As AI workloads push rack densities higher and compress hardware refresh cycles, data centre design is being forced to change. Power availability, cooling limits, and location constraints are now shaping where AI infrastructure can realistically be built and how it operates.
In this interview, Mayank Srivastava, CEO of BDx Data Centers, discusses how these pressures are driving the move toward so-called “AI factories,” why energy constraints now outweigh chips as the limiting factor for AI growth across the region, and how emerging technologies such as quantum computing could affect future data-centre design and workload distribution.
How are data centre architectures evolving as AI models scale and “AI factories” emerge?
The shift to “AI factories” requires a fundamental rethink of data centre architecture, largely because the lifecycle of AI hardware no longer aligns with the 20- to 30-year lifespan of a facility. Nvidia CEO Jensen Huang said at last year’s GTC that H100 GPUs are already obsolete, and today’s GB200 systems are expected to give way in the near future to chips drawing 500 kW to as much as 1 MW per rack. With hardware evolving at this pace, flexibility has become the most critical factor in design.
One of the most immediate pressures is power and cooling density. Traditional air cooling is inadequate for the thermal output of modern GPUs, which is why the industry is moving toward direct-to-chip liquid cooling. This approach is significantly more efficient and can reduce cooling-related energy consumption by up to 50%. Modern AI data centres are increasingly being designed to support rack densities of 200 kW or more to accommodate large training clusters.
A second structural shift is the move toward modular design. Facilities must be built with a modular approach to allow for rapid scalability and adaptation to new hardware. Various AI-focused data centres are now designed around reference architectures such as Nvidia’s DGX SuperPOD, with mechanical and electrical systems that support fault tolerance and allow maintenance to be carried out without disrupting active workloads. This allows the data centre’s infrastructure to remain usable after the initial chips it was built for have become redundant.
How do power constraints shape BDx’s regional capacity deployment?
Power availability has become the primary limiting factor for AI growth, rather than chips or software. As a result, how and where we deploy capacity is shaped by the energy landscape of each market.
In power-rich markets such as Indonesia, which has a national power surplus, we can invest in large-scale AI campuses. We secure reliable, long-term capacity through direct agreements with national providers. We also prioritise access to clean energy, which is why our 500 MW CGK4 campus is powered by renewable hydropower.
In power-constrained markets such as Singapore, where new capacity is limited, we have to take a different approach. Singapore is projected to add more than 300 MW of data centre capacity, which makes direct power sourcing more challenging. Our partnership with HEXA Renewables supports the development of 50 MW of new solar projects in Malaysia intended to supply the regional grid. This creates additionality by introducing new green energy that would not otherwise exist, allowing us to align hyperscale growth with regional decarbonisation goals.
Why is Indonesia emerging as a hyperscale hub for AI infrastructure?
While its power surplus is a significant advantage, Indonesia’s momentum is also shaped by a combination of demographic, political, and economic factors.
One driver is digital demand. Indonesia has a large, young, and highly tech-engaged population, which is driving growth in data consumption and digital services. The expansion of sectors such as e-commerce and fintech, led by platforms including Tokopedia and Shopee, has increased demand for AI capabilities used in areas such as fraud detection and personalised user recommendations.
Government policy is another factor. The Indonesian government is promoting digital transformation through initiatives such as the “Making Indonesia 4.0” roadmap and the “Golden Indonesia 2045” vision, which has contributed to a more predictable environment for long-term investment in technology infrastructure.
Regulation also plays a role. Data sovereignty requirements, including Government Regulation No. 71, mandate domestic data processing and storage for public services. This creates built-in demand for in-country AI infrastructure.
How is rising GPU demand reshaping colocation, cloud, and AI facilities?
The direction is toward a hybrid and distributed model that combines all three. The increasing demands of GPU-based workloads are making a one-size-fits-all approach impractical, and the balance is increasingly defined by the type of AI workload involved.
Large-scale AI model training is highly power-intensive and will continue to concentrate in specialised, purpose-built facilities. These environments are designed to support high-density compute and tend to be located where power is available at scale, at lower cost, and with access to cleaner energy sources.
In contrast, AI inference, the real-time application of trained models, must be pushed closer to users to ensure low latency. This is where cloud platforms and distributed edge infrastructure play a larger role.
How does BDx’s quantum AI testbed shape the role of quantum computing in data-centre design and workload distribution?
The practical role of our hybrid quantum AI testbed is to provide a platform for experimentation, allowing organisations to explore quantum-enhanced AI applications without waiting for the technology to fully mature. It enables government agencies, enterprises, and start-ups in Singapore and across the region to test early use cases and understand where quantum computing may offer practical benefits.
While quantum computing is still emerging, we see it being applied to specific classes of problems that are difficult for conventional computing systems to address. In finance, for example, HSBC has tested quantum systems to improve the accuracy of its trading models, reporting gains of up to 34% in trial conditions. In security, quantum techniques are being explored for new approaches to secure communications. In chemistry and pharmaceuticals, quantum simulation is being used to model molecules for drug discovery and materials research.
From a data-centre perspective, quantum computing represents a significant architectural shift. One notable aspect is its potential energy efficiency, where relatively low power consumption can deliver substantial computational capability compared with large GPU-based systems.
At the same time, quantum systems introduce major operational challenges, particularly the need for cryogenic cooling environments operating near absolute zero (–273°C) alongside conventional high-density infrastructure.
The testbed is intended to explore how these systems can be integrated within a data-centre environment and what that means for future facility design and workload distribution.
