The $94 Billion Re-Pricing of AI Memory: Inside Sandisk’s HBF Architecture and the Future of HBM
Sandisk held its Investor Day on August 13 and laid out a financial model unthinkable for a NAND manufacturer eighteen months ago: roughly 80% non-GAAP gross margin, 75% operating margin, and 50% adjusted free-cash-flow margin for fiscal years 2028–2030, on mid-to-high-teens revenue growth. The stock surged 15.5%, dragging SK hynix (+8.2%), Western Digital (+8.0%), and Micron (+6.3%) higher. Shares have gained about 28% over four sessions and roughly 550% year-to-date.
The Contractual Pricing Machine
Those margins rest on multi-year take-or-pay-style contracts. Sandisk disclosed non-binding memoranda with eight customers covering approximately half of FY2027 bit output and two-thirds of FY2028 bits. Evercore's Amit Daryanani estimated total contract value at about $93.9 billion at floor pricing, spanning up to five years and including three U.S. hyperscale operators. Sandisk cites $91.1 billion in remaining performance obligations and $16.5 billion of financial guarantees.
That $16.5 billion covers only about 18% of the minimum contract value—a detail worth weighing. Sandisk has traded spot-price volatility for counterparty concentration and deployment-risk exposure.
Fiscal Q4 revenue hit $8.97 billion (+372% YoY). Gross margin reached 84.6%, with two-thirds of sequential growth coming from higher pricing. Full-year datacenter revenue grew 437% to $5.15 billion. September-quarter guidance: $10.3–$10.8 billion.
HBF: Real Standard, No Revenue Yet
Sandisk and SK hynix released the first open High-Bandwidth Flash specification through the Open Compute Project. Google and Tenstorrent participated in validation. The spec defines packages up to 512 GB across three bandwidth grades (roughly 0.4–3.0 TB/s), using UCIe interfaces attachable to multiple accelerator types.
Gen-1 targets 512 GB in a 16-die stack with 1.6 TB/s read bandwidth, matching approximately HBM4's footprint and power envelope. Sandisk claims 8–16× HBM capacity at similar stack cost—an order-of-magnitude cost-per-gigabyte reduction for qualifying workloads.
TC-bonder vendors Hanmi Semiconductor, Hanwha Semitech, ASMPT, and Kulicke & Soffa are developing HBF tooling, with Hanmi targeting deliveries in late 2026. Tooling commitments carry more weight than keynote slides. The roadmap, though, still targets memory samples in late 2026 and HBF-equipped inference device samples in early 2027. Sandisk excluded HBF revenue from its FY2028–2030 model entirely.
Technical Evidence Cuts Both Ways
A July preprint called FlashAccel modeled a co-designed accelerator with six HBF stacks and reported 2.54× higher inference throughput per GPU and 1.93× better energy efficiency versus an HBM-only baseline under a 100ms latency constraint.
A paper released August 12 found the opposite for different workloads: substituting HBF for SSD-backed KV-cache storage using real Qwen-Bailian serving traces produced 2–5.5× higher average latency and 1.1–2.7× lower goodput on H100/B200-class systems. Write-heavy transient KV traffic triggered thermal and endurance constraints.
The reconciliation is workload placement. HBF works for persistent, read-heavy data—model weights, cold experts, reusable context, retrieval embeddings. It fails when absorbing write-intensive transient state.
When Two Growth Curves Decouple
SK hynix positions HBF as a capacity layer between HBM and SSD, explicitly describing HBM as the highest-bandwidth tier. The probable architecture: SRAM/cache → HBM hot tier → HBF warm tier → SSD cold tier. SK hynix co-develops HBF while running one of the world's strongest HBM franchises—a structural hedge that binary narratives miss.
HBM demand remains supply-constrained. TrendForce expects HBM wafer input to rise from ~18% of DRAM wafer starts at end-2025 to ~30% in 2027. Micron has locked agreements for nearly all 2026 HBM supply and projects the market approaching $100 billion in 2028.
The disruption is subtler than replacement. If inference nodes pair 192–384 GB of HBM with 512 GB–1 TB HBF tiers, HBM gigabytes per accelerator grow far more slowly than total near-compute memory. HBF attacks content-per-system growth—arguably the most valuable variable in the HBM investment thesis. HBM revenues can keep compounding while HBM's share of near-accelerator memory declines.
The highest-conviction position for the next twelve to twenty-four months: memory hierarchy outperforms memory monoculture. Suppliers profiting across tiers, and the orchestration software deciding which bytes justify DRAM economics, capture disproportionate share regardless of how many HBM gigabytes a 2028 inference node ships. The metric to track is HBM gigabytes per accelerator measured against total near-compute memory per accelerator. When those curves separate, the profit pool has already moved.
not investment advice