AI demand is reshaping data‑centre economics
Rising AI capex is tightening data‑centre capacity and could become a macro cost driver as hyperscalers bid up space, power and specialised plant, prompting calls to zone workloads by latency sensitivity. Commentators recommend keeping venue‑adjacent, exchange‑critical systems in tightly controlled on‑prem/colo environments while moving elastic research and training workloads to cloud or AI‑optimised facilities. (benzinga.com) (fool.com) (markets.financialcontent.com)
Artificial intelligence demand is turning data centres into a bidding war for power, land and electrical gear, with capacity tightening fastest in the facilities built for advanced chips. (jll.com) JLL said in January that global data-centre capacity could nearly double from 103 gigawatts to 200 gigawatts by 2030, with artificial intelligence workloads accounting for half of capacity by then. Goldman Sachs said occupancy could rise above 95% in late 2026 before easing as more sites open. (jll.com) (goldmansachs.com) The squeeze is not just about buildings. Goldman Sachs said United States spending on data-centre construction has tripled in three years, while Bloomberg reported on April 1 that transformers and other electrical equipment are delaying projects. (goldmansachs.com) (bloomberg.com) A data centre is a warehouse for computing, and artificial intelligence training packs far more chips and cooling equipment into each room than a standard cloud setup. That shifts the scarce input from empty floor space to megawatts, substations and the pipes, chillers and backup systems needed to keep dense servers running. (goldmansachs.com) (jll.com) That is changing where companies place workloads. Latency-sensitive systems such as market connectivity, payments processing and other exchange-adjacent tasks still need tightly controlled on-premises or colocation sites near users, while research, model training and bursty inference can be pushed to cloud or specialist artificial intelligence facilities farther away. (benzinga.com) (goldmansachs.com) The money moving into that second bucket is getting larger and longer-dated. CoreWeave said on April 9 that Meta expanded an artificial intelligence infrastructure agreement to about $21 billion through December 2032, and on April 10 CoreWeave announced a multi-year deal with Anthropic to support Claude models. (coreweave.com 1) (coreweave.com 2) Those contracts are helping create a separate market for “neocloud” providers that rent out graphics processing unit-heavy capacity without owning the full consumer cloud stack. CoreWeave disclosed the Meta agreement in a Form 8-K filed with the Securities and Exchange Commission on April 9. (sec.gov) (coreweave.com) The cost pressure is spreading beyond technology budgets. Benzinga reported Wednesday that computer software and accessories rose 4% month over month in the March Consumer Price Index data it cited, while retail electricity prices were running about 5% to 6% higher from a year earlier. (benzinga.com) Utilities are responding with larger spending plans. CBS News reported Tuesday that PowerLines found investor-owned utilities plan to spend $1.4 trillion over five years, with data centres cited as a top capital driver by a majority of the 51 companies reviewed. (cbsnews.com) States are also pushing back on where this buildout lands. Benzinga reported that New York, Maine, Oklahoma and Georgia have all moved in 2026 to restrict or discourage large-scale artificial intelligence data-centre development as local officials weigh power demand, land use and ratepayer costs. (benzinga.com) The near-term result is a more segmented map: keep the systems that cannot tolerate delay close to the exchange, the customer or the factory floor, and send the flexible artificial intelligence jobs to the campuses that can secure the power. (benzinga.com) (jll.com)
