AI capital floods power-hungry data hubs

Artificial intelligence is redrawing the economics of data centres at a speed that is forcing developers, utilities and investors to rethink where digital infrastructure can be built and how it can be financed. A new 2026 marketplace report from Colliers says global investment in data centres topped $580 billion in 2025, up 27 per cent from a year earlier, as hyperscalers and infrastructure groups poured money into facilities built for AI training and inference. The same report says more than $120 billion of hyperscaler debt was issued to support AI infrastructure, underlining how rapidly capital markets have moved behind the sector.

That spending boom is being matched by an equally sharp rise in physical constraints. Colliers says build costs climbed 47 per cent year on year, with power infrastructure now accounting for 40 to 50 per cent of total project costs. In North America, absorption reached 15.6 gigawatts, double 2024 levels, while more than 90 per cent of new capacity was pre-leased before delivery. Those figures point to a market in which demand is outrunning supply long before new buildings are switched on. Colliers’ wider global analysis adds that the United States alone faces an unmet demand gap of about 19GW, showing how AI has shifted data-centre development from a real-estate story into a power and industrial-supply story.

The centre of gravity in the market remains the largest cloud and platform companies. Reuters Breakingviews, citing Morgan Stanley, said Amazon, Microsoft, Alphabet and Meta are projected to spend about $630 billion on data centres and AI chips in 2026 alone. It added that the top 11 cloud and infrastructure providers could together reach $811 billion in capital expenditure. That helps explain why suppliers of land, transformers, cooling systems and substations have become critical gatekeepers in the AI race. A modern 100-megawatt AI data centre can cost more than $4 billion including chips, Reuters reported, with roughly 70 per cent of spending tied to servers and graphics processors.

Yet money is not eliminating bottlenecks. Reuters reported that securing grid access in major hubs such as London can take up to a decade, while transformer lead times in Europe can stretch to around 100 weeks. Nearly 60 per cent of data-centre projects were delayed by more than three months last year, according to nPlan data cited by Reuters, with setbacks hitting everything from foundations to cooling systems and fire-safety installation. Nvidia’s more powerful chip designs are intensifying the problem because they generate more heat and require liquid-cooling systems, new plumbing and more sophisticated electrical architecture. What is emerging is a market in which computing demand can be financed faster than physical infrastructure can be delivered.

Power availability is now the decisive factor in site selection. Reuters reported in February that 46 data centres in the United States were already planning to build their own power plants, mostly gas-fired, representing 56GW of capacity, or about 30 per cent of all planned U. S. data-centre capacity identified by consultancy Cleanview. Utilities are adapting quickly. AES said in February it had signed agreements for nearly 12GW of energy with data-centre customers, including 9GW of power purchase agreements directly with hyperscalers. The company said its Texas arrangement with Google would bring new generation alongside a facility so expansion could proceed without placing as much strain on the local grid.

Large operators are also reshaping commercial risk. In Louisiana, Entergy said this week that Meta would cover the full cost of service for its planned $10 billion hyperscale facility in Richland Parish under a revised agreement expected to deliver nearly $2 billion in savings to customers over 20 years. The arrangement illustrates a broader shift: hyperscalers are increasingly being asked to underwrite the energy and transmission build-out needed for AI campuses rather than relying on legacy grid models. That may reduce political resistance in some regions, but it also highlights how expensive and complex the next generation of AI sites has become.

Another layer of the market is being built by so-called neocloud providers, specialised groups that rent out GPU-heavy compute to larger platforms and enterprise users. BloombergNEF said leases signed by hyperscalers for capacity provided by neoclouds over the six months to March 2026 could together be worth more than $100 billion. That offers a route to faster scale, especially where major platforms need capacity before their own campuses are completed. But it also introduces a risk profile that differs from traditional long-duration infrastructure because some of those contracts are shorter than the useful life of the assets being financed.