Goldman Sachs (GS) one-delta desk has a direct read on China's artificial intelligence market: demand is significant. The complication is that the firm's preferred expression of that demand is hardware investment, a stance that says something specific about where Goldman thinks the value in China's AI buildout actually sits.
The case for hardware starts with the mechanism the one-delta desk identifies. Cheaper inference is pulling more users into AI applications, and in Goldman's framing, that volume dynamic makes hardware the more durable position. If lower inference costs drive adoption, the argument goes, the infrastructure layer absorbs the growth in a way that inference providers may not.
That is a meaningful distinction. Demand readings that point to significant consumption are often taken as a bullish signal across the AI complex. Goldman's desk is making a narrower call: the economic upside from China's AI adoption accumulates at the hardware layer, not in the inference services that are now becoming cheaper to run.
The read-through is pointed. Significant demand in China does not automatically translate into a uniform trade across the AI stack. Goldman's positioning preference suggests the firm is separating the demand signal from the question of who captures the economics.
The counterargument deserves its own hearing. If cheaper inference is the engine of adoption, hardware suppliers are exposed to the same deflationary pressure over time. Goldman's own framing, that lower inference prices are accelerating usage, is simultaneously an argument that cost curves are moving fast. Cost compression in AI has a way of traveling up the stack, and a conviction hardware call assumes the margin story there holds even as the inference layer cheapens. That is not a guaranteed outcome.
On balance, the GS one-delta desk is making a structural argument. Significant demand in China is the signal; hardware exposure is the firm's preferred vehicle for it. The line to watch is whether cheaper inference continues to pull users in at a rate that sustains hardware volume, or whether price compression eventually narrows the infrastructure advantage the trade depends on.