Damnang
· Damnang’s Substack
· August 18, 2026 at 16:24
· ⏱ 12 min read
| Read on Substack ↗
Summary
Sandisk's HBF is technically plausible only in a narrow configuration where read-dominant model weights sit in HBF and KV cache stays in HBM; the all-HBF approach fails on write endurance and thermal limits. That makes any Sandisk upside conditional on design wins, especially at Google and Meta, rather than a broad replacement of HBM.
•The article judges HBF Only to have 'low applicability' to general conversational inference, citing Li et al. 2026 which found write traffic exceeded read traffic in every trace, thermal limits hit before peak bandwidth, and TLC wearout outpaced the SSD pool being replaced.
•HBF (+HBM), with model weights in HBF and KV cache in HBM, is the most realistic base case; at k=4 node memory rises to 2.2TB from 288GB while bandwidth falls to 17.4TB/s from 22TB/s.
•Sandisk's Investor Day claim that four HBF-only accelerators matched eight HBM-only accelerators, with '8x capital efficiency and 2x GPU efficiency,' is challenged because the comparison used 192GB HBM per GPU and BF16 precision.
•HBF must also beat disaggregated designs without HBF, such as GDDR7-based prefill accelerators plus SRAM racks for decode, on cost per token or system efficiency.
•The article argues Google and Meta have a 'relatively strong incentive to evaluate HBF' while NVIDIA is judged unlikely to adopt it.
•The bottom-up model sizes direct HBF revenue from accelerator unit pools, HBF attach rates, and stacks per accelerator at Google and Meta, then adds second-order effects on product mix and NAND wafer absorption.
The article explicitly builds a bottom-up case that HBF 'could move Sandisk's share price if it reaches commercial deployment' and identifies NAND wafer absorption and product mix as second-order reve
The article explicitly builds a bottom-up case that HBF 'could move Sandisk's share price if it reaches commercial deployment' and identifies NAND wafer absorption and product mix as second-order revenue drivers.
Risk: Feasibility is contested; the HBF-only KV-cache path is called unsustainable without SSD-class write management, so commercialization timing and customer commitments remain uncertain.
Article says Google and Meta have a 'relatively strong incentive to evaluate HBF' and uses Google's accelerator unit pool as a primary driver of Sandisk direct HBF revenue.
Article says Google and Meta have a 'relatively strong incentive to evaluate HBF' and uses Google's accelerator unit pool as a primary driver of Sandisk direct HBF revenue.
Risk: Google could favor competing disaggregated or SRAM-based designs, which the article says HBF must beat on cost per token or system efficiency.
Meta is named alongside Google as an incentivized HBF evaluator and is modeled as a source of HBF stacks per accelerator in the bottom-up Sandisk revenue estimate.
Meta is named alongside Google as an incentivized HBF evaluator and is modeled as a source of HBF stacks per accelerator in the bottom-up Sandisk revenue estimate.
Risk: No HBF commitment is disclosed, and Meta may prefer alternative disaggregated memory architectures.
The article judges that 'NVIDIA is unlikely to' adopt HBF, which implies NVIDIA's HBM-centric accelerator memory roadmap is not being disrupted by Sandisk's NAND-based HBF.
The article judges that 'NVIDIA is unlikely to' adopt HBF, which implies NVIDIA's HBM-centric accelerator memory roadmap is not being disrupted by Sandisk's NAND-based HBF.
Risk: If HBF economics improve and hyperscaler ASICs gain share in inference, NVIDIA could eventually face competitive pressure to respond.
This newsletter, published August 18, 2026,
features Damnang
discussing SNDK, GOOGL, META, NVDA.
4 trade ideas extracted by AI with direction and confidence scoring.