Citigroup recently adjusted its outlook on Micron Technology, a major player in the memory semiconductor market. This adjustment has shifted the conversation around AI memory from simple scarcity to a more complex dynamic between technological efficiency and demand growth.
Citi lowered its target price for Micron's stock to $425 from a previous $510. This decision wasn't driven by a drop in Micron's immediate performance—in fact, their recent earnings were strong. Instead, the adjustment reflects new uncertainties on the horizon. Two key factors are at play here.
First is the trend in memory prices. After a strong February, the spot price for DDR5 memory, a key product, softened in March. While spot prices represent a small fraction of the market, they often act as a leading indicator for future contract prices, which made investors a bit more cautious.
Second, and more significant, was a technological development from Google called TurboQuant. This new technique can compress a part of an AI model's memory usage, called the KV cache, by 4 to 6 times with little loss in performance. This “efficiency shock” challenges the assumption that AI growth will require a simple, one-to-one increase in memory demand. If AI models can do more with less memory, the explosive demand growth could be tempered, at least in the short term.
So why didn't Citi change its near-term profit forecasts for Micron? The answer lies in long-term agreements (LTAs). Micron has secured multi-year contracts with major customers, locking in sales and prices. These agreements act as a buffer against short-term volatility in the spot market. Therefore, Citi's move was to reduce the valuation multiple—a measure of how much investors are willing to pay for future earnings—rather than cutting near-term earnings estimates. It's a way of saying, “The immediate future looks fine, but the long-term picture has become a bit less certain.”
- KV Cache: A type of temporary memory used by large language models (LLMs) to store information about the context of a conversation. It helps the AI respond faster and more coherently. Compressing it means using less memory for the same task.
- Spot Price: The price for immediate delivery of a commodity, like a memory chip. It reflects real-time supply and demand, unlike contract prices, which are set for future delivery.
- Long-Term Agreement (LTA): A multi-year contract between a supplier (like Micron) and a customer to supply goods at pre-negotiated terms. This provides stable revenue for the supplier and a secure supply for the customer.