Micron Technology Inc "FQ3:26 (May) Earnings Preview: LTAs + Agentic AI as..."

研报英文原文证据摘录

Micron Technology Inc "FQ3:26 (May) Earnings Preview: LTAs + Agentic AI as..."

Valuation: PT $1,625 (unchanged)

Our price target of $1,625 stays unchanged. This is based on a ~15x NTM P/E multiple

(noting we see no reason why MU should trade a whole lot differently than other semi

companies in terms of P/E), which is in line with the 3-yr average, and applied to our ~

$124 C2029E EPS and discounted back to C2028E using a COE of ~12%. Importantly,

we anchor on C2029E EPS as we believe it best reflects MU’s through-cycle earnings

power under LTAs because, by that point, our model assumes a moderate memory

downcycle, albeit with LTAs buoying the company’s earnings base.

Insights from the MU Teach-In on Agentic AI

Below, we summarize our key takeaways from a recent investor call hosted with MU

management to discuss the implications of agentic AI on memory:

Agentic AI drives a step-function increase in memory intensity, as query

context grows ~30x/yr. MU flagged that, different from traditional “single-

shot” inference today, the workflow behind a single agentic prompt can spawn

~4-5 simultaneous agents, each executing separate CPU-based tasks while

continuously interfacing with GPU/ASIC accelerators. Consequently, this creates a

multiplicative increase in hardware utilization across CPU cycles, DRAM capacity,

GPU time, HBM bandwidth, and SSD storage - all of which imply materially higher

memory content per AI query versus prior (non-agentic) inference architectures.

HBM remains central as the critical enabler of agentic AI. MU emphasized

that active KV cache workloads require extremely high bandwidth memory sitting

directly adjacent to accelerators, with HBM serving as the “intelligence layer” of

the system - as how useful agents are effectively comes down to their