SanDisk的HBF技术进展与商业化前景
SanDisk的HBF技术进展与商业化前景
Conclusion: HBF is a real, standardized memory concept aimed at a real bottleneck, but as of 20 Nov 2026 it is a pre-revenue option, not a contracted growth driver. The evidence supports 2027 samples. Production in 2028 is a broker projection, and no named customer has committed to HBF in a contract. Near-term value to SanDisk still rests mainly on NAND demand from AI inference and on existing customer agreements. Treat HBF as upside that the base case does not need.
Data as of: 20 Nov 2026. The latest HBF-specific source I found is the Morgan Stanley summit transcript dated 19 Nov 2026. I did not fetch current quotes because the question is about technology and commercial timing, not price.
1. The problem HBF targets
- The "memory wall" in inference. Inference needs both high bandwidth and large capacity near the GPU. HBM is limited in capacity, costly, and power-hungry. SanDisk's pitch is HBF at HBM's capacity at similar bandwidth and cost. Its first generation targets read bandwidth and 512GB per 16-die stack (company fact sheet).
- Company simulation. SanDisk reports performance of an unlimited-capacity HBM in one simulated Llama 3.1 405B workload. This is internal modeling, not third-party measurement.
- Complement, not replacement. Mizuho's view, as reported by TrendForce, is that HBF is unlikely to replace HBM outright. It could enable disaggregated architectures. TrendForce expects a hybrid HBM-plus-HBF hierarchy.
- Durability angle. On the 19 Nov call, management said customers want SLC-type durability for KV-cache and scratch-pad writes, and that these use cases could start to drive "tens of exabytes" of demand over a couple of years.
2. How it could affect SanDisk's opportunity
Upside paths
- New product category. Inference products with HBF are targeted for 2027 samples. If they win design slots, HBF becomes a new revenue line.
- NAND bit demand. Critics argue that SLC endurance requirements would sharply increase NAND bit consumption per delivered GB. That is a commentary argument with unverified modeling, not company data. Management did say SLC could be produced from existing products without a new node, so the demand effect can be large.
- Standards position. Publishing the spec through OCP, with Google and Tenstorrent joining, positions HBF as an open standard rather than a single-vendor product.
Constraints
- Not in contracts. Management said on 19 Nov that "HBF is not included in our contracts." The visible demand base is the New Business Model (NBM). Customer agreements cover about of bits in FY2027 and two-thirds in FY2028, per the Q1 FY27 call (29 Oct 2026). Sandisk reports eight NBM customers, including three US hyperscalers, via Reuters as summarized by TrendForce.
- Performance trade-offs. OXMIQ modeled HBF at about HBM's capacity but only its aggregate bandwidth. Its conclusion was "HBM for the rack, HBF for the box." Its best cases are very large MoE expert pools, long-context KV cache, and fitting huge models on fewer GPUs. OXMIQ also says inference software (e.g., vLLM) needs dedicated HBF support.
- Latency and endurance. HBF latency is microseconds versus HBM's nanoseconds. Endurance is bounded (commentary cites about 100,000 cycles). Both shape which workloads fit.
- Manufacturing route unresolved. SanDisk has no independent fab and relies on its Kioxia JV. Blocks & Files reports no sign the JV is producing HBF wafers for SanDisk. Kioxia is not in the HBF consortium and favors a serial, non-stacked design. The source route for HBF wafers is an open question I could not resolve.
- Customer evidence is thin. WCCFtech (28 Apr 2026) reported that NVIDIA is not interested in HBF, and that Google is a reported customer. That report relies on industry sources and a social post, so I rate it weak. A KAIST professor's claim of integration into Nvidia, AMD, and Google products by late 2027 or early 2028 is advocacy, not a company commitment.
3. Evidence timeline
Timing note: The August 2025 blog said memory samples would arrive in 2H26 and inference devices in early 2027. Later statements focus on 2027 for product samples. I did not find a reaffirmed 2H26 memory-sample date. XenoSpectrum notes that the Investor Day materials give no production year. The 2028 production date comes from broker notes.
4. Commercial timing read
- 2026: Standards and tape-out only. No revenue.
- 2027: Samples. Customer evaluation is the gating step. I see no evidence of material HBF revenue in 2027.
- 2028: Earliest plausible production, per brokers. It needs yield, a controller, runtime support, and at least one committed design win.
- Guidance: Management's mid-to-high-teens revenue growth outlook for FY28–FY30 is tied to NBM contracts that, per the 19 Nov call, exclude HBF.
5. Analysis and what would change the view
Bull case: SLC-mode inference samples ship in 2027 with the stated bandwidth; a named accelerator design win (Google is the most-cited candidate); NVIDIA or AMD runtime support; customer prepayment or NRE funding; and disclosure of the wafer-supply route.
Bear case: Samples slip past 2027; TLC-only products hit bandwidth limits and find only niche uses; NVIDIA stays on HBM plus eSSD; software is slow to support multi-tier memory; and HBM capacity expands fast enough that the intermediate tier is not needed.
Checkpoints to watch:
- Cell mode of the 2027 samples (SLC versus TLC).
- Named customers or design wins, and whether any HBF volume enters contracts.
- Disclosure of the HBF wafer source, including any Kioxia JV arrangement.
- Real bandwidth and endurance data versus the simulated figures.
- Commentary in the next two quarterly calls.
My view: HBF is credible optionality with a 2028-plus horizon, not a near-term earnings driver. The investment case should rest on NAND demand and the NBM contracts. HBF deserves little weight in valuation until a customer commits to it in writing.