Analysis
Google, Amazon and Microsoft are each trying to cut China out of their supply chains -- and all three still need Chinese suppliers to build the AI data centers they're racing to bring online, according to an October 5 report from The Information, corroborated by CNBC's earlier reporting on the same supply chain. The dependency runs through batteries, cooling systems, optical transceivers and substation transformers -- the unglamorous equipment that actually determines how fast a new data center can plug into the grid, not the GPUs everyone argues about.
The Hardware Nobody Argues About
Power transformers and substations are running an estimated 15% and 8% shortage respectively in 2026, with the squeeze worst in the 100 MVA-plus segment that large AI campuses need, per CNBC's earlier report on the shortage. US manufacturers simply don't have the capacity to fill that gap on hyperscaler timelines, so operators have been using Chinese-made units to avoid slipping their build schedules -- the same operators publicly pledging to reduce Chinese sourcing everywhere else in their stacks.
“Chinese suppliers win on two things diligence can actually verify: cost, and experience.”
Chinese suppliers win on two things diligence can actually verify: cost, and experience. China's domestic data-center market has matured faster than anywhere else, and its modular-construction techniques can cut build times to months instead of years. Companies like GDS Holdings and 21Vianet already operate interconnected data centers hosting AWS, Azure and other hyperscaler workloads inside China -- 21Vianet is Microsoft's exclusive operator for Azure and Microsoft 365 there -- giving Chinese infrastructure players a working relationship with US cloud providers that predates the current de-risking push.
The stakes go beyond any single supplier. Amazon, Google, Microsoft and Oracle have committed more than $160 billion to AI data centers across Asia, a region where Chinese suppliers already hold the cost and experience edge -- layered on top of Alphabet, Amazon, Meta and Microsoft's combined roughly $725 billion in total 2026 capex guidance, most of it going to GPUs, data centers and grid infrastructure. Microsoft alone is targeting 38 gigawatts of data-center capacity by 2032, a buildout pace that depends on sourcing hardware fast enough to keep up with its own chip orders, not just finding the hardware at all.
What the headline dependency misses: all three companies are making real progress on the parts of the stack that are easier to relocate. Microsoft is pushing toward 80% of its device and data-center component manufacturing outside China by as soon as 2026, Google is steering server production to Thailand, and Amazon Web Services has evaluated buying fewer printed circuit boards from its longtime Chinese supplier SYE. None of that solves the transformer and substation shortage, which is a capacity problem money alone can't fix on a multi-year grid-hardware manufacturing cycle -- but it does mean 'need' today isn't the same as 'need' in 2028.
For founders and GPs, this is the clearest evidence yet that the AI infrastructure bottleneck has moved down the stack -- from GPU allocation, which Nvidia and its partners have been steadily easing, to the physical-infrastructure layer underneath it. Chip startup funding is already running hot behind the AI buildout; the next capital wave is likely to follow power transformers, switchgear and grid-interconnect hardware -- categories almost nobody was pitching VCs on eighteen months ago.
Export controls or a new tariff round on Chinese grid hardware would hit data-center timelines faster than any GPU shortage has this cycle -- a risk none of the three companies' public de-risking plans currently account for.
