August 4, 2026 · Tags: AI, data centers, infrastructure, energy, finance
The AI race gets talked about at the model layer. Which lab is ahead, which benchmark got beaten, whose chips are fastest. But underneath all of that is a physical buildout that has quietly become one of the largest infrastructure programs in history. Disclosed US data center project capex crossed $1 trillion this year, the forward pipeline sits at 331 gigawatts, and the combined 2026 capital spending plans of the five biggest hyperscalers come to roughly $732 billion. That is more than the annual GDP of most countries, spent in a single year by five companies.
What's striking is not the scale. It's that the three big blocs are building the same product in three completely different ways. Different construction, different money, different power.
United States: private capital, and a power problem money can't fix quickly #
The US builds big. Not just megawatt-big, gigawatt-big. Meta's Hyperion campus in Louisiana is 2 GW with room to grow past 5 GW. The Amazon-Anthropic site in Indiana passed 1 GW of operational capacity this spring. Stargate, the OpenAI-Oracle-SoftBank program, has seven sites planned for somewhere between 7 and 11 GW total, with about $400 billion committed.
The construction itself is surprisingly traditional. Stick-built concrete and steel still dominates at hyperscale, with 18 to 36 month timelines. Modular and prefab construction is growing fast but mostly for smaller and speed-critical builds, since it costs 20-30% more per MW. The real cost driver is the electrical system, which is 40-50% of total construction cost. AI-optimized buildings run $15-20 million per MW versus $8-12 million for standard enterprise space, and the binding constraint is no longer land or permits. It's hardware: large power transformers have 128-week lead times and 80% of US supply is imported. Labor is equally tight, with 82% of construction firms reporting craft worker shortages and peak crews of 4,000-5,000 people on the big campuses.
The funding is where the US is genuinely inventing something new. Hyperscaler balance sheets are the core, with Amazon at ~$220B, Google at $195-205B, and Microsoft at ~$175-190B in 2026 guidance. On top of that sits a new layer of special-purpose vehicles and private credit. The Meta-Blue Owl Hyperion deal was a $27B SPV where Blue Owl owns 80%, Meta owns 20%, and PIMCO anchored $27B of A+ rated debt. It's the largest private credit transaction on record. BlackRock is doing the same thing at an El Paso campus with Meta, and Gulf sovereign wealth funds are in as structural partners. Morgan Stanley estimates $800B of private credit is needed through 2028 to fund all of this. The first signs of underwriting discipline are already visible: Blue Owl passed on funding Oracle's Michigan campus.
Power is the constraint that drives everything. Grid interconnection queues run 3 to 6 years, so the US response has been to build power behind the meter. Stargate developers are constructing 6.8 GW of their own generation, GE Vernova's turbine order book is full through 2029, and there's a ~9.8 GW wave of corporate nuclear deals, from the Three Mile Island restart for Microsoft to Meta's agreements with Constellation, Vistra, TerraPower, and Oklo. The grid itself is gas-heavy (about 41% of generation) and the reliability regulator, NERC, issued a rare Level 3 alert in May after 60 Northern Virginia data centers dropped load simultaneously.
Europe: regulation-led, grid-constrained, publicly co-funded #
Europe's campuses are smaller and slower by design. Capacity runs about 16 GW today, heading to 36 GW by 2030, with €176B of cumulative investment expected over the next five years. High-voltage grid connections in Frankfurt and Dublin now take 4 to 7 years, Amsterdam has effectively shut new large builds, and 67% of operators say power access is their single biggest challenge. The growth is pushing north and south: Spain, Italy, and Portugal are surging on new subsea cable capacity, and the Nordics have become the destination for AI training campuses on cheap hydro and free ambient cooling.
What makes European construction distinctive is that efficiency is law. The Energy Efficiency Directive requires annual reporting for any data center over 500 kW, and Germany's EnEfG is the strictest law anywhere: new builds must hit PUE 1.2 within two years of commissioning, reuse 10-20% of waste heat, and run on 100% renewable power from 2027. About 90% of European data center energy already comes from renewables, and heat reuse is turning facilities into district heating sources. Modular construction is growing fastest there, in part because the regulatory target keeps moving and prefab is the cheapest way to chase it.
The funding mix is the most interesting of the three. REITs and colocation operators lead, pension money is arriving (CPP Investments and Equinix paid $4B for Nordic operator atNorth), and private credit is scaling up (AtlasEdge raised €1.2B, Pure DC €1.3B). But the genuinely new piece is public money. The EU is co-funding seven AI gigafactories with up to €2B each, capped at 35% public so state aid rules hold, which with private matching comes to roughly €30B of investment. SoftBank is separately committing up to €75B for 5 GW in France, with EDF supplying a retired power plant site. It's not a US-style buildout, but it's no longer a market that runs on private capital alone.
China: state planning, cheapest construction, compute moved to the power #
China's installed data center capacity is 32 GW and on track for 60+ GW by 2030, with 28 GW of new projects already announced. The build is fast and cheap: mid-spec construction costs run about $6.5M/MW, the lowest in Asia-Pacific (single-source estimate, treat with caution), driven by standardized designs, state-provided land, and a labor force the US doesn't have. Free cooling is designed in from the start, with campuses like Alibaba's Zhangbei base sitting in a county that averages 2.6°C, and liquid cooling penetration jumped from 5% to 20-25% in a single year because PUE rules demanded it.
The organizational innovation is East Data West Computing, the 2022 national program that treats compute location as an energy policy. Eight national computing hubs now hold about 70% of national capacity, with the big AI campuses going to Inner Mongolia, Gansu, and Guizhou where wind and solar are. Ulanqab alone has about 10 GW of announced projects from Huawei, ByteDance, and 21Vianet, and DeepSeek is building a 1 GW campus there. The state has directly invested $6.1B in the hubs so far, with $56-70B planned across the program, and China Development Bank is financing hub projects.
Funding layers private capex on top of that state base. Alibaba committed RMB 380B (about $53B) over three years, Tencent spent $10.7B in 2024 and is spending more in 2025, ByteDance budgeted $20B+ for 2025. The state layer is what differs from the US: policy banks, state telecom operators building much of the national footprint, and local governments competing with subsidies and tax breaks. It's the least transparent of the three blocs, and the only one where the state directly chooses where facilities go.
The power story is coal with a relocation strategy. Coal is about 60.5% of national generation and closer to 70% in the east where most data centers sit, which is why the plan moves compute west and hauls renewable power east over ultra-high-voltage lines running 1,500-1,900 km. The first green power direct supply project, at Ulaanqab, pairs 200 MW of wind, 100 MW of solar, and 45 MW of storage with a 25,000-cabinet facility. PUE caps of 1.25 for new large facilities and 1.2 inside the national hubs are among the strictest anywhere, though the 2022 Sichuan power rationing is a reminder that the grid can still bite when drought hits the hydro provinces.
What to watch #
Three open questions follow from all of this. Can US private credit absorb $800B of AI infrastructure debt without a serious correction? The first refusals are showing up. Can Europe's grid catch up to its own ambitions, when the EU's own report says data center power will not be able to triple as required? And can China square a coal-heavy data center fleet with its climate targets, when renewables and nuclear need to reach 60% of data center supply by 2035?
All three blocs are betting on the same thing: that the power demand is real and will keep growing. Every forecast from LBNL, Rystad, and the IEA has been revised up three years running. Whoever is right about the demand, the construction boom is already here, and it's being paid for in three very different currencies: American private credit, European public-private partnership, and Chinese state finance.
Read More #
- The Energy Geography of AI: How the US, Europe, and China Are Powering the Inference Surge
- The AI race isn't about who has the best model. It's about who keeps the lights on.
- The new data center numbers are in. Here's what a year of real data actually says.
Sources #
- JLL North America Data Center Report YE2025; Wood Mackenzie US data center pipeline Q1 2026
- Platformonomics and Axis Intelligence hyperscaler capex trackers (Q2 2026)
- Meta-Blue Owl Hyperion announcement (Oct 2025); Bloomberg/Business Times; Data Center Frontier (Blue Owl/QIA)
- EUDCA State of European Data Centres 2026; POLITICO (EU AI gigafactories); EUR-Lex (EED 2023/1791); Moduledge (Germany EnEfG)
- DCD (SoftBank France, Microsoft Portugal, Pure DC); IPE Real Assets (CPP-atNorth); AtlasEdge
- Rystad Energy (China 32-60 GW; EU 16-36 GW); Carbon Brief explainer on Chinese data center policy; gov.cn (green data center targets, EDWC investment); Reuters (Tencent, ByteDance capex); Alibaba Cloud (RMB 380B)
- ERCOT, LBNL/SETa 2025, NERC Level 3 alert; Constellation (TMI restart); CNBC (GE Vernova)
- Single-source flags: China mid-spec construction cost; some EDWC investment splits; Oracle lending pullback