SemiAnalysis's facility census puts China's AI-ready datacenter capacity at 24GW

The 24GW comes from a facility-level count of more than 1,000 sites run by more than 60 operators, according to a dev.to write-up of SemiAnalysis's new China Datacenter Model ^[2]. Published estimates of this figure had differed by as much as fifteenfold, and counting site by site is the right way to settle it ^[6].

In this census, a gigawatt measures power and floor space. The write-up says plainly that capacity is not compute. Output depends on the accelerators installed, and US export controls limit those ^[10]. It calls the expansion "chip-gated", with chips coming from domestic suppliers such as Huawei or through gray-market channels ^[10]. The same account says headline comparisons likely overstate China's effective compute today, because some facilities may be running below their intended accelerator density ^[11].

For the 24GW to mean 24GW of compute, the racks have to be filled. Until installed accelerator counts are published, we treat the figure as a ceiling on Chinese AI compute.

The gap figures need the same discount. The US leads by about 32GW today ^[15]. If every committed Chinese project is delivered, China reaches about 44GW ^[16], still below the current US figure ^[1]. To pass the US, China needs the 30GW of announced projects as well, for roughly 74GW ^[17]. The write-up does not include a US pipeline, so that 74GW is being compared with a US number dated to the end of 2026 ^[1].

The two countries are short of different inputs. In China, power is available. Grid companies finished the 14th Five-Year Plan with capital spending 24% over the original blueprint. The 15th plan, covering 2026 to 2030, adds another 40% and takes grid investment past $746 billion ^[12]. In the US, the write-up points to interconnection queues and to local opposition that has already blocked about $130 billion of planned AI datacenters ^[19].

ByteDance shows who fills the halls. It holds about a fifth of China's delivered capacity and rents nearly all of it ^[7]. Alibaba, Tencent and Baidu spent a combined $20 billion in the second quarter of 2026 alone ^[8]. ByteDance is also in talks for another 5 to 6GW in Ulanqab, an Inner Mongolia city the write-up says was once better known for potatoes and yogurt snacks ^[9].

The pricing evidence is local. Oversupply in some Chinese markets cut power-exclusive rates in half from about $80/kW/month ^[13], to roughly $40 ^[18]. The write-up expects pricing to take about 18 months to absorb the reset ^[13]. That is a rate for datacenter power in China. We think it is too early to build it into assumptions about global accelerator pricing. A kilowatt of Chinese datacenter power adds compute to the world's supply only once chips are installed against it ^[10].