Report: China's AI Computing Power to Double Next Year, 60% Reliant on U.S. Chips
nashnova research
A new report from Washington think tank IFP estimates China's accessible AI compute will nearly double to 8.8 million H100-equivalents by 2027, yet roughly 65% will still depend on US-designed chips — capacity is growing, but the structural reliance has not loosened.
Where does the doubling come from?
China's accessible AI compute is projected to grow from roughly 4.6 million H100-equivalents to 8.8 million, nearly doubling.
An H100-equivalent — a benchmark unit pegged to Nvidia's data-center GPU — is the yardstick IFP uses to compare chips of different types on one scale.
IFP's tally covers four sources: domestic chips, legal imports, smuggled processors, and compute accessed remotely via offshore cloud networks.
The fastest-growing slice — why is it "remote cloud"?
Compute accessed remotely through overseas cloud networks is set to jump from about 1 million H100-equivalents this year to 2.8 million by end-2027.
This means → nearly a third of the new capacity will not physically sit inside China; it is called in remotely from foreign data centers.
In plain terms = when you cannot buy the chips, you rent the compute — the cloud channel is becoming the main workaround for hardware shortages.
65% on US chips — has the dependency fallen?
IFP estimates that by 2027 roughly 65% of China's AI compute will still come from US-designed chips.
This means → even as domestic chips ramp up, the overall structural dependency has not meaningfully declined.
This reflects a growth pattern driven more by "sourcing US chips through multiple channels" than by "domestic substitution arriving on time."
Compute gap is 16×, but model gap is only six months — how?
The US maintains roughly a 16× lead in AI training compute, yet China's frontier models trail their US counterparts by only about 6.5 months.
In plain terms = hardware lags by an order of magnitude, but the models are nearly neck-and-neck — Chinese teams are closing the gap through algorithm and engineering efficiency.
This signals that compute is not the sole determinant of model capability; algorithmic optimization, data strategy, and engineering discipline matter just as much.
What happens if export controls tighten?
The report notes that any tightening of US export controls would directly cut into China's actually accessible AI compute.
This means → whether the remote-cloud channel stays open is the key variable for China's future AI capacity growth.
In plain terms = if Washington shuts the cloud pipeline — the "remote IV drip" — China's compute growth curve flattens immediately.
市场有风险，内容仅供研究参考，不构成投资建议。