The more impressive AI capabilities become, the more NAND storage needs to be expanded! Goldman Sachs analyzes SanDisk (SNDK.US) "Long-term Agreement + Massive AI Inference" Bull Market Logic

According to Zhitong Finance APP, at the highly anticipated Goldman Sachs Communacopia+ Technology Conference, leading U.S. NAND memory chip giant SanDisk (SNDK.US) made it clear that NAND supply growth will likely remain severely constrained for the foreseeable future. Meanwhile, the explosion in demand driven by Astra-led high-performance AI inference and the widespread adoption of agent-based AI workflow technologies is propelling AI compute capacity and datacenter NAND storage needs to new heights.

The latest meeting memo from Goldman Sachs analysts covers remarks made by SanDisk CEO David Goeckeler and CFO Luis Visoso, as well as their latest forecasts on the future growth outlook for NAND. It also covers speeches from key technology executives from major companies such as Etsy, NXP Semiconductors, Comcast, and Block on the conference’s second day. The memo shows that as AI large models drive the rapid penetration of AI applications across global sectors, AI inference is simultaneously transforming NAND demand structure and procurement models. Long-term agreements are improving the predictability of revenue and profits, low capital expenditure intensity supports efficient capacity expansion, and share repurchases focus on delivering operating results to shareholders.

Goldman Sachs is also optimistic on SanDisk’s next-gen NAND storage technology under accelerated development—namely High Bandwidth Flash (HBF) memory, and the long-term growth opportunities presented by the continuous layering and offloading of Key Value Cache (KV Cache) onto datacenter SSDs. The firm maintains a “Buy” rating on SanDisk and a 12-month price target of $2,200, implying around 26.6% upside given the closing price of $1,737.99 on September 8, after the stock surged 570% in 2025 and has already risen 600% year-to-date.

The Goldman Sachs memo highlights two long-term opportunities presented by SanDisk's management: First is the HBF technology route based on NAND, which expands flash memory’s role within AI inference storage systems through high capacity and bandwidth design. Second, long-context, multi-round interaction, and high-concurrency intelligent agents drive the layered storage of KV cache, allowing more reusable and inactive cache to be offloaded onto datacenter SSDs, reducing enterprise AI GPU/TPU/XPU memory usage and redundant computation costs, thus increasing demand for enterprise-grade NAND.

Another Wall Street financial giant, Bernstein, recently released a research report indicating that with Astra and AI training operator research automation (i.e., Astra & RSI) providing new sources of semiconductor demand amid this unprecedented NAND chip boom, it is calling for a $3,000 price target for SanDisk. Bernstein maintains an “Outperform” rating on global memory chip leaders—Samsung Electronics, SK Hynix, Micron, and SanDisk—with targets of 440,000 KRW, 3.3 million KRW, $1,300, and $3,000, reflecting Wall Street institutions’ renewed optimistic outlook for the memory chip cycle.

Long-term Agreements Reshape Profit Base: SanDisk’s Long-Term NAND Business Shifts Growth Engine

The Goldman Sachs analyst team stated in the memo that SanDisk is shifting from spot pricing to a business model dominated by long-term contracts. Quoting SanDisk management, the analysts cite plans for 50% and 67% of FY2027 and FY2028 planned shipments to be covered by NBM long-term agreements, with floor-price mechanisms supporting around 80% gross margin across most operations even in downturns, vastly improving earnings visibility. Goldman projects company revenue for FY2027 and FY2028 at approximately $53.15 billion and $71.04 billion, up 162.5% and 33.7% year-over-year, with EPS of $231.84 and $284.98, respectively.

Persistently lagging supply growth and capital allocation form the second leg of SanDisk’s long-term bull thesis. Goldman notes that SanDisk management believes the proliferation of AI inference and applications across industries will continue to drive up demand, while NAND supply growth is still limited for the foreseeable future, with new capacity from Chinese competitors mainly absorbed by the domestic market.

The company has extended its joint venture with Kioxia through 2034, emphasizing that its proprietary IP, R&D investment, and BiCS technology roadmap will support bit output growth for years to come, while maintaining a low capital expenditure intensity of around 5%. Goldman notes this signifies SanDisk aims to boost sellable capacity through manufacturing efficiency and technology upgrades, raising output per capital unit invested. The company has already executed around $4.5 billion in share buybacks, remains focused on buybacks as the primary way to return excess capital, and is open to dividends in the future.

Nvidia’s Q2 fiscal 2027 revenue reached $96.2 billion, up 106% year-on-year; datacenter revenue was $89 billion, up 117%, with guidance for Q3 revenue to reach $108 billion, plus or minus 2%. For next fiscal year, Nvidia projects 70% growth. Meanwhile, pre-IPO Anthropic reportedly signed a $45 billion, six-year compute deal with Nscale for 460MW capacity, and an approximately $35 billion cloud computing deal with Lambda. These latest signs of surging global AI compute demand show that leading model companies are locking in long-term compute supply early to meet growing AI application demand, spurring large-scale procurements of AI accelerator clusters, server memory, enterprise SSDs, high-performance network devices, and server CPUs for datacenter infrastructure.

From a technology standpoint, incremental NAND demand from AI inference comes from more data retention and greater storage workloads in the inference pipeline. Agents need to repeatedly read enterprise knowledge bases, code, documents, and multimodal materials, and save task status, tool outputs, and reusable context; in the typical Transformer architecture, longer context windows and higher concurrency amplify KV cache. Active computation is prioritized in HBM, DRAM on the CPU side extends memory, while reusable and temporarily inactive cache can be offloaded to enterprise SSDs to reduce recomputation costs.

As such, leading NAND players—SanDisk, Kioxia, Micron, and Samsung—stand to benefit through capacities, read throughput, and workload-adaptive products. SanDisk management also sees HBF as a long-term option to alleviate the “DRAM/HBM memory wall” with higher densities, estimating that AI datacenter storage capacity could reach around 1.2ZB by 2032, with KV cache-related demand accounting for roughly 35%, or 0.42ZB. However, these are capacity forecasts from SanDisk management, not the company’s revenue expectations, nor does it mean datacenter NAND will directly replace all HBM use cases.

From Answering Questions to Sustained High Output: Astra Unlocks New Compute Demand, NAND Absorbs Massive Data Loads

The GPT-6 Astra large model launched by OpenAI and the RSI technology path focused on by AI leaders are expected to be the two main drivers of exponential growth in AI compute demand—i.e., stronger performing large AI models, more widespread use of AI application tools, and more compute-intensive next-gen AI training paths are boosting persistent demand for AI compute infrastructure.

Astra represents the leading edge of performance demand expansion: improved large model capabilities move previously unreliable tasks into commercially viable territory. In addition, Astra could shift the entire demand curve outward—i.e., as AI models become smarter, enterprises can attempt tasks previously off limits; competitors must continue investing in R&D and training, providing powerful support for new investment cycles in AI.

Nvidia CEO Jensen Huang went so far as to announce on social media that the release of GPT-6 Astra means “AGI is here.” Nvidia’s strong confirmed revenue range and solid shipment guidance, coupled with AI model development stepping into the “Recursive Self-Improvement (RSI)” phase—representing another surge in compute demand growth—suggest Astra could further expand commercial AI compute demand, and the trajectory of “AI building AI” may intensify upfront operator experimentation, evaluation, and sustained training, prolonging the compute investment cycle.

The significance of Astra for the demand curve lies in raising the completion rate for complex workflows, making previously non-commercializable tasks economically viable. In recent OpenAI OSWorld 2.0 tests, Astra scored 72.6%, above GPT-5.6 Sol’s 65.7%. Its capabilities span long-term tasks in computer operation, software engineering, and scientific research.

This implies that leading-edge research or global IT/web firms may deploy more parallel AI agent workflows, assign them longer, more data-intensive work, and expand inference, cache reuse, and persistent storage needs. This aligns with Morgan Stanley’s focus shift to supply constraints in compute, power, and materials. OpenAI’s product head mentioning a “potential pause on new Pro subscriptions” also reflects short- and medium-term pressure on AI compute and high-performance datacenter storage driven by surging demand.

Research agents built around Astra further showcase the trajectory of compute demand entering a new exponential expansion phase. Large numbers of agents continually generate and reuse context, code, experimental results, and checkpoints, highlighting the long-term growth space for high-performance SSDs, potential HBFs, and research agents benefiting NAND leaders like SanDisk. OpenAI disclosed that its Navier-Stokes study used about 10,000 concurrent agents, delivering a solution in 88 hours; total research tasks produced about 300 billion output tokens, with around 130 billion from the Navier-Stokes section. The models used were stronger than Astra, while Astra accomplished Lean-format verification in about 17 hours.

Global equity market semiconductor sectors have witnessed a “July deleveraging sell-off, August sentiment recovery, and September model advancements driving a new bull market.” The Philadelphia Semiconductor Index dropped nearly 29% from June to July, but rebounded about 20% from the July 29 low to August 13, hitting the technical bull market threshold. South Korea’s KOSPI rose around 27.5% from 5,593 on July 30 to 7,129.34 on September 8. On September 7, Samsung Electronics and SK Hynix surged 5.7% and 8.1% respectively, showing memory leaders remain a bullish force in Korea’s market. Astra’s release adds new demand drivers on top of robust profit and order backing—capital is reconsidering how much further enterprise compute and storage spending can scale now that models can accomplish increasingly complex tasks.