SanDisk was the best stock in the S&P 500 in 2025 with a 559% return, and the company is continuing this strong performance in 2026 with shares up 41% in barely two weeks. Much of this performance is rumored to be linked to Nvidia’s CES presentation where the GPU leader discussed context windows as the next bottleneck for AI inference, and hinted at rack-scale and network solid state drives (SSDs) becoming key components to address this.
On a broader level, data center/enterprise SSDs are often overlooked but equally critical as HBM when it comes to AI training and inference. This is because data center SSDs offer higher read-write speeds critical for accessing and transferring data rapidly, along with higher performance and energy efficiency, vital factors for larger-scale AI training and inference workloads.
SanDisk operates independently after being spun out of Western Digital in February 2025. The company expects to ride enterprise SSD demand tailwinds with management projecting sequential growth in its data center segment through 2026, with two hyperscaler qualifications underway and an additional hyperscaler expected in 2026.
However, data center remains a smaller portion of revenue, contributing $269 million last quarter or less than 12% of revenue, with client (PC/smartphones) and consumer products (SD cards/USB) remaining core to its business.
The Context Bottleneck, and Extending KV Cache to SSDs
Moving to some longer-term tailwinds for SSDs, Nvidia discussed how context windows could soon be the new bottleneck for AI inference performance at CES this week, as it unveiled its new Inference Context Memory Storage platform (ICMS) to address growing key-value (KV) cache capacity limits. KV cache capacity is a known pain point when working to balance long-context reasoning and memory capacity.
Put simply, the KV cache is a memory optimization technique that stores calculations during the inference phase, allowing the model to remember those prior calculations instead of repeating them, thus enabling faster response times; without it, latency would be much higher and computation much slower as every new token would require recalculation of all prior tokens. KV cache essentially serves as a model’s long-term memory that is reused and extended throughout many steps or requests.
However, the KV cache has a substantial memory footprint, especially for long contexts, and during deployment it can consume ~30% of GPU memory, making it a major bottleneck for large-context applications, such as coding, natural language processing, or handling simultaneous requests from many users on large models.
Let’s first take a look at KV cache from a tensor parallelism perspective, and how distributing memory across tens to thousands of GPUs significantly increases available KV cache memory, translating directly to larger maximum token context windows.
For example, a single AMD MI300X GPU running a Llama-70B model would have ~17GB of memory capacity available for the KV cache, after accounting for 140GB to store model parameters (2 bytes per parameter on FP16, so 70 billion * 2), and 35GB for the activation buffer (estimated ~25% of model storage), per TensorWave. With an estimated 2.6MB required per token (or 2.6GB per 1K tokens), a single GPU can handle a max request of ~6,500 tokens.
When you distribute model parameter and activation buffer across an 8-GPU server, or ~17.5GB and 4.4GB per GPU, this frees up 170.6GB per GPU for the KV cache, a ~10X increase; for the server, KV cache memory availability is now 1,360GB. This means that an 8-GPU server could now handle a max request of 523,000 tokens, an ~80X increase, with gains that only compound as server size and memory increase. This can then be optimized for longer contexts, or 8 requests of 64k context lengths, for high-throughput, or 64 requests of 8k context lengths, or other combinations.
Here’s where SSDs fit in – by extending or offloading KV cache to local SSDs, model prefill and time to first token can be significantly reduced, thus significantly decreasing latency and increasing throughput.
AI inference acceleration startup WEKA states that when it tested Llama-3.1 70B with no optimizations, a 100K token prompt took 24 seconds to prefill into the model before any output could be generated, but “extending GPU memory to ultra-fast storage [NVMe SSDs] can dramatically improve token processing efficiency.” When configuring an Nvidia DGX H100 server with an 8-node exabyte-scale NVMe SSD pod, WEKA says its “tests demonstrated a staggering 41x reduction in prefill time on LLaMA3.1-70B, dropping from 23.97 seconds to just 0.58 seconds,” significantly improving model efficiency with zero optimizations – simply from adding SSDs to extend GPU memory.
Google ran tests using an 8-GPU H100 server on Llama-3.3 70B, extending system memory to larger, lower-cost CPU RAM and enterprise SSD tiers. For a ~4 million token cache, Google found that utilizing CPU RAM and SSDs in conjunction with HBM decreased end-to-end latency and time to first token by 64% to 79%, while increasing throughput by 179-264% on 50k to 100k token prompts.
Nvidia Working to Tackle the Context Bottleneck with its ICMS Platform
With its new ICMS platform, Nvidia is working to mitigate context windows becoming a major bottleneck as agentic AI and physical AI scale over the coming years. Scaling of models to trillions of parameters and the shift to multi-step reasoning, multi-agent workflows or advanced multimodal applications will generate substantial volumes of context data and require significant KV cache reuse to maintain accuracy and context in prolonged interactions.
As such, Nvidia believes that “AI factories need a complementary, purpose‑built context layer that treats KV cache as its own AI‑native data class rather than forcing it into either scarce HBM or general‑purpose enterprise storage.” For example, the current inference context hierarchy begins with HBM (G1), providing near-instant access to latency-critical context in active generation, down to SSDs (G3) in the third tier to handle ‘warm’ data, or data that is used regularly but less frequently and still requiring efficient, cost-effective storage. Enterprise or shared storage sits at the bottom of the hierarchy (G4), handling ‘cold’ data, or data stored for long-term retention but much less frequently accessed.
Source: Nvidia
Nvidia is essentially proposing an architectural redesign of this hierarchy, positioning the new ICMS platform between G3 and G4, or as it calls it, G3.5. ICMS is a new Ethernet-attached NVMe SSD storage tier, likely integrated into the fabric and optimized specifically for KV cache usage at the pod level. It is powered by Nvidia’s new BlueField 4 data processing unit (DPU) packing 512GB of on-board SSD capacity, a 4x increase from BlueField3’s 128GB, and is combined with Nvidia’s GPUDirect Storage, which bypasses the CPU and provides direct memory access from GPU to SSDs to reduce latency.
ICMS will provide petabytes (millions of GB) of shared KV memory capacity per GPU pod, capable of storing context for many models or agents simultaneously, while being located close enough to GPUs to frequently share inference context with lower power consumption and better efficiency versus shared storage. Nvidia claims ICMS can enable up to 5x improvements in power efficiency and 5x increases in tokens per second versus shared storage.
It is this new platform and Nvidia CEO Jensen Huang’s comments relating to the storage hierarchy redesign that have fueled optimism for SanDisk and SSDs, as Huang believes it is a completely new market, integrating SSDs to the fabric, that could ultimately become the largest storage market:
“For storage, that is a completely unserved market today. The way that storage works is SQL. SQL is structured data. Structured database is lightweight. AI database KV caches insanely heavy weight. You're not going to hang that off of your north-south network. I mean that's just a horrible waste of network traffic. You want to put it right into the computing fabric, which is the reason why we introduced this new tier.
This is a market that never existed. And this market will likely be the largest storage market in the world, basically holding the working memory of the world's AIs. And that storage is going to be gigantic, and it needs to be super high performance.”
Because Nvidia is positioning NVMe SSDs to become the backbone for this new shared memory tier, there is the potential for SSD suppliers to see solid medium/long-term tailwinds from increased SSD capacity requirements in inference-optimized deployments over the next few years. For example, Bernstein estimates that Huang’s CES comments on SSDs and KV cache requirements suggest an additional 16TB per GPU, compared to 3-4TB per GPU today, or 4-5X growth. This will be more weighted towards year-end and into 2027 as ICMS rolls out with Rubin.
SanDisk’s BiCS8 Tech, Kioxia JV and Data Center ‘Stargate’ SSD Line
SanDisk is eyeing strong growth in the enterprise SSD market with its ‘Stargate’ NVMe SSD products, based on its BiCS8 architecture jointly-developed with Kioxia, offering industry-leading capacity, energy efficiency and performance.
NVMe (Non-Volatile Memory Express) is a protocol designed specifically for NAND-flash based SSDs that optimizes performance by reducing latency and increasing data transfer speeds by utilizing the PCIe bus, enabling high throughput and fast data transfer speeds necessary for AI training and inference.
SanDisk and Kioxia’s joint venture is one of the longest-standing JVs in the industry, signed in 2000 and lasting through 2034. It is primarily a shared manufacturing and capex strategy, with the two both splitting JV capex and wafer output and then selling NAND products independently. For example, the mega-fab in Yokkaichi, Japan produces nearly one-third of all global NAND bits (500K wafers per month), with the JV taking 80% capacity, split 50/50 between SanDisk and Kioxia, and Kioxia taking the remaining 20% (for an overall split of 60-40 for Kioxia and SanDisk).
The two also recently started operations at their new second fab in Kitakami, Japan, which is geared towards BiCS8 3D NAND and future advanced 3D NAND production, with output expected to ramp meaningfully in the first half of 2026. This will help scale SanDisk’s enterprise SSD line, based on BiCS8, and potentially aid in future development of high-bandwidth flash. BiCS8 accounted for 15% of bits shipped in fiscal Q1 and is expected to reach majority of bit production exiting FY26, providing a clue into the ramp profile for the year.
BiCS8 is the duo’s eighth-generation BiCS (bit cost scalable) 3D NAND architecture, which stacks NAND cells vertically, creating more layers and reducing costs per bit. BiCS8 scales to 218 layers from 162 layers in BiCS6, with SanDisk saying that BiCS8 increases memory density by more than 50%, program and read bandwidth by 35% and 26%, and data transfer speeds by more than 80% versus BiCS6. The two also have previewed the next generation of BiCS, scaling to 332 layers and further improving interface speeds by ~33%.
Additionally, SanDisk believes that its BiCS8 QLC (quad-level cell) die underpinning its Stargate data center SSDs delivers substantial performance, latency and efficiency advantages over competitors: 11% to 67% faster input/output speeds in Gb/s, along with 27% to 34% lower latency. Management expects its BiCS8 QLC line to go from ~20% to 40% of its data center business by the end of FY26.
Source: SanDisk
SanDisk’s ‘Stargate’ line, built on BiCS8, debuted this year with 64TB and 128TB capacities now shipping. 256TB products are scheduled for launch in mid to late 2026 and 512TB targeted in 2027, with the combination of fast performance, low latency and high capacity making the SSDs suitable for managing massive AI datasets and workloads. SanDisk says Stargate “is growing in demand with 2 hyperscaler qualifications underway and a third hyperscaler along with a major storage OEM planned for calendar year '26,” with current qualification focused on 128TB.
SanDisk also says that its SN861 NVMe SSDs were the first to be certified to support Nvidia’s GB200 NVL72, though this does not mean that its shipments will be correlated 1:1 with Nvidia’s racks. Management explained that the certification puts them on the approved vendor list, and “when the ODMs are picking their design, they pick vendors from the approved vendor list. So that's how we are getting into qualifications with the system partners that are building on top of Nvidia, not necessarily the Nvidia build.”
‘AI SSDs’ and a Path to ~33X Increase in Performance
Nvidia’s ICMS also could create a new market for SLC-based ‘AI SSDs’ in the data center, which had predominantly focused on TLC and QLC-based SSDs. SLC, or single-level cell, stores one bit of data per cell, offering the fastest data retrieval and best performance, though it is typically the most expensive; TLC and QLC (triple-level and quad-level) store three and four bits per cell, increasing storage capacity significantly and reducing cost but with slower performance versus SLC. Overall, SanDisk’s role in these first performance-based ‘AI SSDs’ remains somewhat unclear, though the company is playing a much more integral role in high-bandwidth flash (HBF) development.
The AI SSD push is currently being spearheaded by SK Hynix and Kioxia, and ultimately aims to boost SSD performance by up to 33X by 2027. SK Hynix is creating a three-product family, ‘AI-N’, with its AI-N P focusing on maximizing performance, AI-N B on maximizing bandwidth, and AI-N D on maximizing storage density. Reports suggest that Hynix’s first-gen AI-N P product will target approximately 25 million input/output operations per second (IOPS), or how many read/write operations a device can perform per second, with a higher number meaning faster performance. This is about an 8X leap from current SSDs at ~3 million IOPS today, while Hynix’s second-gen and Kioxia’s product in conjunction with Nvidia are said to be targeting 100 million IOPS by 2027, a 33X increase.
High-Bandwidth Flash Could Boost GPU Memory by 21X
HBF is being proposed as a future alternative/replacement for HBM memory on GPUs, stacking up to 16 3D NAND BiCS8 dies using through-silicon vias (TSV) to deliver up to a 21X capacity boost with similar bandwidth and cost as HBM.
For example, SanDisk’s first-gen HBF could pack capacity of 512GB per stack with HBM-like bandwidth. At partial HBM replacement, such as in six out of the eight dies on the GPU package, HBF could boost total GPU memory to ~3,120 GB per GPU, a nearly 17X increase versus individual Blackwell GPUs featuring 186GB HBM per GPU. In full replacement of HBM, HBF could provide 4,096 GB of total memory per GPU.
The significance of this is that it could allow frontier models to be stored entirely within a single GPU, rather than needed to be partitioned across a rack. SanDisk explains, “Think about a frontier large language model, let's say, something like GPT-4. GPT-4 has 1.8 trillion parameters with 16-bit rates. And that model alone would take 3.6 terabytes or 3,600 gigabytes of memory space. I can put that entire model on a single GPU now. I don't need to shuffle around data any longer.”
In terms of potential commercialization of HBF, Hynix’s aforementioned AI-N B line is built on HBF and expected to be developed in collaboration with SanDisk. The first ‘alpha’ version sample could be released as early as January with the first proof-of-concept samples in 2027, followed by full-scale evaluation afterwards.
SanDisk stated that the “gating item indeed is going to be enabling the ecosystem, aligning with the customers at their system level, integrate it and then bring it to the market,” but the product is highly executable as it is based on its existing NAND architecture. CEO David Goeckeler clarified in Q3 that the company “announced a time line last quarter of having the memory later in '26 and then having the controller for that in '27 we're still working towards that time line.”
However, timelines for HBF are still unclear given the newness of the technology, with some estimates suggesting 2027 to 2028 as a possibility. It is far too early to tell whether HBF will be commercially viable or successful.
Training and Inference are Long-Term SSD Demand Drivers
This section includes a brief excerpt from our 6,000+ word thematic deep dive into the current AI memory boom recently published for our Pro subscribers:
AI training and inference are two main long-term drivers for SSD demand, which is projected to rise ~6X from 2024 to 2030, from 181 exabytes (EB, or equal to 181,000,000 TB) to 1,078 EB, under McKinsey’s base case scenario. Training demand projected to rise at a 62% CAGR to from 7 EB in 2024 to 127 EB by 2030. On the flipside, demand from AI inference is expected to grow at a 105% CAGR from 6 EB to 447 EB by 2030, giving inference a 41% share of demand versus less than 12% for training. base case scenario. Training demand projected to rise at a 62% CAGR to from 7 EB in 2024 to 127 EB by 2030. On the flipside, demand from AI inference is expected to grow at a 105% CAGR from 6 EB to 447 EB by 2030, giving inference a 41% share of demand versus less than 12% for training.
This is not only driven by development of more LLMs, but also the increasing size and complexity of frontier models, where training data sets and context windows for inference are getting increasingly large.
For example, EpochAI estimates that training data set sizes are rising 3.7X per year on average, or nearly doubling every six months, though there are some models that are scaling much quicker. For example, Meta’s Llama2-70B from 2023 was trained on 2 trillion tokens, while Llama3-70B, from 2024, was trained on 15 trillion tokens, a 7.5X increase. Multi-modal models, those integrating audio, video, image or more, are also likely to require significantly more SSD storage, with McKinsey estimating in the hundreds of TBs depending on the mix of data needing to be stored. estimates that training data set sizes are rising 3.7X per year on average, or nearly doubling every six months, though there are some models that are scaling much quicker. For example, Meta’s Llama2-70B from 2023 was trained on 2 trillion tokens, while Llama3-70B, from 2024, was trained on 15 trillion tokens, a 7.5X increase. Multi-modal models, those integrating audio, video, image or more, are also likely to require significantly more SSD storage, with McKinsey estimating in the hundreds of TBs depending on the mix of data needing to be stored.
Source: EpochAI
The increasing size and complexity of models also ties directly to a major pain point when it comes to inference: “As models grow in complexity and require longer contexts, their memory footprint expands beyond what a single GPU can handle. This results in inefficiencies where GPUs are memory-starved, causing significant bottlenecks in AI token generation.” inference: “As models grow in complexity and require longer contexts, their memory footprint expands beyond what a single GPU can handle. This results in inefficiencies where GPUs are memory-starved, causing significant bottlenecks in AI token generation.”
This is exactly what Nvidia is addressing with Rubin and ICMSP, creating a new storage tier within the cluster fabric that is designed to extend GPU memory and facilitate high-speed KV cache distribution among racks.
There are also tailwinds to SSD growth from increasing cluster sizes, with compute-focused eSSDs seeing a 1:1 attach rate per GPU. For example, SanDisk says that a real-world 32,256 GPU cluster (or eight pods of 252 16-GPU racks) would require 4,032 compute eSSDs such as its SN861 product. This could create a strong tailwind for SSD growth as clusters scale to 100K+ GPUs towards 1 million, assuming the correlation for compute eSSDs to GPU remains 1:1.
Financials
Revenue
SanDisk reported a strong sequential revenue acceleration in its fiscal Q1, driven by NAND demand outpacing supply and increasing demand in its data center, edge and consumer end markets. Q1 revenue increased 22.6% YoY and 21.4% QoQ to $2.31 billion, accelerating from 8% YoY and 12.2% QoQ growth in fiscal Q4. Higher-than-expected bit growth drove the outperformance in the quarter relative to guidance of $2.1-2.2 billion, per management.
SanDisk’s Edge segment was the primary growth driver in Q1 with revenue up 30% YoY and 26% QoQ to $1.39 billion, driven by increasing NAND content in PCs and smartphones and a positive PC refresh cycle. Consumer revenue rose 27% YoY and 11% QoQ to $652 million, while data center revenue was down (10%) YoY but up 26% QoQ to $269 million.
Q2 revenue was guided to be $2.55 to $2.65 billion, up 38.6% YoY and 12.6% QoQ at midpoint. CFO Luis Visoso clarified that “the key message is most of the growth in revenue will be pricing driven in the quarter.”
Revenue growth is then expected to accelerate further to 55% YoY in fiscal Q3 (even with a seasonal slowdown in consumer products following the holidays) and then decelerate slightly to 51% in Q4. Pricing tailwinds could strengthen significantly in FQ3 on reports that NAND prices for enterprise SSDs could rise ~100% QoQ in the March quarter, according to supply chain checks by Nomura. Citi estimates SSD prices will rise ~32% QoQ in the March quarter following a 21% QoQ increase in the December quarter.
For fiscal 2026, SanDisk is currently expected to generate revenue of $10.6 billion, up 44.1% YoY. SanDisk sees demand outpacing supply through the entire year, currently estimating supply to support mid-teens demand growth, and potentially lead to strong pricing tailwinds from this tight/tightening environment:
“We saw supply growth in calendar year '25 of about 8%. We see it at about 17% in '26. We see demand — constrained demand around 14% [mid-teens] because that's all that's out there from a supply point of view. But unconstrained demand is in the — literally, a couple of weeks ago, we thought it was 20%, it's probably mid-20s by now. So we see the supply pretty much being able to service that kind of mid-teens level demand for '26.”
This tightening environment comes despite fabs running at 100% utilization, with management adding that they do not plan on adding capacity to any end market, but rather remain prepared with the optionality to shift capacity as visibility into product mix strengthens.
AI Segment Growth
SanDisk’s data center revenue, as mentioned above, declined (10%) YoY but rose 26% QoQ to $269 million, driven by increasing demand for its ‘Stargate’ enterprise SSD product line. However, revenue contribution remains small, at less than 12% of revenue.
SanDisk did not provide a numerical guide for Q2 for data center, but management noted that they are expecting sequential growth throughout fiscal 2026 with faster growth in the back half, driven by the current hyperscaler qualifications planned and underway. SanDisk did clarify that they are “working with 5 major hyperscale customers through active sales and strategic engagements” across its data center portfolio.
Data center growth is supported by solid visibility, with management explaining that they are either “striking deals that are multi-quarters, let's say, through the first half of next calendar year” from customers looking to lock in supply, or working with customers with demand visibility through 2027 to align supply with those demand forecasts. Management also sees undersupply conditions extending potentially into 2027 now, supporting strong pricing in deal negotiations.
Management also increased their forecast for data center exabyte growth, explaining that last quarter, exabyte growth expectations were in the mid-20% range, but now are in the mid-40% range. As a result, data center is expected to be the largest market in NAND on an exabyte basis in 2026, surpassing mobile.
Earnings
SanDisk stands out for its strong expected earnings growth through fiscal 2026 and fiscal 2027, with adjusted EPS expected to reach more than $21 by then, or >7X higher than the $2.99 it earned in fiscal 2025.
Q1 GAAP EPS was $0.75, a strong improvement from a ($0.16) loss in Q4, though this was down (49%) YoY from $1.46 in the year ago quarter as margins remained lower YoY. Adjusted EPS was $1.22, up 321% QoQ but down (33%) YoY.
For Q2, SanDisk guided for adjusted EPS of $3.00 to $3.40, up more than 162% QoQ. Adjusted EPS is expected to further increase to $3.78 in fiscal Q3 and $4.82 in fiscal Q4.
For fiscal 2026, SanDisk is expected to generate $13.29 in adjusted EPS, up 344.6% YoY, while GAAP EPS is projected to be $11.53, up from ($11.32) in FY25 due to the spin off. Fiscal 2027 is expected to see earnings power surpass $21, with GAAP EPS estimated to be up 86% to $21.47 and adjusted EPS up nearly 62% to $21.50.
Margins
Margins are lower YoY compared to pre-spinoff margins, but Q1 saw strong sequential margin expansion that is expected to accelerate in Q2.
- Q1 GAAP gross margin was 29.8%, down 8.8 points YoY but up 3.6 points QoQ. Adjusted gross margin was 29.9%, down 9 points YoY but up 3.5 points QoQ.
- GAAP operating margin was 8.3%, down 8.3 points YoY but up 5.6 points QoQ. Adjusted operating margin was 10.6%, down 8.2 points YoY but up 5.3 points QoQ.
- GAAP net margin was 4.9%, down 6.3 points YoY but up 2.7 points QoQ, and adjusted net margin was 7.8%.
For Q2, SanDisk guided adjusted gross margin to be 41-43%, or up just over 12 points QoQ at midpoint on higher pricing and cost reduction tailwinds, while adjusted operating margin is implied to be 24.2% at the midpoint of opex guidance, or up 13.6 points QoQ. Fab startup costs are expected to transition from headwinds to tailwinds during the quarter, potentially aiding more margin expansion into fiscal Q3 and Q4.
Cash
SanDisk noted that in Q1 it reached a net cash position, six months ahead of schedule, though debt is still almost equivalent to its cash on hand. Cash flows were quite strong, and adjusted FCF margin showed strong expansion.
- Operating cash flow was $488 million in Q1 for a 21.1% margin, up from a (7%) margin in the year ago quarter and a 4.9% margin in Q4.
- Adjusted free cash flow was $438 million in Q1 for a 19% margin, up from a (10.5%) margin in the year ago quarter and 2.6% in Q4.
- SanDisk’s total gross capex to support the JV was $387 million in Q1, though its cash capex spend was only $40 million (1.7% of revenue) as the remainder was funded through external sources such as subsidies or tool depreciation recorded in COGS.
Cash and equivalents totaled $1.44 billion while debt totaled $1.35 billion.
Valuation
SanDisk’s valuation is somewhat hard to pin down given the company’s limited history on the public markets after its February spinoff, and its rapid 362% ascent since the end of August.
SanDisk trades at 5.3x forward PS, surpassing its prior peak at 4x in November and a substantial re-rating higher from 0.6x in the summer. For comparison, this is now on par with former parent Western Digital at 5.1x forward PS, though the two are focused on different memory market segments with WDC primarily in hard disk drives.
For forward PE, SanDisk currently trades at an 28.7x multiple, nearly double its 15.8x average from the second half of fiscal 2025 prior to its fiscal year readjustment in June. Shares traded as low as 3x in July and August due to the sharp earnings increase expected in fiscal 2026.
Notable Risks
The NAND flash market has historically been quite volatile, and is shifting from significant oversupply in 2023 to expectations for substantial supply shortages through 2026 and potentially into 2027. However, if NAND capacity begins to come online quickly through next year, or if demand for PCs and smartphones falters due to rising memory prices, the NAND cycle could reverse and lead to pricing pressures cutting into revenue growth and margins.
Competition is also quite stiff in enterprise SSDs, and SanDisk is a small player with <4% market share, versus Samsung at >35%, SK Hynix/Solidigm at nearly 27%, and Micron and Kioxia in the 14% range. Jefferies analysts also warned that there is “no idea” what market share China’s YMTC could take as it ramps up output.
There’s also the risk of having a limited viewpoint on where normalized earnings could land if/when the cycle peaks and reverses, as SanDisk is currently benefiting from strong pricing and a tightening supply-demand environment. Combined with the sharp 1,000%+ rally since its summer lows and peak valuation multiples, there could be a higher degree of risk if the supply-demand imbalance and pricing revert sooner than expected.
Conclusion
SanDisk has a multi-faceted growth opportunity ahead over the next few quarters, with supply-demand imbalances widening with strong enterprise SSD demand, a potential doubling of prices in the March quarter supporting more upside for revenue, and multiple hyperscaler qualifications on deck.
Nvidia’s CES keynote discussion around the context window becoming the next bottleneck could have positive implications for the SSD market from Nvidia’s ICMS platform utilizing NVMe SSDs to significantly boost KV cache memory and increase throughput for inference applications. HBF is also a potential long-term opportunity later in the decade as it could dramatically boost total GPU memory to allow frontier models to run on a single GPU, though it is too early to tell if it will be viable.
Damien Robbins, Equity Analyst at I/O Fund contributed to this analysis.
Please note: The I/O Fund conducts research and draws conclusions for the company’s portfolio. We then share that information with our readers and offer real-time trade notifications. This is not a guarantee of a stock’s performance and it is not financial advice. Please consult your personal financial advisor before buying any stock in the companies mentioned in this analysis. Beth Kindig and the I/O Fund do not own shares in SNDK at the time of writing and may own stocks pictured in the charts.
Recommended Reading: