SanDisk secures $93.9 billion in data-center supply agreements through 2028

The shift toward long-term, contract-based revenue models marks a fundamental change in how memory suppliers manage the capital-intensive NAND cycle. By locking in 80% margin floors with hyperscalers, SanDisk is prioritizing earnings stability over spot-market gains, a move that justifies its valuation premium against competitors like Micron. For the broader AI ecosystem, the development of High-Bandwidth Flash with SK hynix suggests that capacity constraints in high-context inference are becoming the next major infrastructure bottleneck. Industry observers should monitor whether Nvidia adopts the HBF standard, as its participation would solidify the technology's role in the next generation of AI data centers.

SanDisk's pivot toward long-term data-center contracts places it in direct competition with established NAND leaders who are pursuing their own AI-storage strategies. SK hynix, SanDisk's HBF collaboration partner, has simultaneously been expanding its dominance in high-bandwidth memory for AI training, while Kioxia and Samsung have accelerated enterprise SSD roadmaps targeting hyperscaler procurement cycles that mirror the multi-year supply agreements SanDisk has now locked in. The competitive dynamic is sharpened by Micron's parallel efforts to secure design wins in AI inference storage, where the company has positioned its 232-layer NAND as a cost-efficient alternative for data-center workloads that do not require the full bandwidth of HBM stacks.

The business model SanDisk is adopting, with pricing floors and multi-year volume commitments, reflects a broader shift in semiconductor supply chains away from spot-market exposure. Nokia's Autonomous Network Fabric partnership with AWS and Databricks, announced at DTW Ignite in June 2026, illustrates how infrastructure vendors are restructuring their commercial relationships around platform-level integration rather than transactional hardware sales. In the memory sector specifically, the trend toward contracted capacity has accelerated as hyperscalers like Alphabet and Meta seek to guarantee supply for AI training clusters that require petabytes of flash storage for checkpointing, dataset caching, and inference model loading. SanDisk's eight-customer portfolio, covering up to 67% of projected bit shipments, represents one of the most aggressive attempts to de-risk NAND revenue in the industry's history.

On the technical front, the High-Bandwidth Flash architecture that SanDisk is co-developing with SK hynix targets a specific gap between conventional SSDs and HBM in AI inference pipelines. Ericsson's AI in RAN commercial software subscription, launched in June 2026, demonstrated that embedding AI models directly into existing silicon can deliver up to 20% throughput gains without additional accelerator hardware, a philosophy that parallels SanDisk's argument that HBF can improve inference throughput without requiring operators to add expensive HBM capacity. The HBF specification aims to deliver bandwidth approaching HBM levels at a fraction of the cost per gigabyte, which would make it attractive for large-context inference workloads where model weights and key-value caches exceed available HBM capacity. Tenstorrent, one of the mentioned entities, has been an early advocate of alternative memory hierarchies for inference, and its RISC-V-based accelerator designs could benefit from a standardized HBF interface if the technology reaches commercial volumes by 2027.