There is a major shift happening in the artificial intelligence (AI) infrastructure market. While a significant chunk of the spending on AI hardware, such as chips and networking components, has been directed toward the training of large language models (LLMs) so far, inference workloads are now gaining significant traction.
The training phase requires a lot of computing power and huge datasets to ensure that the model is trained accurately and is fit for real-world usage. The inference phase, on the other hand, is putting the trained AI models to work in the real world by feeding LLMs new data. In simpler words, training models is like preparing for an exam, going over a lot of course material. Inference is akin to answering questions in an exam based on what one learned from the course material.
This explains why chip designers such as Broadcom and Intel have been witnessing healthy demand for their inference-focused AI processors. Broadcom is the leading designer of custom AI chips, collaborating with major hyperscalers, such as Google, to develop inference-focused processors. Intel, meanwhile, has received a big boost in the server central processing unit (CPU) and custom processor markets thanks to the growing tilt toward inference workloads.
However, both these chipmakers rely on memory chips to ensure that their data center accelerators perform to their full potential. That’s why I believe that the biggest winners of the AI inference era will be memory manufacturers — Micron Technology (MU 3.23%) and Sandisk (SNDK 6.06%).
Image source: Micron Technology.
AI inference workloads are going to boost memory demand
Consulting giant Deloitte estimates that inference workloads will account for two-thirds of the AI data center computing power this year, up from 50% in 2025. The shift toward inference will create demand for more compute and storage. As Western Digital CEO Irving Tan noted on the company’s April earnings call:
As AI workloads extend from training to large-scale inferencing, data generation is at an inflection point. This year, inference is expected to account for roughly two thirds of all AI compute. This larger focus on inference increases the amount of data generated, which in turn increases the need for data storage.
It is easy to see why Tan thinks so. While inference requires less computing power than the training phase, a massive increase in the number of inference requests can be expected as consumers and enterprises use AI applications to unlock productivity gains. Also, as inference happens on edge devices such as smartphones, personal computers (PCs), and cars, as well as in data centers, memory becomes a key factor in ensuring that trained models respond quickly to requests.
As a result, the inference phase requires faster data movement between the processor and storage. This explains why smartphones and PCs that can run AI workloads locally are packed with more compute and storage. Also, AI chip designers looking to deliver inference-focused chips are procuring and integrating more high-bandwidth memory (HBM) to reduce any potential bottlenecks during inference workloads.
Moreover, we have already seen that inference workloads will require more data storage. All this bodes well for Micron and Sandisk, which are in the business of manufacturing storage and compute memory.
Sandisk and Micron’s tearing stock market rally could be long-lasting
Micron stock has jumped by 639% in the past year, while Sandisk’s gains are way bigger at almost 3,400%. Micron manufactures both compute-oriented dynamic random-access memory (DRAM) and storage-focused NAND flash. Sandisk, meanwhile, is a pure-play flash storage manufacturer.
Today’s Change
(-3.23%) $-23.38
Current Price
$701.28
Key Data Points
Market Cap
$817B
Day’s Range
$697.22 – $756.73
52wk Range
$90.93 – $818.67
Volume
582K
Avg Vol
44M
Gross Margin
58.54%
Dividend Yield
0.07%
The AI-fueled demand for memory chips has created a severe supply shortage. Memory industry participants are warning that there may not be enough supply to meet demand, at least until next year. This is likely to lead to further price hikes in DRAM and NAND flash. Gartner is predicting a 2.9x increase in the memory industry’s revenue to $633.3 billion in 2026. That’s well above the 17% growth anticipated in the non-memory semiconductor market’s revenue this year.
Given that memory shortages could last until the end of the decade due to AI-fueled demand from data centers, PCs, and smartphones, the favorable environment that’s powering incredible growth at Sandisk and Micron should continue.
Data by YCharts
The impressive memory demand and supply constraints explain why these two companies are anticipated to deliver terrific earnings growth.
Data by YCharts
Both Sandisk and Micron should be able to sustain their eye-popping growth rates beyond next year as AI inference workloads drive greater memory demand. Even better, both semiconductor stocks are trading at attractive valuations despite jumping substantially over the past year.
Micron has a forward earnings multiple of 7.6, while Sandisk’s forward multiple stands at 24. The valuation suggests that both stocks are undervalued, given their stellar earnings growth. Both stocks deserve to trade at a significant premium to the tech-focused Nasdaq-100 index’s forward earnings multiple of 26.2, given their eye-popping earnings growth rates, which are significantly higher than the broader market’s.
The S&P 500 index companies could see a 12% increase in earnings per share in 2026, according to Goldman Sachs. Sandisk and Micron are on track to deliver way stronger growth than that, which is why buying these AI stocks is a no-brainer right now, given their attractive valuations and the lucrative opportunity in the AI inference space.
Leave a comment