SanDisk said AI data center buildouts are reshaping flash memory demand, with data centers expected to overtake edge devices as the biggest source of demand by 2026 as the flash market reaches about 1.2 zettabytes. In investor day materials released on Aug. 14, the company projected industry revenue total addressable market, or TAM (total addressable market), climbing from a historical cyclical midpoint of about $60 billion to more than $300 billion by 2027 and approaching $500 billion. The company argued that AI inference is the key driver, with long context, RAG (retrieval-augmented generation), multimodal models and agent applications increasing demand for KV cache (stored inference context), which could elevate SSDs from low-value storage hardware into a strategic layer of AI infrastructure. SanDisk described SSDs as a "Token Battery," saying stored completed computation can reduce redundant processing, lower energy use and improve throughput. To reduce traditional NAND market volatility, SanDisk said it is signing new multi-year commercial agreements with large customers. It said eight customers are already participating, with an average contract term of more than four years; about half of bit supply is expected to be covered by the model by FY2027, rising to about two-thirds by FY2028. The agreements carry total contract value of $93.9 billion at floor prices, with $91.1 billion in remaining performance obligations and $16.5 billion of financial guarantee support. SanDisk also outlined aggressive financial targets, including FY2026 revenue of $20 billion, up 175% year on year, non-GAAP gross margin of 71.6%, and adjusted free cash flow of $8.7 billion. Its long-term model for FY2028 to FY2030 calls for mid- to high-teens revenue growth, non-GAAP gross margin of about 80%, operating margin of about 75% and adjusted free cash flow margin of about 50%. The company also pointed to HBF (high-bandwidth flash) as a longer-term opportunity for large-model inference, saying it is designed to address the memory wall (limits on moving data fast enough for compute) and can offer up to 8 to 16 times the capacity of HBM at similar read bandwidth. SanDisk said the first HBF inference product samples are expected in 2027, a development it said could prompt the market to reassess its long-term upside through shifts in AI memory architecture.