SanDisk CEO: KV Cache Migration to NAND Is Already Happening, and Data Center Demand Forecasts Keep Getting Revised Up
Bernstein 42nd Annual Strategic Decisions Conference, May 28, 2026 — SanDisk Chairman and CEO David Goeckeler lays out the demand inflection, the $11 billion in contract commitments, and the coming HBF architecture play
Data Center Demand Forecasts Have Been Revised Up Four Times Running
The most striking disclosure from SanDisk's presentation at the Bernstein Strategic Decisions Conference was not a single number but a sequence of them. CEO David Goeckeler revealed that over roughly three forecast cycles, SanDisk's internal projection for data center NAND demand growth in calendar year 2026 has been revised upward four consecutive times — from mid-20s percent, to mid-40s, to mid-60s, and then higher still. "I don't think it's a mystery to anybody — data center is really, really, really growing aggressively," Goeckeler said, adding that this pace of upward revision has caught even his own management team off guard in its intensity.
The driver is structural, not cyclical. As hyperscalers and cloud infrastructure builders work through the architecture design process for large-scale inference deployments, they are consistently landing on higher NAND requirements than their initial estimates. Goeckeler explained that early AI infrastructure builds were essentially unconstrained resource grabs — "give me all the compute, give me all the memory, give me all the power" — but that the move toward scaled global inference forces a reckoning with economics. DRAM and HBM, while technically superior on latency, simply cannot scale to meet the density requirements of serving billions of inference queries globally. NAND can. "NAND is very scalable. It's the most scalable semiconductor technology in the world," Goeckeler said. "DRAM is spectacular technology. HBM is spectacular technology. It has unbelievable characteristics. It doesn't have the scale to solve a global inference issue."
KV Cache Migration to NAND: "It's What's Happening in Reality"
Bernstein analyst Mark Newman pressed Goeckeler on Jensen Huang's CES vision of roughly 17 terabytes of KV cache storage per GPU migrating to NAND-based storage. Goeckeler's response was unambiguous: this is not a forward-looking thesis but a current commercial reality. "It's much more than a vision. It's what's happening in reality, and it's been happening in reality for quite some time now," he said. The KV cache, which stores intermediate attention states for large language models during inference, is being moved to NAND because context windows are expanding and model sizes are growing — and DRAM simply cannot match NAND on cost per gigabyte or area density at the required scale.
Goeckeler cautioned against expecting a simple linear formula for how much NAND demand each GPU generates, noting that the final number depends on model size, token count, KV cache dimensions, cache hit rates, and the specific use case being built for. But the direction of travel is clear: as infrastructure designers dial in their architectures, the answer keeps coming back as needing more NAND, not less. "As companies need to take this brilliant AI technology and scale it so we can all use it," he said, "that's what's driving the market."
$42 Billion in Remaining Purchase Obligations, $11 Billion in Contractual Backstops
SanDisk's new business model disclosures deserve careful parsing. As of the most recent quarter-end, the company had signed three long-term supply contracts carrying approximately $42 billion in remaining purchase obligations — the minimum volume commitments across the life of those agreements. Two additional contracts were signed after quarter-end and are therefore not yet in the reported figures. Across all five contracts, the aggregate amount placed with third-party escrow holders as termination backstops — what Goeckeler calls the financial commitment in case a customer exits — totals $11 billion.
The structure is deliberate. SanDisk does not receive these funds upfront; they sit with a neutral third party under a predefined release mechanism. If a customer walks away from a contract, the escrow is released to SanDisk — no litigation, no negotiation. "We don't sue our customers," Goeckeler said flatly. The intent is alignment of incentives: customers think twice before forfeiting billions in escrow, and SanDisk gets a cash cushion precisely when it is likely to need one most, during a demand downturn. "If you do have to exit this contract, you're going to have to forgo billions. And so you better be sure," he said.
More than one-third of SanDisk's current volume sits inside these new business model agreements. Goeckeler declined to name customers but acknowledged they span a range of contract durations, by design. He wants staggered maturities — some one-year, some three-year, some five-year — to avoid concentration risk. Whether these are exclusively hyperscalers was not confirmed. Asked whether the remaining two-thirds of customers not yet in long-term agreements risk supply access, Goeckeler was candid: "I'm not the only supplier in the market" and "I can't run their business."
HBF: NAND Die Expected by Year-End, Full System in 2027
High-Bandwidth Flash, SanDisk's proprietary architecture designed to deliver NAND at near-DRAM-level access patterns for inference workloads, is now moving from concept to hardware. Goeckeler confirmed that the NAND die for HBF is currently in development and expected to be ready by end of 2025, with the full system — including the custom controller built on top — expected sometime in 2026. Customer integration work is already underway, though Goeckeler was explicit that this is not a plug-and-play component swap. "You got to get your customers to adopt it into what they're building, and we're going through that process," he said.
HBF is positioned as a density-optimized, read-dominant architecture suited specifically for inference, which Goeckeler characterized as "predominantly a read-based activity and a deterministic read-based activity." He was careful not to overstate it as a DRAM substitute or a replacement for enterprise SSDs. Instead, it is a new category within the memory hierarchy — one that SanDisk believes becomes economically necessary as inference scales globally and the bandwidth-per-dollar constraints of HBM become binding.
CapEx Efficiency Is the Under-Appreciated Part of the Growth Story
One of the more important points Goeckeler made, and one that tends to get lost in the headline demand discussion, is the relationship between nodal transitions and capital intensity. Each new NAND node delivers more bits per wafer at a compounded rate that exceeds SanDisk's stated mid-to-high-teens exabyte growth target. The implication: if SanDisk simply ran the same number of wafers through each successive node, it would systematically oversupply the market. Instead, it calibrates wafer starts downward per node while still growing output, which means CapEx as a percentage of revenue is declining even as the top line grows.
"If you look at our CapEx as a percent of revenue, it continues to go down as revenue goes up," Goeckeler said, describing this as the core mechanism by which the franchise generates free cash flow efficiently. For investors, this is the structural argument for why NAND economics are not as capital-destructive as historical cycles might suggest — provided supply discipline holds.
On Cyclicality: Goeckeler Is Trying to Break a 40-Year Industry Pattern
Goeckeler was unusually candid about the corrosive nature of the boom-bust cycle his industry has historically followed, and unusually direct about the ambition to structurally change it. His three-part framework — earn a fair return, reduce cyclicality, and grow — is simple, but he acknowledged the tensions between them explicitly. Pushing for higher near-term economics risks demand destruction; accepting more cyclicality in exchange for growth destroys value; and growing faster than mid-to-high teens may be feasible but at a cost to the other two objectives. "You have to do all three," he said. "And if you start messing with one of them, a different one goes in an opposite direction."
He also noted with some wry satisfaction that less than twelve months ago, analyst reports were projecting SanDisk would miss its December quarter numbers due to pricing declines. "I was reading reports as late as late summer that said, SanDisk is going to miss their numbers in December because pricing is going to be down. So it didn't quite work out that way." The ASP trajectory — Bernstein's Newman estimated approximately 140% quarter-on-quarter improvement in price per gigabyte in the most recent quarter — has since validated the demand thesis SanDisk was arguing for throughout 2024.
On whether capacity will be added aggressively given the current pricing environment, Goeckeler offered a structured answer: SanDisk is always adding capacity, it makes those decisions years in advance, and the nodal productivity dynamic means it can grow output without proportional wafer start increases. He was reluctant to imply any near-term acceleration, pointing out that it would be operationally incoherent to reverse course on investment levels less than a year after the market was calling those same investment levels excessive.