We have this thing where supply and demand create cycles, and they're wonderful cycles and they're terrible cycles all at the same time.
Those were the words of Solidigm’s Scott Shadley at VAST FWD as he, along with VAST’s Glenn Lockwood tracked the (not too distant) past, present, and future of the “dry year” for SSDs ahead.
The funny thing is that throughout other cycles in the IT industry, the expectation is that they will eventually self-correct. We’ve seen this plenty of times in the last few decades across various components where supply ramps up when demand is strong, then prices fall when there is too much capacity, forcing companies pull back investment, then the system tightens again when demand returns. Rinse and repeat.
That assumption has held through several big technology shifts. For example, moving from planar NAND to 3D NAND disrupted production, reduced supply for a time, forcing factories to retool but it still followed the same pattern (capacity dip, prices rise, new output comes online, supply catches up). And even major changes were treated as temporary disruptions inside a system that ultimately behaves in predictable ways.
But this predictable cycle only works if demand behaves in ways the system/industry has seen before. When that demand changes form instead of just scale, the cycle stops being predictable, leaving the industry with a lagging response to a problem that’s already moving faster than the system can adjust to.
And You May Ask Yourself…Well…How Did I Get Here?
And you thought pandemic hoarding narratives in IT were so 2022?
“If we all remember, there was a huge shortage for no real reason other than people got scared about things… they hoarded components, SSDs, NICs, servers, you name it,” Shadley reminded the crowd.
He pointed to how the pandemic created a demand signal that looked real at the time but was actually kind of misleading. Companies went on an infra buying spree as if failure rates were about to spike and resupply would be impossible. This meant pulling forward years of expected demand into a much shorter window.
And what followed was predictable, even if it was built on a shaky premise. When systems held up better than expected and the feared failures did not materialize, demand dropped bigtime, leaving the industry with excess inventory and reduced consumption. In response, vendors cut back, fabs slowed production, utilization dropped, investment tightened…the classic squeeze that’s the hallmark of similar reactions in a downcycles, but in this case they were based on a distorted view of demand rather than an actual decline in long-term need. Oops.
And what’s more, this happened at the same moment different kind of demand was about to emerge.
Couple All This to the AI Demand Shock
Storage demand follows directly from this new demand, of course. Those expensive GPUs only matter if the data behind them is there. Training and inference pipelines create multiple versions of data, and most of it gets kept. And as Lockwood explained to the audience, “both of those come with a floor on a minimum amount of data that is required to be stored to make those GPUs worthwhile.”
This created a different scaling problem because storage was no longer something you could size independently. It scales with GPU deployments, and those deployments were (and still are) accelerating. At the same time, data is staying active longer. Instead of being archived or deleted, it remains accessible because future models or workflows may extract new value from it.
You can see this in how quickly demand projections are still rising. Even in 2026, capacity estimates jumped significantly in just a few months but, as Shadley and Lockwood emphasize, the conversation now is less focused on demand and more on time. Because even if the industry reacts immediately, supply can’t expand fast enough to meet what AI is pulling through the system.
“A fab takes three years to get the building built, another three to four years to get the equipment in the building, and another couple of years to get the output at a reasonable level,” Shadley reiterates. That means the capacity available today reflects decisions made way before AI demand was clear.
He adds there are also limits inside the manufacturing process itself. NAND and DRAM aren’t easily interchangeable at the fab level, requiring different production lines. He says DRAM demand is rising at the same time because of GPUs and memory requirements across the system. What’s more, SSDs themselves depend on DRAM internally, which further ties storage supply to memory supply. The result is that multiple parts of the system are competing for resources at once.
So now, at the system level, this creates a lockstep effect. GPUs drive demand for DRAM and NAND together and you can’t deploy compute without storage, and you also can’t build storage without memory components that are also under pressure. That coupling makes the shortage harder to isolate or solve in one place.
And by the time new fabs begin to contribute meaningful output, demand has already moved. The cycle/system is just chasing a moving target with a delayed response.
There Are Some Answers: Inefficiency in How Data Is Stored
It stands to reason that if supply can’t expand fast enough, the only option is to use what you have better.
The problem is, most systems were built assuming storage would keep getting cheaper. That shows up in all kinds of places, including at a more granular level like how data is protected. Keeping multiple copies is simple and safe, but uses a lot of space. Even erasure coding, which is more efficient, still adds overhead that becomes harder to justify when NAND is the rare gem. As Lockwood points out, keeping three copies of data starts to look wasteful when capacity itself is limited.
There’s also a processing inefficiency. AI pipelines create multiple versions of the same data as it moves from raw input to usable form with each step spitting out a new dataset with earlier versions often kept. They aren’t identical, but they do share a lot of the same content and is pretty much duplicate.
Deleting data is one response, Lockwood says, but that nixes future flexibility. And data that seems unimportant now might be useful later, especially as models and tools improve. Once it is gone, you cannot get it back.
The takeaway? How you protect data and how much duplication exists inside pipelines directly determine how far your existing storage can stretch.
For data protection, Lockwood explains, the goal is to use less overhead without making recovery riskier. That depends on rebuilding data faster and more precisely so instead of, say, reconstructing everything after a failure, there are newer methods that rebuild only the missing pieces.
“There are mathematical ways that you can read only part of a stripe and reconstruct just the bits that you're missing without having to read all that data back,” he explains, which lets systems use less parity while still recovering quickly.
This is where VAST’s approach shows up in practice, turning efficiency into something the system enforces automatically rather than something users manage manually.
As for duplication, the focus shifts to finding similarity across data instead of just exact matches. AI pipelines create different versions of the same data at each step and while traditional methods treat them as separate, VAST can look across files and store only what is actually different.
Doing things this way changes how SSDs are used too since they are no longer generic components. Their behavior is shaped by how data is written and managed, which makes it possible to use higher-density, lower-cost media more effectively by controlling write patterns and buffering.
The Managed Shortage
Both Shadley and Lockwood agree we’re going to keep seeing vendors choosing how to distribute capacity instead of selling everything to the highest bidder, which shows up as allocations and longer lead times.
“There are vendors that are being the capitalist type, and we have chosen not to be that kind of person” Solidigm’s Shadley says. Orders that once took weeks now take months, and not every customer can get what they want when they want it.
He says too that demand also becomes uneven as large buyers try to secure extra supply to avoid future shortages, which adds even more pressure. Vendors have to decide how much to give to a few big customers versus spreading it across a wider base, which turns the shortage into both a supply problem and a distribution problem.
But, he adds, never fear. More capacity will come, new fabs will ramp and drive sizes will increase. But the system/cycle will not go back to how it worked before. The changes made under constraint remain.
The result is a system that has learned to operate with limits, even as those limits start to ease and VAST reflects a broader shift toward systems that use storage more efficiently rather than relying on supply to expand.
