Title: High Bandwidth Flash (HBF): A Capacity Tier Beside HBM for Read-Mostly Inference Data. Sandisk and SK hynix Standardizing, Samples ~Late 2027, HVM 2029 "If All Goes Well," and Optane as the Cautionary Template
Author / source: Jon Y (Asianometry), solo narration
Source title: High Bandwidth Flash Is Coming
Source URL: https://www.youtube.com/watch?v=3nTpW52nioI
Video ID: 3nTpW52nioI
Published: 2026-09-24 (upload)
Duration: 20:53
Memo date: Sunday, September 27, 2026 (America/Toronto)
Transcript paths: /workspace/youtube-transcripts/3nTpW52nioI.md · 3nTpW52nioI_plain.txt · Brief: 3nTpW52nioI_brief.md
Caption quality: Creator-uploaded manual English captions (manual-en.json3), near-verbatim to the script, with section headers. This isn't ASR, so name and number rendering is reliable as narrated. The only oddities are deliberate script jokes ("defill and precode," "the state of Montana"). Spec numbers are as narrated and not independently verified (External check needed).
Source / conflict flags: Jon says the video idea was suggested by a Sandisk employee who "walked with me to Hot Chips across the Stanford lawn" (19:30–20:00). Sandisk is the lead HBF proponent, so treat the framing as potentially Sandisk-influenced, even though the video includes skeptical material (Optane, "HBF Sucks?", Agrawal & Giduthuri). The SemiAnalysis Q&A quote is relayed "via Irrational [Analysis]," and Jon says "I was there to confirm it happened." Irrational Analysis is described as a "good friend of the show."
Product: Memory-hierarchy / semis research map. Not investment advice. No buy/sell recommendations.
Source discipline: This transcript is the only primary source. The brief was corrected where it disagrees (Appendix B). The earlier Asianometry memo (2026-09-14-asianometry-silicon-valley-starving-compute.md) is cross-referenced as other desk memo, and no figures are imported. No market data was fetched. Live marks come only from TradingView/IBKR read-only, and none were pulled. The video names no stock prices, valuations or earnings, so every equity implication below is my inference.
Name locks (clean captions; for reference): Sandisk (SNDK); SK hynix; Samsung ("zNAND-O"); Micron (MU); Intel (Optane / 3D XPoint); Google; Tenstorrent; Open Compute Project (OCP); David Patterson; Raja Koduri; Pat Gelsinger; Kim Ho-sik (SK hynix VP); Alper Ilkbahar (Sandisk EVP & CTO, ex-VP/GM Intel Optane Group); A. Agrawal & R. Giduthuri (Hot Chips 2026 presenters); Koji Sakui & Takayuki Ohba (Honda Research Institute Japan, 2019); SemiAnalysis (unnamed analyst); Irrational Analysis; Peking University / Fudan ("HBF Sucks?").
Stance one-liner (source-locked): Inference is memory-bound (decode), and both model weights (Kimi K3 ~2.8TB at 8-bit) and KV caches (long-running agents, ~1M-token contexts) are exploding. HBM hits a capacity-vs-bandwidth-per-GB wall (TSV-limited). HBF stacks NAND like HBM stacks DRAM, giving ~8–16x the capacity per stack (~512GB vs. ~36GB HBM4) at comparable headline bandwidth, but with ~two orders of magnitude slower read/program, limited endurance and more power. It fits read-mostly data (weights; maybe shared KV/prompts; MoE experts) and needs heavy software work. The timeline is samples ~late 2027, HVM 2029 if all goes well. "Not something coming anytime soon, but it will come sooner than you think."
Overall conviction: High that the technical framing is accurate (clean source, multiple independent skeptic inputs). Med that HBF reaches HVM roughly on schedule (standardization momentum: SK hynix MoU, OCP v0.7.0, Google/Tenstorrent. Against it: the Optane precedent and software dependence). Low on any 2026–27 P&L impact for memory names. The earliest material revenue is 2029+ (inference).
Takeaway 1: The demand driver is inference memory pressure from weights and KV caches, and it's structural. Source (01:00–04:30): Prefill is compute-bound. Decode "happens one by one… more memory bound," so "the GPU's logic circuits waste their time waiting for data." "Kimi K3 has 2.8 trillion total parameters… about 2.8 terabytes" at 8-bit, versus K2.7 (~1TB) and DeepSeek V3 (671GB). Long-running agents make KV caches "absolutely massive." Context is "about 1 million tokens," and mitigations include compaction (lossy) and offloading less-read KV blocks to SSDs. Inference: Near-term NAND demand from KV-cache SSD offload is a separate, earlier channel than HBF itself. It's the 2026–27 NAND read, while HBF is the 2029+ read. Conviction: High on direction, Med on magnitude (External: model sizes, context norms).
Takeaway 2: HBM's scaling trade-off is getting publicly challenged, including by SK hynix's own VP. Source (04:30–07:30): HBM "turned all of Korea into degenerate semiconductor stock traders" and "remains a money-maker." At the Hot Chips 2026 Q&A, a SemiAnalysis analyst told SK hynix that ~20 TB/cm² at the cell level becomes ~4 TB/cm² in a 20-high stack: "You've diluted the throughput enormously… Why… taller rather than going faster?" Capacity comes only by cutting bandwidth per GB (TSV-limited), so "the price of incremental capacity and bandwidth [rises]." Gelsinger called HBM "lousy memory." SK hynix VP Kim Ho-sik "partly agreed," saying HBM is "not the final answer" but "the best thing on the market right now," and hynix is "working on other solutions." Irrational Analysis: HBM "is a mistake." Inference: This is a narrative overhang for the HBM-scarcity premium over the long term, but not a demand signal today. The incumbent (SK hynix) is co-sponsoring the alternative, which hedges rather than threatens its franchise. Conviction: Med.
Takeaway 3: HBF specs trade huge capacity against latency, endurance and power. Source (07:30–10:00): HBF "vastly increases the per-stack memory capacity by 8-16 times (512 gigabytes compared to 36 or so… per HBM4 stack)." The spec can do "up to 3 terabytes per second effectively," which is "higher than Micron's HBM3E's 1.2" and "comparable to… HBM4." But "adjusted for capacity, HBF's bandwidth numbers are far smaller." Downsides: (1) read/program "roughly two orders of magnitude" slower. "You will need to wait longer… end of story." (2) Endurance of "several thousands of cycles," versus DRAM's "essentially… no endurance limit." (3) Power and thermals, "not a trivial consideration." Inference: HBF is a capacity tier, not an HBM substitute, for hot data. The binding question is how much interposer shoreline architects cede from HBM to HBF. Conviction: High on the trade-off framing (External: spec values).
Takeaway 4: The ecosystem and timeline are standardization-first, with a multi-vendor design, and HVM is 2029 at best. Source (10:00–13:00, 20:00):
The spec lets competitors cooperate on the interface while competing internally, and gives buyers "a credible second source." Inference: Second-sourcing is the explicit anti-Optane design choice (Optane died partly on single-sourcing). Conviction: Med on schedule, High on the standardization facts (External: dates, OCP document).
Takeaway 5: Placement and use cases pit Tier-1 capacity against "cheap capacity ≠ cheap tokens." Source (13:00–16:30): Sandisk pitches HBF "side by side with HBM at the Tier 1 level": HBM for hot data, HBF for warm. Example: 8× 24GB HBM gives ~192GB, while 6 HBF + 2 HBM gives ~3.12TB. Data must be "read-mostly" (most people think weights; "some have suggested KV Cache too"). Agrawal & Giduthuri (Hot Chips 2026): "just because HBF has cheap memory capacity does not one-to-one mean cheaper tokens." If the accelerator waits, output suffers, and it "will require extensive software support." Peking/Fudan's "HBF Sucks?": "no drop-in for SSDs," and software must choose data carefully (e.g., shared prompts). MoE "might work better," since it holds all experts but reads only a few per token. Inference: The winners are software and systems integrators (inference-serving stacks, schedulers) as much as NAND vendors. MoE's prevalence in frontier models is a leading indicator of HBF fit. Conviction: Med–High.
Takeaway 6: Optane is the cautionary template, and Sandisk's CTO ran Optane. Source (16:30–19:30): 3D XPoint (PCM) had DRAM-like speeds at a mid price, with a "warm data" pitch that required application software support. 3D NAND's cost-per-bit decline "invalidated much of Optane's economic justification." Micron exited, leaving Intel as sole vendor: "it was over." Alper Ilkbahar, Sandisk EVP & CTO, "spent five years at Intel as VP and General Manager of the Intel Optane Group." A software-side Optane veteran told Jon that "Optane just couldn't do what Intel claimed." Counterpoint: "Optane's actual PCM cells worked brilliantly." It was the controller and go-to-market that failed. "Considering the DRAM shortage today," some muse Intel should have kept Optane. Inference: The HBF failure modes to watch are controller quality, software ecosystem, cost curve against HBM/DRAM, and single-vendor perception. The CTO's Optane history cuts both ways: lessons learned, or pattern repetition. Conviction: Med.
Non-obvious angle #1: HBF's sampling window (late 2027) coincides with the H2'27 compute-glut window Jon flagged on Sep 14. Cross-memo inference (other desk memo): In the Sep 14 memo, Jon faded "two more years of scarcity," with H2'27 capacity tipping shortage into glut. If that happens, architects evaluating HBF samples in late 2027 will face cost-per-token pressure (favoring cheaper capacity tiers) and capex caution (disfavoring new, unproven memory tiers). That's a two-sided adoption risk worth tracking. Speculative; neither video links these.
Non-obvious angle #2: the incumbents are hedged, and the cleaner single-name exposure is NAND-heavy and HBF-leading. Inference: SK hynix co-leads both HBM and HBF. Samsung has zNAND-O. Micron is not named as an HBF participant, only as the HBM3E/HBM4 bandwidth comparator and as Optane's exiting partner. If HBF succeeds, share shifts within the memory complex, and the relative read is Sandisk (HBF proponent, pure NAND) positive versus HBM-concentrated DRAM exposure at the margin in 2029+. This is a long-dated narrative, not a 2026–27 earnings effect. External: whether Micron has an HBF program not mentioned here.
Why now: Hot Chips 2026 memory-session debate (Gelsinger's "lousy memory," the SemiAnalysis Q&A), OCP v0.7.0 public spec (Aug 2026), the SEMICON Taiwan schedule reveal, the "DRAM shortage today" (spoken), and agentic workloads inflating KV caches.
Horizon: Now to 2027 for software ecosystem signals, spec iterations (v0.7 → v1.0), customer PoC announcements and the KV-cache SSD offload channel. Late 2027 for engineering samples. 2029+ for HVM and any P&L materiality.
Sleeve relevance (desk inference):
Overall conviction: High on technical framing. Med on schedule and adoption. Low on near-term financial impact.
What the PM can do from this summary alone (research agenda, not trades):
What this is not: It isn't a call that HBM is obsolete (HBF is "challenger and/or companion"). It isn't a near-term earnings thesis for any memory name. It isn't an endorsement of Sandisk's roadmap (a Sandisk employee suggested the video). No buy/sell.
Chronological. Timestamps are ~30s caption blocks. Quotes are ≤20 words. Single narrator.
A. Demand node: inference workloads (Source). Model weights (TB-scale, e.g., Kimi K3 ~2.8TB) + KV caches (agents, ~1M-token contexts) → memory-bound decode → accelerators stall on memory. Inference: Agentic inference intensity is the root driver. It's the same thesis as the Sep 14 memo's "long-running agents/agent swarms" (other desk memo).
B. Current memory hierarchy (Source). Tier 0 SRAM (on-die) → Tier 1 HBM (interposer shoreline; TSV-limited bandwidth per GB) → Tier 2 system DRAM → SSD (KV offload, capacity). Inference: The interposer shoreline is the scarce real estate. Every slot is a zero-sum choice among HBM, HBF and future options.
C. HBF supply chain (Source + Inference). NAND die (Sandisk, SK hynix; Samsung zNAND-O) → TSV stacking and advanced packaging (inference: OSAT/foundry packaging content) → controller/base logic (the Optane failure point) → OCP spec (interface standard, second source) → accelerator integration (Google TPU and Tenstorrent in the consortium; GPU vendors implied, not named) → software stack (data placement, schedulers, MoE-aware serving).
D. Economics node (Source → Inference). The HBM incremental cost of capacity and bandwidth is rising (Source). HBF gives cheap capacity, but cheap capacity doesn't equal cheap tokens if accelerators wait (Source). Inference: The HBF value proposition is $/token for capacity-bound, bandwidth-tolerant workloads (MoE experts, weights, shared prompts). It isn't a general-purpose HBM replacement. The 3D NAND cost curve helps here (it was Optane's killer, and it's HBF's tailwind).
E. Competitive node (Source). SK hynix hedges HBM with HBF, Samsung has zNAND-O, and Micron isn't named (it's the HBM comparator and Optane exiter). Intel is out (Optane). Inference: HBF success moves mix within memory vendors more than it moves share across them, except where a vendor lacks HBF (External: Micron).
F. Risk / execution node (Source). Latency (~100x), endurance (thousands of cycles), power and thermal, software dependence, controller quality, go-to-market, schedule ("if all goes well"). The Optane precedent and the CTO's Optane history.
G. Source-influence layer. The idea was suggested by a Sandisk employee, Irrational Analysis is a friend of the show, and SemiAnalysis is quoted through Irrational (Jon was present). The skeptical inputs are independent (Agrawal & Giduthuri; Peking/Fudan; the Optane software insider).
Chain 1: Exploding KV caches → SSD offload today → enterprise SSD/NAND demand before HBF exists (Source → Inference). 1st: Agents inflate KV caches (Source). 2nd: Less-read KV blocks get offloaded to SSDs (Source). 3rd (Inference): 2026–27 NAND demand uplift from inference storage tiers is a channel independent of HBF, and a nearer-dated positive for NAND suppliers. Confirm: enterprise SSD bit growth tied to inference, inference-serving software adopting KV offload (External). Disconfirm: compaction or architectural changes shrink KV footprint.
Chain 2: HBM taller stacks → diluted bandwidth per GB → rising incremental cost → HBF as a pressure valve (Source → Inference). 1st: TSV limits mean capacity costs bandwidth per GB (Source). 2nd: The industry openly questions HBM's trajectory (Gelsinger, SemiAnalysis, the hynix VP) (Source). 3rd (Inference): HBF puts a long-dated ceiling on HBM pricing power for capacity-driven use cases from ~2029, while HBM keeps hot-data pricing power. The desk read is a gradual narrative shift, not a cliff. Disconfirm: HBM4/HBM4E+ architectures (e.g., wider interfaces, hybrid bonding) fix bandwidth per GB, and HBF becomes a niche.
Chain 3: Standardization + second source → avoid the Optane single-vendor death → broader adoption odds (Source → Inference). 1st: Sandisk–SK hynix MoU, OCP v0.7.0, Google/Tenstorrent (Source). 2nd: Buyers get a credible second source (Source). 3rd (Inference): Adoption probability rises versus Optane's path. But standard interfaces commoditize HBF, so vendor margins could compress relative to HBM's scarcity economics. The winners may be volume NAND vendors with packaging scale. Watch: spec v1.0, the number of consortium members, Samsung joining versus forking (zNAND-O) (External).
Chain 4: Software gating → MoE and data-placement stacks decide HBF's value → the software ecosystem becomes the kingmaker (Source → Inference). 1st: "Extensive software support" is required, and HBF is no drop-in (Source). 2nd: MoE and shared-prompt workloads fit best (Source). 3rd (Inference): Inference-serving frameworks and hyperscaler in-house stacks determine the adoption pace, so Google (consortium member, TPU) is a likely first mover and merchant GPU adoption is less certain. Disconfirm: no major serving stack adds HBF-aware placement by 2028.
Chain 5: Compute glut risk in H2'27 (Sep 14 memo) × HBF sampling in late 2027 → adoption decision under cost pressure (Cross-memo inference). 1st: The Sep 14 memo's H2'27 shortage-to-glut tip (other desk memo). 2nd: HBF samples reach customers in late 2027 (Source). 3rd (Inference): A glut raises the focus on $/token efficiency (a potential HBF tailwind) but cuts new-architecture capex appetite (a headwind). Net: adoption variance widens. Watch: capex guidance in 2027, customer PoC counts.
Horizons: near is now to end-2027 (spec and samples). Medium is 2028–2030 (HVM and ramp). Probabilities are my inference; the video gives none.
Assumptions: Samples arrive on time in late 2027 and HVM in 2029. Google TPU and at least one merchant accelerator roadmap adopt HBF. MoE-heavy frontier models dominate. Serving stacks add HBF-aware placement. Endurance and power are acceptable for weights. Samsung converges on the standard. Winners (hypotheses): Sandisk (HBF proponent), SK hynix (hedged in both), NAND bit demand, advanced-packaging content (TSV stacking), inference-serving software. Losers: HBM capacity-tier pricing power at the margin (2029+), DRAM-only exposure without an HBF program (Micron status External). Leading indicators: Spec v1.0, customer PoC announcements, accelerator roadmaps citing HBF (External).
Assumptions: Schedule slips 6–18 months ("if all goes well" doesn't hold). HBF lands as a niche capacity tier for specific inference deployments (weights for MoE) at hyperscalers with in-house stacks. HBM stays dominant for hot data. The near-term NAND read is driven by KV-cache SSD offload, not HBF. Winners: NAND vendors via SSD offload (2026–27), HBM incumbents (little displacement before 2030). Losers: Aggressive "HBM is obsolete" narratives. Leading indicators: A slow cadence of spec revisions, limited merchant GPU commitments (External).
Assumptions: It repeats Optane. Controller or software shortcomings, latency and endurance limit use cases, power and thermal constraints bind in dense racks, HBM4E+ closes the capacity gap, or DRAM/HBM cost curves undercut the economics. Samsung forks (zNAND-O) and fragments the standard. Winners: HBM incumbents' franchise, conventional SSD tiers. Losers: Sandisk's HBF optionality, the consortium narrative. Leading indicators: Sample delays, absent customer PoCs, more negative academic or industry benchmarks in the vein of "HBF Sucks?" (External).
Hypotheses, not tickets. The video gives no prices or valuations, so External check needed on all financial data.
| Asset / sleeve | Thesis link (Source) | Metrics to watch | Catalysts | Risks / disconfirm |
|---|---|---|---|---|
| Sandisk (SNDK) | HBF originator (patent May 2024; Feb 2025 investor day; SEMICON Taiwan roadmap); CTO ex-Optane GM | HBF sample timing, customer PoCs, enterprise SSD (KV offload) mix | Spec v1.0; late-2027 samples | Optane repeat; schedule slip; source influence (a Sandisk employee suggested the video) |
| SK hynix | HBM leader and HBF co-standardizer (MoU Aug 2025); VP "not the final answer" | HBM vs. HBF roadmap commentary; HBM pricing | Hot Chips / earnings commentary | HBM narrative overhang; hedged |
| Samsung | zNAND-O rival | Standard convergence vs. fork | Product disclosures | Fragmentation |
| Micron (MU) | HBM3E 1.2 TB/s comparator; Optane partner that exited; not named as HBF participant | Any HBF/high-capacity tier program (External) | — | Absent from HBF ecosystem (if confirmed) |
| Google (TPU) | HBF consortium member; Patterson link | TPU memory roadmap | TPU announcements | In-house only |
| Tenstorrent (private) | Consortium member | Product roadmap | — | Scale |
| GPU/ASIC vendors (NVIDIA etc.; implied, not named) | Shoreline allocation decision | Roadmaps citing HBF | GTC-type events | Non-adoption |
| Advanced packaging / TSV (SMH-adjacent) | HBF uses "advanced packaging voodoo" TSV stacking | Packaging capacity, content per stack | HBF ramp | Timeline |
| Inference-serving software (IGV-adjacent) | "Extensive software support" required; data placement | HBF-aware features in serving stacks | Framework releases | Indirect |
| DRAM / SMH sleeves | HBM "money-maker" now; long-dated narrative risk | HBM ASPs, shoreline mix | — | Near-term: none |
| Enterprise SSD / NAND (near-term channel) | KV-cache offload to SSDs | Inference-driven SSD bit demand | Agent adoption | Compaction reduces need |
Thesis risk: HBF could repeat Optane: good cells, failed controller, software or go-to-market. HBM roadmaps could close the capacity gap. "Cheap capacity ≠ cheap tokens" could cap the use cases to niches.
Timing risk: "If all goes well" signals slippage risk. Memory products routinely slip. Nothing affects P&Ls before 2029 at the earliest. Equity narratives can front-run and then disappoint.
Execution / expression risk: Incumbents are hedged (SK hynix), so "HBF winner vs. HBM loser" pairs are impure. Sandisk is the cleanest proponent, but its near-term results are driven by NAND pricing and SSD demand, not HBF. The equity read needs External financials, and none were pulled.
External / identification risk: Clean captions, but spec numbers are narrated, not verified. The SemiAnalysis quote is relayed via Irrational Analysis. The "Optane couldn't do what Intel claimed" line is anonymous. Micron's HBF status is unknown.
Source-influence risk: The idea was suggested by a Sandisk employee. The skeptical sections mitigate, but don't eliminate, the framing bias.
Process risk: Don't import Sep 14 memo figures without labels. Don't email, publish, forward or message.
| Item | Narrated | Timestamp | Caveat |
|---|---|---|---|
| Kimi K3 | 2.8T params (~2.8TB at 8-bit) | 03:00 | External |
| Kimi K2.7 / DeepSeek V3 | ~1TB / 671GB | 03:00 | External |
| Context window | ~1M tokens (~1,500–3,000 pages; audio 11–12h; ~1h HD video) | 04:00 | Approximate |
| HBM stack heights | 4/8/12/16 core dies | 05:00 | — |
| Cell vs. stack bandwidth density | ~20 TB/cm² vs. ~4 TB/cm² (20-high) | 06:00 | SemiAnalysis Q&A via Irrational |
| HBF capacity | ~512GB per stack (8–16x) vs. ~36GB HBM4 | 08:00 | Spec narration |
| HBF bandwidth | Up to ~3 TB/s effective vs. HBM3E 1.2 TB/s; ~HBM4 | 08:30 | Capacity-adjusted far lower |
| Latency gap | ~2 orders of magnitude slower read/program | 09:00 | — |
| Endurance | Several thousand cycles | 09:30 | — |
| Shoreline example | 8× 24GB = 192GB → 6 HBF + 2 HBM = 3.12TB | 14:30 | Arithmetic consistent (6×512 + 2×24) |
| Dates | 2019 paper; May 2024 patent; Feb 2025 unveiling; Jul 2025 board; Aug 2025 MoU; Aug 2026 OCP v0.7.0 | 10:00–12:30 | External |
| Schedule | Samples ~late 2027; HVM 2029 "if all goes well" | 20:00 | Sandisk slide |
| Optane partnership | Intel–Micron 2015; Micron exit | 17:30–18:00 | Historical |
| Ilkbahar Optane tenure | 5 years VP/GM | 18:30 | External |
Where the transcript disagrees with or adds to the brief / Monica's angle:
External checks needed: OCP spec; Sandisk roadmap slide; HBM4/HBM3E specs; Kimi K3 and other model sizes; Micron HBF status; Samsung zNAND-O; Agrawal & Giduthuri and "HBF Sucks?" papers; Ilkbahar bio; KV-cache SSD demand data; any equity data (prices and valuations), via TV/IBKR only.
Document control: Saved to /workspace/pm-memos/2026-09-27-asianometry-high-bandwidth-flash.md. Published to the desk library on 27 Sep 2026: https://andrepow.here.now/memos/asianometry-high-bandwidth-flash/. Not emailed, forwarded or messaged. No trades.
End of memo.
Desk copy · not a trade recommendation · Erica · 27 Sep 2026