Goldman moves its China AI trade onshore, from Hong Kong platforms to A-share hardware

Goldman Sachs has changed how it expresses the China AI trade, steering investors toward mainland-listed hardware makers instead of Hong Kong-listed tech giants, according to a Bloomberg-sourced report ^[1][2]. The consequence for anyone running a China book is that the position moves from platform equity, where the AI story is a narrative about future monetisation, to component suppliers, where it is a shipment schedule.

The bank names three priority areas: the power supply chain, hardware infrastructure including printed circuit boards, optical modules and data centres, and a bucket it calls "physical AI applications" covering industrial intelligence and humanoid robots ^[3]. Goldman's stated catalyst is exports. Chinese chip exports rose 111% year over year in May 2026 ^[4], which means shipments slightly more than doubled against the prior-year month ^[5].

The load-bearing argument is not the export headline but what sits behind it. Goldman's case, as reported, is that these segments offer robust earnings visibility over the next couple of years because the companies have order books and supply contracts that make revenue relatively predictable, unlike software or broader internet names ^[6]. That is a claim about contract duration, not about end demand. It holds only as long as the orders are firm and the customers keep taking delivery.

The rotation is dated. The preference for mainland stocks over Hong Kong-listed equivalents is described as a tactical change as of June 2026 ^[7], and it sits inside a framework the bank labels "Go Global 3.0", which recasts China's next export cycle around AI-enabled products rather than the low-cost manufactured goods of earlier phases ^[8]. The accompanying index view is that MSCI China and the CSI 300 could deliver 12% to 20% upside by the end of 2026, attributed to AI momentum, export growth and supportive government policy ^[9]. Measured from the June 2026 date on the rotation, that is roughly six months of runway for a double-digit move ^[10].

Two gaps are worth naming. The source material gives a growth rate for chip exports but no absolute value and no base, so the durability of a 111% print cannot be assessed from it ^[4]. And a third of the projected index upside is attributed to policy ^[9], which is the least forecastable of the three drivers and the one that no order book protects against.

What to watch: whether monthly customs prints hold a triple-digit growth rate as the comparison base rises through the second half of 2026 ^[4]; whether the mainland-over-Hong-Kong preference survives contact with A-share liquidity and valuation, given it is framed as tactical rather than structural ^[7]; and whether the physical AI bucket, which includes humanoid robots ^[3], produces revenue on the same two-year visibility horizon Goldman applies to power and PCBs ^[6]. The power supply chain and optical module names are the cleanest test, because they ship into data centre builds that are already funded.