AI Data Centre Delays Put 35%-50% of Capacity Behind Schedule
Oxford research finds 35%-50% of announced AI data-centre capacity is delayed, pushing operators toward on-site gas, fuel cells and batteries while exposing commodity markets to overpricing.
By Nathan Brooks
4 min read

What's News
- 35%-50% of announced AI data centre capacity is behind schedule, Oxford research says.
- By 2028, off-grid generation could produce a substantial, uncaptured rise in natural-gas usage.
- Three exposed markets—copper, transformers and electrical steel—face short-term overpricing before possible oversupply.
- Marex funded the research and serves commodity producers, traders, banks, hedge funds and asset managers.
- The white paper has had at least one internal referee review and may be submitted to academic journals.
By Nathan Brooks
Between 35% and 50% of announced AI data centre capacity is running behind schedule, according to a white paper from the University of Oxford’s Smith School of Enterprise and the Environment. The range means more than one-third, and potentially one-half, of announced capacity faces delays; it measures capacity, not the share of individual projects delayed.
Grid interconnection queues and transmission upgrades are not keeping pace, the paper says. Operators are turning toward on-site gas turbines, fuel cells and battery systems to secure power.
What does the delivery gap change for investors?
The report, The Great AI Infrastructure Buildout: Impact on Power and Commodity Markets, was written by Dr Antif Ansar, Dr Stephen Lezak and Dr Guy Wolf. Ansar is programme director of the Oxford Programme on the Sustainable Future of Capital Intensive Industries. Wolf is global head of market analytics at Marex, which funded the research.
Ansar said: “Markets appear to be pricing in the announcement curve, not the delivery curve.” The paper says copper, transformer and electrical steel markets are reacting to “lofty buildout claims that may materialise years late or not at all.”
The paper warns of short-term overpricing followed by oversupply. Lezak said: “Media and markets are expecting a rapid, exponential ramp-up of Al infrastructure. The evidence tells a different story: a rollout that is delayed, lumpy, and constrained by physical realities.” It calls inflated or postponed announced capacity “bragawatts.”
For investors and suppliers, the practical risk is that prices peak on announcements before physical demand catches up. The paper’s focus makes the exposure measurable in copper, transformers and electrical steel rather than in abstract AI forecasts.
That distinction makes delivery data more important than headline capacity when assessing commodity and equipment exposure. Markets for copper, transformers and electrical steel can price an infrastructure boom before the underlying data centres operate, the paper says.
Grid constraints push generation onto site
The paper links the delivery gap to grid interconnection queues and transmission upgrades that are not keeping pace with AI capacity. It details a move toward gas turbines, fuel cells and battery systems for on-site generation.
By 2028, the authors say, off-grid generation for data centres could produce a “substantial uptick in natural gas usage not captured in official power sector statistics.” The change would raise local power costs and natural-gas consumption, the authors say, while complicating energy-market planning and climate accounting.
For operators, on-site generation offers an alternative to waiting for grid upgrades, though the paper does not quantify the cost trade-off. It also makes natural-gas demand harder to capture in official power-sector statistics, a problem the authors expect to intensify by 2028.
The economic case remains unproven
The authors conclude that the economic justification for the current scale of the buildout has not been demonstrated. Wolf warned that weak enterprise adoption could strand assets: “If enterprise AI adoption does not scale as expected, the result could be stranded assets or sharply lower utilization.”
The paper’s warning limits how far investors should extrapolate the current buildout. A smaller enterprise rollout would leave capacity underused even if each project reaches completion.
The authors framed the investment discipline this way: “The winners will be those who balance ambition with realism — investing in the future of AI without losing sight of the practical constraints of the present.” That conclusion puts delivery constraints, not ambition alone, at the center of the investment case.
Funding and publication status
The paper forms part of an ongoing programme on the sustainable future of capital-intensive industries funded by Marex. Marex, a diversified financial services firm, provides clearing, execution, market-making and hedging services across global commodity and financial markets.
The paper lists major commodity producers, consumers and traders, banks, hedge funds and asset managers among Marex’s clients, describing them as parties with direct exposure to the infrastructure and energy markets analysed. Marex’s status as both the funder and an active market intermediary gives readers material context when assessing the findings.
The white paper has undergone at least one internal referee review and may be submitted for publication in academic journals. The authors state that the views are theirs and do not necessarily represent the Oxford Smith School, any other institution or the funder.
Through 2028, grid connections and actual project delivery will determine whether AI infrastructure follows the announcement curve or the delivery curve. If enterprise adoption disappoints, the paper warns, even delivered assets could face sharply lower utilization or become stranded.
Source: epyhia.com
More from Nathan Brooks
Show full bio
News editor covering regulation, policy and their impact on entrepreneurship, startups and venture funding at Business Bearings.
2,611 articles