Keppel and Microsoft flag power infrastructure as the binding constraint on AI data-center growth
Редакция ЦОД:Главное
At Shell’s Co-Creating Progress Future forum in Singapore, executives from Keppel and Microsoft said AI-driven data-center demand is advancing faster than the power and other physical infrastructure required to support it. Keppel also said its floating data-center project took more than five years of development work before entering construction.
- Keppel and Microsoft said AI infrastructure demand is outpacing available power and other supporting infrastructure.
- Keppel said regulators are increasingly asking data-center and AI operators to secure additional power rather than depend fully on the grid.
- Microsoft said it absorbs additional energy costs associated with data-center development rather than transferring them to local consumers.
- Keppel’s floating data-center project entered construction after more than five years of preparation.
- Rapid GPU refresh cycles may require infrastructure systems inside data centers to be replaced well before the building reaches the end of its design life.
Keppel and Microsoft executives have identified power availability and broader physical infrastructure as the principal bottlenecks to scaling AI data-center capacity.
Speaking on 9 October at Shell’s Co-Creating Progress Future forum at Marina Bay Sands in Singapore, Wong Wai Meng, chief executive of Keppel’s data-center business, said existing power systems are insufficient for the AI demand Keppel is seeing. He said AI infrastructure must be planned as an interconnected stack spanning power, water, data centers, chips, applications and AI models.
Wong said regulators in multiple countries are increasingly requiring data-center operators and AI companies to bring additional power resources rather than rely entirely on the shared grid. Mark Souza of Microsoft said the company bears incremental energy costs associated with building data centers rather than passing those costs to local consumers and communities.
Souza also said demand for AI infrastructure exceeds supply and that the supply shortfall is likely to persist for several years. That is an executive assessment rather than a defined market forecast.
Keppel’s floating data-center project illustrates the extended development cycle associated with unconventional infrastructure. Wong said the project moved into construction only after more than five years of preparation. Keppel did not disclose the project’s capacity, CAPEX, precise location, construction start date or expected commissioning date.
The panel also highlighted the mismatch between long-lived facilities and rapid compute-platform refresh cycles. Wong said some technologies within newly built data centers may need replacement after about five years as successive GPU generations emerge, while a data center built seven or eight years ago could have a design life of up to 30 years. These were illustrative observations, not specifications for a particular project.
For operators and investors, the comments suggest that sites with secured power resources, fuel logistics and network infrastructure could become more valuable. If self-provisioning of incremental power becomes a more common regulatory expectation, projects may face higher capital requirements and longer delivery timelines. More frequent technology refreshes could also support investment in modular cooling and electrical systems and more flexible capital-planning models.
The discussion underscores a core constraint on AI data-center deployment: compute demand is scaling faster than generation, fuel and grid infrastructure. A shift toward requirements for operators to secure incremental power could raise project capital intensity and development lead times, while faster GPU refresh cycles may favour modular power and cooling architectures.
- — / 14 октября 2026 г. · КитайскийПервоисточник →
Материал подготовлен редакцией ЦОД: Главное на основе указанных первоисточников.