✦ High Bandwidth FlashNarrow moat

Sandisk (SNDK) — the future bets

High Bandwidth Flash is the only thing on these pages that could turn a commodity supplier into a memory supplier — and the sample date has already moved once.

High Bandwidth Flash is the only thing on these pages that could turn a commodity supplier into something else.

High Bandwidth Flash roadmap, per stack512GBGen 1 capacity (GB)1TBGen 2 capacity (GB)1.5TBGen 3 capacity (GB)1.6 TB/sGen 1 bandwidth(GB/s)>3.2 TB/sGen 3 bandwidth(GB/s)Gen 2 at 0.8x gen 1 power, gen 3 at 0.64x. Claimed 8-16x HBM capacity at similar cost.
Taped out and specified through the OCP with SK hynix — and two years from production, against an incumbent on its fourth generation.

The concept: stack NAND rather than DRAM in an HBM-style package and put it next to the accelerator. Per Sandisk's own fact sheet, the first generation targets 512GB per stack from sixteen 256Gb die at 1.6 TB/s of read bandwidth, claiming up to 8 to 16 times the capacity of HBM at similar cost, in a package closely matching HBM4's footprint, stack height and power — and, because it is NAND, non-volatile with no refresh power. The roadmap runs to a second generation above 2 TB/s and 1TB per stack at 0.8× the power, and a third past 3.2 TB/s and 1.5TB at 0.64×.1

The target workloads are stated plainly: mixture-of-experts models, long context lengths and large KV caches — the places where inference is capacity-bound rather than compute-bound. Sandisk describes three deployments: HBF augmenting HBM in some stack positions, replacing it outright, or sitting disaggregated while a smaller HBM tier acts as cache.2 That flexibility matters, because it lets HBF into a socket without an accelerator vendor abandoning HBM.

The progress is real. Sandisk taped out the first HBF die, shown at its 13 August 2026 investor day, with first inference product samples listed for 2027 and production reported for 2028.3 On 3 August it and SK hynix published the first HBF technical specification through the Open Compute Project.4

Rated narrow, and the trajectory is widening. A taped-out die and an open specification with a second major producer behind it are much more than a slide.

Watch the sample date. In August 2025 samples were targeted for the second half of calendar 2026; in August 2026 they are listed as coming soon in 2027.

Moat trajectory: Widening

First HBF die taped out and shown at the August 2026 investor day, with the first technical specification published through the OCP with SK hynix. A year ago this was a slide; the sample date slipping from 2H 2026 to 2027 is the offset.

The number that tests this moat
Reported
Operating income
$12,389M in fiscal 2026 (61.3% of revenue)

High Bandwidth Flash has taped out its first die, with samples due in 2027, and is funded from current profits. Profits at this level pay for the bet; a sharp fall would test how long it can be funded.

Source: Sandisk Form 10-K, FY2026 ↗
References
  1. ReportedThe roadmap runs to a second generation above 2 TB/s and 1TB per stack at 0.8× the power, and a third past 3.2 TB/s and 1.5TB at 0.64×. The target workloads are stated plainly: mixture-of-experts models, long context lengths and large KV caches — the places where inference is capacity-bound rather than compute-bound.
    Coverage of Sandisk's 2026 Investor Day of 13 August 2026 and its High Bandwidth Flash programme. Sandisk has taped out the first High Bandwidth Flash memory die, shown on an HBF Roadmap slide next to a die photograph, with the first HBF inference product samples listed as coming soon in 2027; a tapeout means the design is finished and committed to a mask set, and is several steps short of a product, with wafers still to return from the fab, yields to climb, and the die to survive stacking into 8-high and 16-high configurations with a working logic die and controller, followed by thermal and endurance qualification, accelerator software work and customer qualification cycles. Per Sandisk's HBF fact sheet, the first generation targets 512GB per stack built from sixteen 256Gb die at 1.6 TB/s of read bandwidth, which Sandisk claims lands at up to 8 to 16 times the capacity of HBM at a similar cost, in a package closely matching HBM4's footprint, stack height and power profile, and which is non-volatile and spends no power on refresh; a second generation targets more than 2 TB/s and up to 1TB per stack at 0.8 times the first generation's power, and a third pushes past 3.2 TB/s and up to 1.5TB per stack at 0.64 times the power. The Investor Day deck frames the target workloads as mixture-of-experts LLMs, long context lengths and large KV caches, with the architecture developed using input from major cloud and AI customers, and lays out three deployments: HBF augmenting HBM around an xPU, replacing HBM stacks outright in a similar footprint, or sitting disaggregated holding decode weights and KV cache while a smaller HBM tier acts as cache. Sandisk showed an inference token output comparison from which it draws an 8x capex efficiency claim (one HBF GPU against eight HBM GPUs) and a 2x GPU efficiency claim (four HBF GPUs delivering the same token output as eight HBM GPUs), both labelled as based on internal testing; the fact sheet adds a simulation result of HBF landing within 2.2% of unlimited-capacity HBM when reading pretrained weights for Llama 3.1 405B. These are vendor numbers on unreleased silicon and the chart carries no axis values. On timing, in August 2025 Sandisk said the first HBF samples were targeted for the second half of calendar 2026 with AI inference devices expected in early 2027; as of the August 2026 Investor Day the first HBF inference product samples are listed as coming soon in 2027, and mass production is reported for 2028. On 3 August 2026 Sandisk and SK hynix released the first HBF technical specification through the Open Compute Project, six months after the consortium formed, with Google and Tenstorrent among the contributors. — August 2026 · publ. 2026-08-18 · source ↗
  2. ReportedSandisk describes three deployments: HBF augmenting HBM in some stack positions, replacing it outright, or sitting disaggregated while a smaller HBM tier acts as cache. That flexibility matters, because it lets HBF into a socket without an accelerator vendor abandoning HBM.
    Coverage of Sandisk's 2026 Investor Day of 13 August 2026 and its High Bandwidth Flash programme. Sandisk has taped out the first High Bandwidth Flash memory die, shown on an HBF Roadmap slide next to a die photograph, with the first HBF inference product samples listed as coming soon in 2027; a tapeout means the design is finished and committed to a mask set, and is several steps short of a product, with wafers still to return from the fab, yields to climb, and the die to survive stacking into 8-high and 16-high configurations with a working logic die and controller, followed by thermal and endurance qualification, accelerator software work and customer qualification cycles. Per Sandisk's HBF fact sheet, the first generation targets 512GB per stack built from sixteen 256Gb die at 1.6 TB/s of read bandwidth, which Sandisk claims lands at up to 8 to 16 times the capacity of HBM at a similar cost, in a package closely matching HBM4's footprint, stack height and power profile, and which is non-volatile and spends no power on refresh; a second generation targets more than 2 TB/s and up to 1TB per stack at 0.8 times the first generation's power, and a third pushes past 3.2 TB/s and up to 1.5TB per stack at 0.64 times the power. The Investor Day deck frames the target workloads as mixture-of-experts LLMs, long context lengths and large KV caches, with the architecture developed using input from major cloud and AI customers, and lays out three deployments: HBF augmenting HBM around an xPU, replacing HBM stacks outright in a similar footprint, or sitting disaggregated holding decode weights and KV cache while a smaller HBM tier acts as cache. Sandisk showed an inference token output comparison from which it draws an 8x capex efficiency claim (one HBF GPU against eight HBM GPUs) and a 2x GPU efficiency claim (four HBF GPUs delivering the same token output as eight HBM GPUs), both labelled as based on internal testing; the fact sheet adds a simulation result of HBF landing within 2.2% of unlimited-capacity HBM when reading pretrained weights for Llama 3.1 405B. These are vendor numbers on unreleased silicon and the chart carries no axis values. On timing, in August 2025 Sandisk said the first HBF samples were targeted for the second half of calendar 2026 with AI inference devices expected in early 2027; as of the August 2026 Investor Day the first HBF inference product samples are listed as coming soon in 2027, and mass production is reported for 2028. On 3 August 2026 Sandisk and SK hynix released the first HBF technical specification through the Open Compute Project, six months after the consortium formed, with Google and Tenstorrent among the contributors. — August 2026 · publ. 2026-08-18 · source ↗
  3. ReportedSandisk taped out the first HBF die, shown at its 13 August 2026 investor day, with first inference product samples listed for 2027 and production reported for 2028. On 3 August it and SK hynix published the first HBF technical specification through the Open Compute Project.
    Coverage of Sandisk's 2026 Investor Day of 13 August 2026 and its High Bandwidth Flash programme. Sandisk has taped out the first High Bandwidth Flash memory die, shown on an HBF Roadmap slide next to a die photograph, with the first HBF inference product samples listed as coming soon in 2027; a tapeout means the design is finished and committed to a mask set, and is several steps short of a product, with wafers still to return from the fab, yields to climb, and the die to survive stacking into 8-high and 16-high configurations with a working logic die and controller, followed by thermal and endurance qualification, accelerator software work and customer qualification cycles. Per Sandisk's HBF fact sheet, the first generation targets 512GB per stack built from sixteen 256Gb die at 1.6 TB/s of read bandwidth, which Sandisk claims lands at up to 8 to 16 times the capacity of HBM at a similar cost, in a package closely matching HBM4's footprint, stack height and power profile, and which is non-volatile and spends no power on refresh; a second generation targets more than 2 TB/s and up to 1TB per stack at 0.8 times the first generation's power, and a third pushes past 3.2 TB/s and up to 1.5TB per stack at 0.64 times the power. The Investor Day deck frames the target workloads as mixture-of-experts LLMs, long context lengths and large KV caches, with the architecture developed using input from major cloud and AI customers, and lays out three deployments: HBF augmenting HBM around an xPU, replacing HBM stacks outright in a similar footprint, or sitting disaggregated holding decode weights and KV cache while a smaller HBM tier acts as cache. Sandisk showed an inference token output comparison from which it draws an 8x capex efficiency claim (one HBF GPU against eight HBM GPUs) and a 2x GPU efficiency claim (four HBF GPUs delivering the same token output as eight HBM GPUs), both labelled as based on internal testing; the fact sheet adds a simulation result of HBF landing within 2.2% of unlimited-capacity HBM when reading pretrained weights for Llama 3.1 405B. These are vendor numbers on unreleased silicon and the chart carries no axis values. On timing, in August 2025 Sandisk said the first HBF samples were targeted for the second half of calendar 2026 with AI inference devices expected in early 2027; as of the August 2026 Investor Day the first HBF inference product samples are listed as coming soon in 2027, and mass production is reported for 2028. On 3 August 2026 Sandisk and SK hynix released the first HBF technical specification through the Open Compute Project, six months after the consortium formed, with Google and Tenstorrent among the contributors. — August 2026 · publ. 2026-08-18 · source ↗
  4. ReportedOn 3 August it and SK hynix published the first HBF technical specification through the Open Compute Project. Rated narrow, and the trajectory is widening.
    Coverage of Sandisk's 2026 Investor Day of 13 August 2026 and its High Bandwidth Flash programme. Sandisk has taped out the first High Bandwidth Flash memory die, shown on an HBF Roadmap slide next to a die photograph, with the first HBF inference product samples listed as coming soon in 2027; a tapeout means the design is finished and committed to a mask set, and is several steps short of a product, with wafers still to return from the fab, yields to climb, and the die to survive stacking into 8-high and 16-high configurations with a working logic die and controller, followed by thermal and endurance qualification, accelerator software work and customer qualification cycles. Per Sandisk's HBF fact sheet, the first generation targets 512GB per stack built from sixteen 256Gb die at 1.6 TB/s of read bandwidth, which Sandisk claims lands at up to 8 to 16 times the capacity of HBM at a similar cost, in a package closely matching HBM4's footprint, stack height and power profile, and which is non-volatile and spends no power on refresh; a second generation targets more than 2 TB/s and up to 1TB per stack at 0.8 times the first generation's power, and a third pushes past 3.2 TB/s and up to 1.5TB per stack at 0.64 times the power. The Investor Day deck frames the target workloads as mixture-of-experts LLMs, long context lengths and large KV caches, with the architecture developed using input from major cloud and AI customers, and lays out three deployments: HBF augmenting HBM around an xPU, replacing HBM stacks outright in a similar footprint, or sitting disaggregated holding decode weights and KV cache while a smaller HBM tier acts as cache. Sandisk showed an inference token output comparison from which it draws an 8x capex efficiency claim (one HBF GPU against eight HBM GPUs) and a 2x GPU efficiency claim (four HBF GPUs delivering the same token output as eight HBM GPUs), both labelled as based on internal testing; the fact sheet adds a simulation result of HBF landing within 2.2% of unlimited-capacity HBM when reading pretrained weights for Llama 3.1 405B. These are vendor numbers on unreleased silicon and the chart carries no axis values. On timing, in August 2025 Sandisk said the first HBF samples were targeted for the second half of calendar 2026 with AI inference devices expected in early 2027; as of the August 2026 Investor Day the first HBF inference product samples are listed as coming soon in 2027, and mass production is reported for 2028. On 3 August 2026 Sandisk and SK hynix released the first HBF technical specification through the Open Compute Project, six months after the consortium formed, with Google and Tenstorrent among the contributors. — August 2026 · publ. 2026-08-18 · source ↗
Sources
Generated September 23, 2026