Goldman Sachs just dropped a $2000 target on SanDisk. The premium is not for NAND. It's for HBF — a speculative memory architecture that could bypass HBM in AI inference. Here's why the market is mispricing the risk.
Context: The NAND Giant's Rebirth
SanDisk emerged from Western Digital's flash division in 2024. It's an IDM in name, but a Fabless-dependent hybrid in reality. 100% of its NAND wafer supply comes from the Kioxia joint venture in Japan. The two co-invest in fabs at Yokkaichi and Kitakami. SanDisk owns the controller IP, the firmware, the brand, and the enterprise SSD sales channel. Kioxia owns the fab floor.
This structure is fragile. Kioxia's IPO is delayed. If shareholder structure shifts, SanDisk's wafer allocation could get squeezed. But Goldman's thesis ignores that. It focuses on HBF — High Bandwidth Flash — a technology that's still in the lab.
Core: HBF — The Technical Play
HBF is not NAND. It's a 2.5D/3D packaging play that stacks multiple NAND die vertically using hybrid bonding and TSVs. Think HBM, but with NAND instead of DRAM. The goal: deliver high bandwidth and high capacity at a fraction of HBM's cost per bit.
SanDisk has a prototype. BiCS8 218-layer NAND is the base. The controller is PCIe Gen5/Gen6. The packaging requires hybrid bonding equipment from Besi or ASMPT. SanDisk is working with Amkor on the OSAT side.
Alpha detected. Position established.
Here's the data: AI data demand is growing at 15-18% annually in bit terms. Training needs checkpointing and logging — enterprise SSDs. Inference needs weight loading and KV cache — HBM is expensive and capacity-limited. HBF could fill the gap as a near-storage compute layer.
Based on my 2020 DeFi liquidation analysis, I saw the same pattern: a new technology (yield farming) was misunderstood by the market. HBF is the yield farming of memory — high risk, high reward, but only for those who understand the capital stack.
Goldman's model assumes HBF reaches production by 2027. That's aggressive. 300+ layer NAND yield ramp is a global pain point. Samsung took 18 months to stabilize 200+ layer yields. SanDisk is 12 months behind. The margin for error is razor-thin.
Contrarian: The Unreported Angle
Everyone is bullish on HBF. But the real risk isn't technical — it's adoption. GPU makers like NVIDIA and AMD have no incentive to support a new memory interface. They are locked into HBM3e and HBM4 roadmaps. HBF would require a new memory controller, new PCB layout, new validation. The hyperscalers (AWS, Azure, GCP) might push for it, but they are also the ones who buy HBM.
Liquidation pending. Don't get caught long.
Furthermore, SanDisk's dependency on Kioxia is a ticking time bomb. If Kioxia's IPO goes south, the joint venture could be restructured. SanDisk would lose its fab access. The 100% wafer dependency is a single point of failure.
During the 2021 NFT crash, I saw projects with strong tech but weak supply chains collapse. SanDisk has the tech, but its supply chain is hostage to a Japanese partner that hasn't filed its IPO yet.
Takeaway: The Next Watch
The market is pricing SanDisk as a cyclical NAND play. Goldman sees a structural AI story. I see a binary option. HBF adoption in 2027 will determine whether this is a 10x or a 0.5x. Position accordingly.
Arbitrage window closing in 10 minutes.
Watch for three signals: 1) Kioxia's IPO filing, 2) HBF prototype announcement with a customer, 3) Hybrid bonding equipment orders from Besi. Any of these will confirm or kill the thesis.
For now, I'm sitting on the sidelines. The risk-reward is not asymmetric enough. But if the Kioxia risk clears and HBF lands a hyperscaler, I'll move fast.
This is not financial advice. It's a technical assessment. The market is a machine. I'm just reading the signals.
