Computer Weekly: datacentre squeeze hits APAC
- Computer Weekly says APAC’s AI datacentre boom has hit a hard physical limit on May 11, 2026 — power, not land, now decides builds. - The clearest number is $800 billion: Deloitte’s estimate for Asia-Pacific datacentre investment by 2030, colliding with grid delays, permits, and scarce AI-ready space. - That matters because GPU expansion now depends less on chip supply and more on electricity, queueing, and where workloads can move.
Datacentres are having a very old-school problem. Not chips. Not software. Electricity. That is the real news in Computer Weekly’s May 11 piece — the AI buildout across Asia-Pacific is running into hard limits in grid capacity, connection timelines, and site readiness, even as planned investment keeps ballooning. The result is simple but brutal: more money wants in than the physical system can absorb. ### Why is power the bottleneck? A modern AI datacentre is basically a machine for turning electricity into tokens. Computer Weekly’s core point is that operators are now picking sites based first on energy access, not tax breaks or geography. In APAC, Deloitte’s projected $800 billion of datacentre investment by 2030 is crashing into the fact that many grids cannot add capacity fast enough. Global datacentre electricity demand was about 460 TWh in 2022 and is expected to top 1,000 TWh by 2026. (computerweekly.com) ### Why does AI make this worse? AI racks are much denser than older enterprise compute. CBRE’s figures, cited in wider coverage of the same regional trend, put AI-optimized racks at more than 40 kW each, versus roughly 5 to 15 kW for legacy setups. That changes everything — cooling, floor loading, network design, and especially utility hookups. A site that looked fine for cloud workloads can turn out to be the wrong building for GPU clusters. (computerweekly.com) ### How bad is the APAC gap? Pretty bad. One estimate puts Asia-Pacific on track for a 15 to 25 GW datacentre capacity shortfall by 2028, equal to roughly 25% to 42% of forecast demand. That is despite supply expanding sharply. So this is not a story about no construction happening. It is a story about demand rising even faster than construction, while power and permitting slow delivery. (networkworld.com) ### Which markets are getting squeezed first? The established hubs are feeling it first. Singapore has already shown what constraint looks like in practice — a pause in new approvals from 2019 to 2022, and even now tight planning conditions remain. Computer Weekly says Meta-linked projects have faced delays there, with officials effectively signaling that some energy won’t be available before 2028. Tokyo has another version of the same problem — grid access timelines stretching to 48 to 60 months. (networkworld.com) ### So where does capacity go instead? It starts drifting to places with cheaper, more available power and fewer land constraints. CBRE has been flagging Mumbai and Seoul as beneficiaries, while Computer Weekly points to India as a major greenfield target, including Reliance Industries’ huge AI datacentre plans and Microsoft’s continued regional expansion. Basically, the map of AI infrastructure is being redrawn by substations and transmission lines. (computerweekly.com) ### Can operators just build around the grid? Only partly. The article points to on-site generators, microgrids, and battery storage. Broader industry work also points to cleaner but harder options — hydrogen, ammonia, nuclear, and bigger storage systems. But these are expensive and slow. They help at the margin. They do not magically create utility-scale power where the grid is years behind. (computerweekly.com) ### What does this change for buyers? It means compute planning gets more operational. If GPUs are scarce and datacentre slots are pre-leased years ahead, companies have to queue workloads, move jobs between regions, and be much stricter about what deserves premium AI infrastructure. The hidden constraint is no longer just “Can I buy the accelerator?” but “Can I get powered, cooled space for it?” Pre-construction leasing three to four years ahead is already becoming normal in tight markets. (computerweekly.com) ### Bottom line? The APAC datacentre squeeze is not a temporary hiccup. It looks structural. AI demand is rising into a physical system — grids, permits, substations, cooling water — that moves much slower than software hype. So the winners are not just the companies with chips or capital. They are the ones that can actually secure megawatts. (computerweekly.com) (dcbyte.com)
