IFP: China's AI compute to nearly double by end of 2027

China's available AI computing resources are projected to nearly double from an H100-equivalent of 4.6 million at the end of this year to 8.8 million at the end of next year, according to a report published by the US think tank Institute for Progress (IFP), as reported by aitimes.com.

The report estimates that about 65 percent of that AI computing capacity will rely on chips designed by US companies, even as Beijing pushes domestic semiconductor production.

Computing power obtained by renting overseas data-center GPUs or space is projected to grow from a 1 million H100-equivalent at the end of this year to 2.8 million at the end of next year. Cumulative capacity from Chinese-made chips is projected to rise from 1.7 million H100-equivalent to 3 million over the same period, about 35 percent of the total.

Report urges tighter overseas controls

The IFP recommends that the US government tighten controls on the use of overseas training clusters, in addition to existing export and smuggling restrictions. The report notes its figures are estimates, since overseas rental contracts and per-customer usage are often not disclosed.

In September, AI research institute Epoch AI said its analysis of Huawei's semiconductor supply chain identified a shortage of high-bandwidth memory as the key constraint on expanding production. Epoch AI estimated that the total computing capacity of chips Huawei produces this year would be less than 4 percent of Nvidia's, and could stay around 1 percent of Nvidia's in 2028 if only Chinese-made HBM is used.