contracts.

Power and land constraints turn data centre capacity into a premium

Singapore stopped approving new data centres in 2019 and reopened only through controlled allocation. A pilot call in 2022 awarded 80 megawatts (MW) to 4 operators in 2023. The Green Data Centre Roadmap of May 2024 targets at least 300 MW of additional capacity in the near term, on top of more than 1.4 gigawatts across over 70 facilities (IMDA, 2024). The second allocation round, announced in December 2025, offers at least 200 MW and requires at least 50% green power (Morgan Lewis, March 2026).

The obvious overflow site is Johor, across the causeway. In November 2025 the state stopped approving Tier 1 and Tier 2 data centres, routed applications through a committee of 20 agencies and required reclaimed water (New Straits Times, November 2025). Scarcity in these hubs is a policy choice, and it is priced into every rack and cloud hour sold there.

The same AI workload costs more in APAC cloud regions than in the US

We checked list prices on the official Azure Retail Prices API on October 8, 2026, for Linux pay-as-you-go. A general purpose D4s v5 virtual machine costs US$0.192 per hour in East US, and US$0.24 in Southeast Asia (Singapore) and Australia East. That is 25% more. For AI capacity the gap widens.

Azure region	NC24ads A100 v4, US$ per hour	Premium vs. East US
East US	3.673	Baseline
Southeast Asia (Singapore)	4.775	+30%
Central India (Pune)	5.142	+40%
Japan East (Tokyo)	5.326	+45%
Australia East (Sydney)	5.326	+45%

These are list prices before reservations or enterprise discounts, and other GPU families differ. Still, a cost case built on US prices understates APAC GPU cost by roughly a third. Over a year of continuous use, a single A100 instance in Tokyo costs about US$14,500 more than the same instance in Virginia.

Data residency rules keep workloads in the expensive regions

Moving inference to a cheaper region is often not allowed. The Reserve Bank of India has required since April 2018 that payment system data be stored only in India; processing abroad is permitted, but the data must be deleted there within 1 business day or 24 hours (Reserve Bank of India FAQ).

For AI this matters more than for classic workloads. Retrieval pipelines, vector indexes, prompt logs and fine tuning data all contain the regulated records they draw on. If the data must stay local, so must most of the AI stack, at the local price.

USD licences and local budgets carry a currency risk most cost cases ignore

Microsoft aligns its cloud prices to US dollar levels and reviews local currency prices twice a year. In 2024 that meant increases of 20% in Japanese yen, 8% in Korean won and 6% in Indian rupees for cloud services (Microsoft, December 2023). Local currency budgets absorb that without any change in usage.

On top of that come list price changes. From July 1, 2026, Microsoft 365 E3 rose from US$36 to US$39 per user per month and E5 from US$57 to US$60, with Copilot Chat included. The full Microsoft 365 Copilot licence stays an add-on at US$30 per user per month (Directions on Microsoft, December 2025). For a group with 3,000 users, Copilot for everyone adds about US$1.08 million a year at list price, before currency effects. Our view on that renewal decision is in Microsoft 365 Copilot ROI at renewal.

What to do now

1. CFO: Split every AI cost line into 2 buckets. Value spend has a named business owner and a metric agreed before go-live. Capacity premium covers regional uplift, residency, idle reservations and currency. Report both quarterly.
2. CIO: Reprice all AI cost cases with APAC list prices for the region the data must live in, not US reference prices. Add a currency buffer that matches the vendor review cycle.
3. CISO and Legal: Map which datasets in each AI use case are subject to residency rules. Only those need the local region.
4. IT Operations: Measure GPU and Copilot utilisation monthly. Reserved capacity below 60% use and licences without regular activity go back on the table at the next true-up.
5. Procurement: Negotiate price protection in local currency or fixed exchange rates for multi-year AI commitments, and ask for regional uplift to be shown as its own line.

We start AI cost work with read access to billing, licences and usage, separate the capacity premium from spend that moves a business number, and report against the client's own figures. That is the core of our Cloud & AI Cost Reduction work, and the starting point of the Executive IT & AI Review. Who should own the number internally is covered in AI cost management needs an owner.