Big Tech $700B capex pipeline stalls
- U.S. AI data-center construction has started to slow, not because demand faded, but because power, permits, and electrical gear cannot arrive fast enough. (cbre.com) - The clearest number is this: only about 5 GW of the 16 GW slated for 2026 is under construction, and 30%–50% may miss year-end. (sightlineclimate.com) - That matters because Big Tech still plans to spend more than $650 billion this year, so the bottleneck has shifted from money and chips to the grid. (bloomberg.com)
AI data centers are running into a very old-economy problem. Not software. Not chips. Wires, transformers, permits, substations, and enough grid capacity to feed buil(cbre.com) billion capex pipeline stalls” line — the money is still there, but the physical system underneath it is starting to seize up. (b([sightlineclimate.com)ows-due-to-power-grid-limits-wood-mackenzie-says)) ### What actually slowed? The U.S. buildout itself. CBRE says capacity under construction in primary(bloomberg.com) happened even as vacancy hit a record-low 1.4% and leasing stayed extremely strong. So this is not a demand problem. It is a delivery problem. (cbre.com) ### Where does the “half the pipeline” idea come from? From project-level tracking. Sightline Climate says at least 16 GW of capacity is supposed to come onlin(bloomberg.com)herwise not visibly moving. Their base read is that 30%–50% of the 2026 pipeline will not come online before year-end. Bloomberg framed the same point even more bluntly — almost half of planned U.S. data centers this year are expected to be delayed or canceled. (sightlineclimate.com) ### Why is power the hard part? (cbre.com)zie says developers added projects representing about 25 GW of electricity demand in the fourth quarter of 2025 — only about half the pace of the prior quarter. That slowdown lines up with a system hitting physical limits. In plain English, companies can decide to spend billions in one board meeting, but utilities cannot conjure new transmission, substations, and interconnection approvals on startup timelines. (bloomberg.com) ### Why(sightlineclimate.com) lets all the glamorous stuff turn on. Bloomberg’s April 1 report says shortages of transformers, switchgear, and batteries are now delaying projects directly. These parts are needed both inside data centers and across the wider grid. That means AI builders are competing with utilities, EV charging, heat-pump electrification, and other industrial demand for the same equipment. (bloomberg.com) ### So is Big Tech (bloomberg.com) and Microsoft had committed to spending more than $650 billion this year alone. The catch is that capex no longer maps cleanly to near-term capacity delivery. Money can secure land, GPUs, and contracts — but not instantly accelerate a constrained electrical supply chain. (bloomberg.com) ### What do companies do instead? They move closer to power. Sigh(bloomberg.com)nced capacity. That is why giant campuses with dedicated generation — like the Abilene, Texas site Bloomberg highlighted — matter so much. Hyperscalers are starting to treat power procurement as part of the product, not just a utility bill. (sightlineclimate.com) ### Does this kill the AI buildout? No — but it changes the shape of it. The near-term winners are the d(bloomberg.com)longer constrained only by chips from Nvidia or budgets from Big Tech. It is constrained by whether the electrical system can catch up. (cbre.com) ### Bottom line? The stall is real, but it is not a collapse in appetite. It is a collision between software-speed investment plans and infrastructure-speed reality. Big Tech can still announce a $700 billion wave. The grid gets a vote. (sightlineclimate.com)
