China's Cheap AI Models May Be Micron's Bull Case, Not the Threat

Bank of America sees open Chinese AI as a memory demand story, not a Micron threat

Micron's setup is being read less as a geopolitics write-down and more as a memory-flow trade. The stock jumped over 7% in trading after BofA reiterated a Buy rating and a $1,550 price target. That target still implies roughly 66% upside from recent levels, which helps explain why investors are leaning into the thesis while the debate is still open.

The bear case is not gone. If cheap Chinese AI ends up suppressing memory pricing or keeping demand too centralized to liftMicron's mix, the stock could still stumble. BofA's opposing view is that open models multiply memory demand because downloads push inference onto customer hardware, where each deployment carries its own memory load.

Kimi K3 shows why low API pricing does not mean low memory demand

Kimi K3 carries 2.8 trillion parameters, but the key point for Micron is not model scale alone. BofA says it still needs about 1.4 terabytes of HBM per serving instance. In the bank's framing, open-weight models change where memory demand sits: instead of one provider serving many users from centralized hardware, each customer that deploys the model must provision its own stack.

Lower API pricing reflects business choices, not lighter hardware needs

The confusion comes from mixing API economics with hardware economics. Some Chinese models are priced up to 350 times cheaper than Western alternatives, but BofA argues that gap reflects business-model choices, lower energy and labor costs, and possibly state-linked capital subsidies rather than a proportionally cheaper memory requirement.

In that view, "open" refers to weight distribution, not deployment cost. Once the weights move from one central provider to many customer systems, the memory purchase follows the deployment footprint.

What would weaken the argument

The bear rebuttal is straightforward: if Chinese labs can serve models cheaper through efficiency, better architecture, or subsidized operating costs, then low API prices could win share without creating a matching jump in memory content.

BofA's counter is narrower but more testable. Models may cut compute intensity, yet memory requirements still rise as model weights and active parameters increase. So the bull case depends less on model prestige and more on whether open-weight adoption expands across enterprises, governments, and cloud environments.

What the market still has to price in

At 19.4x earnings, Micron still looks cheaper than the broader semiconductor sector at about 57x. That gap can support a rerating, but only if investors follow BofA's logic more broadly: cheap model access is not the same as light hardware demand.

One bridge point matters because it ties the argument together. BofA says Kimi K3 still needs about 1.4 terabytes of HBM per serving instance, and that open models multiply memory demand as customers deploy weights on their own hardware. If that framing gains traction, cheap Chinese AI starts to look less like a threat to Micron and more like a potential amplifier of memory demand.