AI's Next Bottleneck Is Power, Water and Grid Capacity

Artificial intelligence now ranks with the steam engine, electricity and the internet as a general-purpose technology. Its weight comes less from the AI industry than from how it spreads. IMF research estimates GCC AI investment could lift real GDP growth by 0.1 percentage point a year over the next decade. BNEF reckons R&D staff remain AI development’s largest single cost. The binding constraint sits underneath: power, water and grids.

Global data-center electricity use was about 500 TWh in 2025 and is projected to grow ninefold by 2050. The IEA projects global electrification of about 35 percent by 2035, driven by electric vehicles, cooling demand and data centers. BNEF finds energy availability, not land, fiber or tax regimes, is the binding constraint; hundreds of megawatts of reliable baseload are hard to assemble. In Hong Kong, Singapore and the GCC, data centers could take 10–15 percent of peak urban grid demand, and close to 25 percent in the United States. A facility can be built in two to three years; grid connections and clean generation take roughly twice as long.

Density sharpens the squeeze. Rack densities are moving from 8 kW toward 1,000 kW, making heat rejection and water management as limiting as chip supply. In the hot, water-stressed Gulf, cooling can collide with energy transition targets.

Ambition ahead of the environmental ledger

The UAE and Saudi Arabia show the tension most clearly: sovereign capital is being committed at gigawatt scale, often ahead of the clean power and water systems meant to support it.

ADNOC and Masdar signed a 2024 agreement with Microsoft to explore supplying its data centers from Masdar’s clean-energy portfolio. Barakah, at 5.6 GW and roughly a quarter of UAE electricity, is increasingly treated as AI’s baseload. DEWA’s solar-powered Moro Hub in Warsan won approval for a second phase in January 2026. Khazna Data Centers is testing treated sewage effluent instead of potable water for cooling. MGX closed MGX Fund 1 at $49 billion in mid-2026, linked to the $100 billion Global AI Infrastructure Partnership with Microsoft, BlackRock and Global Infrastructure Partners.

Saudi Arabia leans on cheap energy; Aramco argues low-cost power makes the Kingdom a natural AI hub. DataVolt’s $5 billion first phase of a planned 1.5-GW “net-zero AI factory” at NEOM’s Oxagon is billed as fully renewable. Yet the Kingdom runs some of the world’s largest desalination plants, still mostly fossil-powered — the water any build-out depends on. A net-zero facility on a grid and water system with a fossil footprint is not yet net-zero. The unit of accounting is the grid, not the data center.

Financing is split the same way. PIF has issued roughly $11 billion in green bonds over four years, and Middle East sustainable bond issuance is forecast at $15–20 billion in 2026. Abu Dhabi has positioned itself as a climate finance hub since COP28. Little of that money is earmarked for AI infrastructure.

A capex supercycle, and bubble risk

PwC and Oxford Economics forecast $31.6 trillion in cumulative data-center capex through 2050, rising from about $800 billion a year today to $1.8 trillion by mid-century. Chips and servers need refreshing every four to six years, so more than 90 percent of 2050 capex goes to hardware replacement. Hyperscalers are expected to issue over $1 trillion in new debt. A circular loop — developers raising capital to buy compute while hyperscalers invest in those developers to create demand — can conceal concentrated exposure. Cheaper Chinese open-source models undercut US pricing assumptions.

Data centers are a test case for adaptation, not just mitigation: they must stay resilient to heat, water scarcity and grid instability while decarbonizing. Treat each facility as an integrated energy hub, not isolated real estate, and a hyperscaler’s balance sheet can speed the wider grid’s clean transition. Public-private partnerships can de-risk transmission, grid connections and water infrastructure; green bonds, sustainability-linked debt, sukuk and infrastructure funds need earmarking for the power, cooling and water AI depends on. Environmental performance should be measured at the grid, not the facility.