US data-centre delays threaten to choke AI expansion
Source · Technology desk
— Summary
Nearly 40% of US data-centre projects scheduled to complete in 2026 are at risk of running more than three months late, according to satellite-analytics firm SynMax. The delays threaten the infrastructure behind AI rollouts by Microsoft, OpenAI, Oracle and others, raising concerns that the returns on hundreds of billions of dollars in announced capex (capital expenditure) will arrive later than expected.
Hyperscalers (the largest cloud operators running mega data centres) are racing to build sites drawing 1 gigawatt or more — roughly a nuclear reactor's output. Only a handful will complete this year, including campuses by Amazon Web Services, Meta and Elon Musk's xAI. A flagship 1.4GW, 1,200-acre Oracle campus in Shackelford County, Texas (which will equip OpenAI with chips and compute) has cleared land for six of 10 planned buildings but shows visible construction on only one. SynMax now sees the first building ready in December at the earliest, while comparable-project benchmarks point to late 2027. A 1.2GW OpenAI-linked site in Milam County, Texas has one building under construction. Only a project in Abilene is on track.
Bottlenecks: shortages of specialist workers (electricians, pipefitters), strained grid capacity, missing gas turbines and transformers, slow permits and rising local opposition. Remote sites are pushing labour costs up as much as 30%. SynMax estimates more than 60% of 2027 projects have not yet begun construction. Capstone's Josh Price describes a "regulatory lag" against the pace of developers. Source: Financial Times, 17 April 2026, Rafe Rosner-Uddin, Martha Muir, Nassos Stylianou and Aditi Bhandari.
The story in one line: Almost 40% of US data-centre projects due in 2026 are now at risk of missing completion dates by more than three months, constraining the AI rollout for Microsoft, OpenAI and peers.
Key numbers
Projects at risk of >3-month delay in 2026: ~40% (SynMax).
Share of 2027 projects not yet started: more than 60%.
Typical scale of top new sites: 1 GW or more — equivalent to a nuclear reactor.
Oracle-OpenAI Shackelford County (Texas) campus: 1.4 GW, 1,200 acres, 10 buildings planned; construction visible on 1.
Milam County (Texas) OpenAI campus: 1.2 GW; construction on 1 building.
Remote-site labour cost premium: up to +30%.
Why it matters
The AI arms race has shifted from model training to physical capacity, and the industry is discovering that turning billions of dollars of announced capex into working gigawatts takes years, not months. Electricians, transformers and grid capacity are now the binding constraints — not GPUs or dollars. A late-2027 slip on flagship OpenAI sites would mechanically delay revenue conversion, extending the gap between AI spending and AI monetisation that already worries investors.
Takeaway
Expect more project “blow-ups” through 2026, as Applied Digital’s Wes Cummins warned. For AI investors, data-centre operators with proven delivery track records, in-house construction teams, and secured power are now the scarce asset — more than those with the biggest announced pipelines.
Source: Financial Times, 17 April 2026, Rafe Rosner-Uddin, Martha Muir, Nassos Stylianou and Aditi Bhandari.