China's AI computing power may nearly double by 2027, still built on US chips
A new report projects Chinese access to AI computing will roughly double, yet about 65 per cent will run on American-designed chips, showing a deep reliance.
The short version
- Chinese entities' AI compute: about 4.6 million to 8.8 million H100-equivalents in 2027
- About 65 per cent of that capacity is expected to use US-designed chips
- Remote access alone is projected at 2.8 million H100-equivalents, from roughly 1 million this year
- The US holds a roughly 16-fold lead in AI training compute
- Leading Chinese models trail American ones by only 6.5 months
Why it matters
The report suggests that chip limits alone have not stopped China from gaining AI computing power, and that Chinese AI still leans on American hardware. For readers, it explains why chip rules and cloud access keep showing up in AI news.
What the report projects
The Washington-based Institute for Progress (IFP) has published research on how much AI computing power Chinese entities can use. As South China Morning Post reported, the total is expected to grow from about 4.6 million to 8.8 million "H100-equivalents" in 2027.
That is close to a doubling in a single year. The catch is where the chips come from: about 65 per cent of the capacity is expected to run on hardware designed in the US.
What an H100-equivalent means
The H100 is a popular data-centre chip from Nvidia. Researchers use it as a yardstick, so a mix of different chips can be added up into one comparable number.
It works a bit like measuring engines in horsepower. It lets you compare very different setups without arguing over model names.
Renting chips abroad
Part of the growth comes from remote access, meaning Chinese users tap chips housed in data centres outside China. The report expects this to reach 2.8 million H100-equivalents in 2027, compared with roughly 1 million this year.
That is nearly triple. It shows computing power does not have to sit inside a country's borders to be used by its companies.
A big gap, a small lag
The report says the US keeps a roughly 16-fold advantage in AI training compute, the raw computing used to build models. Even so, the leading Chinese models trail their American rivals by only 6.5 months.
Put together, those two figures suggest that having far less computing power has not meant being far behind in results.
What to watch
- Whether remote access to chips outside China draws new rules or scrutiny
- Whether the 6.5-month lag between Chinese and American models grows or shrinks
- How reports of restricted chips turning up in China, like the B300 case, fit into these projections
Try it
Next time you see a computing-power figure in the news, check two things: whether it is counted in H100-equivalents, and whether the chips are owned or rented.
Sources and links
Liked it? Say why
Add a line for other readers.
No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.
Listen to the AI Wire headlines
Listen to the AI WireListen to our daily AI shows every morning: on your commute, in the shower, wherever you are
Listen to the daily showsFirst reported by South China Morning Post · Read the original. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.