Washington has spent years working to limit America's dependence on Chinese AI chips, Chinese-made servers, and Chinese software in critical infrastructure. A Bloomberg investigation published this week found that the physical electrical infrastructure of the US data centre expansion is heavily dependent on Chinese manufacturing: transformers, switchgear, and the industrial electrical equipment that actually powers the facilities where AI runs. It is a supply chain exposure that has largely escaped the scrutiny applied to semiconductors, and one that sits awkwardly alongside the broader political effort to ring-fence the AI industry from Chinese influence.
The dependency is structural. American electrical equipment manufacturers do not have the production capacity to supply the volume of transformers and high-voltage switchgear that a data centre build-out at this scale requires. Lead times for domestically manufactured power transformers now run to two to three years in some cases. The data centres cannot wait. So they buy from wherever the equipment is available, and Chinese manufacturers have been expanding capacity into exactly this gap. The irony is precise: facilities being built to develop and deploy AI that the US government considers a strategic national asset are wired up with equipment from the country it is trying to strategically outcompete.
The risk is not trivial. Electrical infrastructure embedded in a data centre is not easy to swap out after the fact. A transformer or switchgear assembly that is part of the building is there for decades. Security researchers and intelligence officials have raised concerns about the possibility of remotely exploitable vulnerabilities in industrial control systems from Chinese manufacturers, a category of risk that received significant attention after the SolarWinds and Volt Typhoon incidents. Whether those concerns apply specifically to electrical equipment in data centres is a separate question, but the fact that the conversation is happening at all signals that the exposure is being taken seriously in some quarters, even if it has not yet surfaced prominently in public policy debate.
The energy financing picture adds a separate pressure to the infrastructure story. A Bloomberg report from the end of March found that Asian financial institutions are becoming cautious about the $800 billion data centre pipeline across the region, with energy price volatility and geopolitical risk emerging as core underwriting concerns. The concern is not that data centres are a bad investment; it is that the energy contracts that underpin them are increasingly uncertain, as grid operators in multiple countries face structural imbalances between baseload supply and the new peaks created by AI workloads. Lenders who were comfortable with the financial model twelve months ago are now building in larger risk buffers, which slows the financing cycle.
Neither story is catastrophic on its own. The US data centre build-out is continuing at pace and the Chinese equipment dependency, while real, has not yet prompted the kind of policy response that semiconductor export controls produced. The Asian financing caution is a sentiment shift, not a freeze. But together they suggest that the physical and financial infrastructure layer of the AI industry is entering a more complicated phase than the capital expenditure headlines imply. The software and model capability story has been the dominant frame for AI coverage. The hardware and infrastructure story is starting to demand equal attention.
There is a broader pattern here that is worth naming. Each layer of the AI stack turns out to have its own supply chain, geopolitical exposure, and systemic risk: chips depend on TSMC and ASML, energy depends on gas peakers and constrained grid capacity, and now the electrical infrastructure depends on manufacturers whose national affiliations are politically inconvenient. The industry has built fast and has not always looked carefully at what it is built on.