The AI infrastructure trade has followed a predictable sequence — GPUs first, then HBM, then power, then networking. But a quieter shift is now underway, one that hasn’t yet captured the attention it deserves: enterprise SSD storage is emerging as a genuine bottleneck in AI inference workloads. Three analysts weigh in on what this means for investors.
Everyone is still talking about HBM. That’s the problem.
The bottleneck migration in AI infrastructure follows a consistent logic: wherever capital concentrates, supply eventually catches up, and the constraint moves downstream. We saw it with GPUs. We’re watching it play out in memory — SK Hynix’s pending U.S. listing targeting roughly ₩45 trillion in fresh capital tells you everything about where the money is flowing right now. But the question nobody is asking loudly enough is: what comes after memory tightness gets partially resolved?
The answer is storage. And the mechanism matters here. AI inference is not a compute-only problem. A deployed AI agent must continuously retrieve training data, inference logs, vector databases, and KV cache — all of which must flow from storage to GPU without interruption. Industry data suggests that roughly 30% of machine learning training time is consumed by data I/O operations, with some models spending as much as 65-70% of epoch time on input/output. This creates a structural paradox: as GPU performance improves, the gap between compute speed and storage throughput widens, not narrows. The faster the GPU, the more painful the storage bottleneck becomes. What we’re witnessing now is the AI bottleneck tree branching outward — from GPU to HBM to networking to CPU, and now into enterprise SSD. The money will follow. The investor implication is straightforward: storage names that serve hyperscalers are currently priced for a secondary role. That mispricing has a limited shelf life.
The bull thesis on enterprise SSD is real, but let me be precise about what’s actually happening before we assign valuation multiples to it.
Start with the demand signal, because the data is unambiguous. Dell’s most recent infrastructure solutions group revenue grew 181% year-over-year — and critically, that wasn’t just AI-optimized servers. Traditional server and storage revenue grew 92% in the same period. HPE confirmed a similar pattern: networking revenue up 148% year-over-year, with traditional storage seeing meaningful price increases that customers are accepting without pushback. That last detail is significant. When buyers absorb price increases under supply constraint, you’re not looking at cyclical demand — you’re looking at structural urgency. Now look at the NAND side. Sandisk’s most recent quarter showed gross margins approaching 67% with operating margins near 55% on guidance — numbers that, when plotted against a decade of quarterly data, look like a completely different company operating in a completely different industry. The HDD comparison is equally instructive: Seagate’s gross margins are now entering territory the company has never seen in its history, driven precisely by the inference-era shift from cold data storage toward hot and warm data access patterns that favor SSD.
The risks, however, deserve equal attention. First, the hyperscaler data center mix for SSD-focused players remains concentrated in a small number of customers — Sandisk’s datacenter revenue mix was only around 14.5% as of recent quarters, which means the margin story is being driven partly by consumer/client strength, not purely enterprise AI demand. Second, compression algorithms like Google’s TurboQuant, which aims to reduce AI memory consumption, are a genuine wildcard — if model weight compression accelerates, some of the derived storage demand could soften faster than the market expects. Third, valuation across the storage chain is no longer cheap. The easy money in identifying the bottleneck has likely already been made; execution on the actual supply ramp is what separates the trade from here.
Let me drill one layer below the consensus framing here, because I think both of my colleagues are partially right but missing the most important structural distinction.
The storage bottleneck story has two very different components, and conflating them is how investors get hurt. The first is capacity demand — the sheer volume of data that AI training, inference, and agentic workflows are generating. This is real, it’s quantifiable, and it’s driving the revenue lines we see at Dell, HPE, Seagate, and Sandisk. The second is performance demand — the requirement that storage systems match GPU throughput without creating idle cycles. These two demands do not necessarily favor the same companies or the same investment timeline. Capacity demand benefits HDD players like Seagate, whose gross margin trajectory has gone parabolic precisely because inference-era data centers are building out massive cold storage infrastructure alongside hot data SSD tiers. Performance demand favors enterprise NVMe SSD and increasingly specialized controller architectures — which is where a company like Padu (파두) becomes interesting again. Padu has been developing proprietary datacenter SSD controllers that sit at exactly the intersection of performance bottleneck and supply constraint. Jensen Huang’s reported comments about the company weren’t random flattery.
What concerns me about the current market framing is the tendency to treat “storage bottleneck” as a monolithic trade. The SemiAnalysis bottleneck migration sequence — from NVDA to power to logic and memory — has been remarkably accurate as a roadmap, but it doesn’t tell you which layer of the storage stack captures the economic value. My working hypothesis is that the controller and firmware layer — the intelligence inside the SSD — will prove more durable than the raw NAND commodity, for the same reason HBM outperformed standard DRAM: differentiation wins when performance specifications matter more than price. The uncertainty I’m holding openly is whether the inference demand acceleration is fast enough to prevent the storage trade from becoming crowded before the fundamental earnings inflection arrives. Dell’s 757% year-over-year growth in AI-optimized servers suggests the infrastructure buildout is still in early innings. But early innings and cheap entry points are not the same thing.
There is genuine convergence here beneath the surface disagreement. All three perspectives confirm that the storage bottleneck is structural, not cyclical — driven by the physics of inference workloads rather than a temporary demand spike. The debate is really about precision: which layer of the storage value chain captures durable economic rent, and at what point does the market’s recognition of the theme eliminate the investable edge? The macro framework says the money will move here regardless; the valuation discipline says it already has for the obvious names; and the structural analysis suggests the real opportunity may lie in the less-covered controller and software layer that nobody is yet pricing correctly. As with every prior AI bottleneck rotation, the window between “nobody is talking about this” and “everyone is crowded into this” has been compressing. Storage is no longer the former. Whether it has yet become the latter is the question worth answering before making a position decision.