Who’s really solving AI’s data center energy crisis in Asia?

Each new chatbot response or AI-generated video comes with a hidden cost: the massive power needed to run the servers behind them. Data centers now consume so much electricity that they are straining grids in US cities and raising questions about sustainability in Asia, where many facilities still run on fossil-heavy energy mixes.

But the next breakthrough in managing AI’s energy crunch in Asia may not come from Big Tech. Instead, the brunt may fall to colocation providers, the companies that operate these massive facilities.

In the US, hyperscalers like AWS or Google Cloud often build and run their own sites. But in Asia, they rely heavily on third parties, renting facilities from colocation providers like AirTrunk and Iron Mountain. According to S&P Global, this helps them distribute the upfront costs while still providing services to end-users.

Image credit: Timmy Loen

S&P notes that Asia Pacific now leads the world in leased data center capacity.

This dependency is opening doors for startups such as Singapore’s Red Dot Analytics, UK’s EkkoSense, and US-based Phaidra to work their tech into the AI goldrush without actually selling to the hyperscalers.

Collectively, these firms are chasing a market that’s worth more than US$3.2 billion and is projected to more than double by 2030 as operators look for ways to cut ballooning power bills and meet climate commitments.

The spiky load problem

By 2030, the demand for global data center capacity is expected to reach 219 gigawatts, up from 82 gigawatts in 2025, according to an estimate by McKinsey. Of this, 156 gigawatts, or about 71% of the total, would come from AI workloads.

To put that into context: If AI data centers run at full load all year, 156 gigawatts of capacity would require approximately 1,366 terawatt-hours of power annually.

Of course, data centers never run at 100% load all the time, but even at 50%, the figure would still be over 12x higher than Singapore’s total electricity consumption in 2023 (55 terawatt-hours).

See also: Malaysia’s thirst for AI data centers could leave it high and dry

While training AI models is energy intensive, the computing load is relatively predictable and easy to optimize. On the other hand, inference workloads – which refers to the processing required for AI models to generate a response – are more volatile.

McKinsey expects that inference will become the dominant AI workload by 2030.

But despite the continued work of AI giants like OpenAI in optimizing and reducing the computational power required by their models, the variety of ways in which people are using AI means optimization can’t just stop at model design or new chips.

Making AI run efficiently with AI

Keeping costs down is just the start