At the highly anticipated Goldman Sachs Communacopia+ Technology Conference, SanDisk (SNDK.US), the U.S. NAND memory chip giant, clearly stated that supply growth of NAND memory chips is likely to remain severely constrained for the foreseeable future. Meanwhile, high-performance AI inference led by Astra, along with the widespread adoption of agentic AI workflows, is continuously and explosively driving up demand for AI computing capacity and NAND storage in data centers.

Goldman Sachs analysts' latest meeting notes from the Communacopia+ conference cover remarks from SanDisk CEO David Goeckeler and CFO Luis Visoso, along with their latest outlook on NAND growth prospects. The notes also encompass presentations from executives at Etsy, NXP Semiconductors, Comcast, and Block during the second day of the conference. According to the notes, as AI applications driven by large models accelerate penetration into various global sectors, AI inference is driving simultaneous changes in NAND demand structure and procurement methods. Long-term agreements enhance the predictability of revenue and profits, low capital expenditure intensity supports effective capacity expansion, and expanded share buybacks focus on delivering operational results to shareholders.

Goldman Sachs also expressed optimism about SanDisk's next-generation NAND storage technology currently under accelerated development—namely, the High-Bandwidth Flash (HBF) storage technology roadmap—as well as the long-term growth opportunities arising from the continuous tiering and expansion of Key-Value (KV) cache and its offloading to data center SSDs. The firm maintains a "Buy" rating on SanDisk with a 12-month price target of $2,200, which, relative to the September 8 closing price of $1,737.99 used in the report, implies a potential upside of approximately 26.6% for a stock that has surged 570% in 2025 and approximately 600% year-to-date.

One of the key highlights of Goldman Sachs' notes is two long-term opportunities identified by SanDisk management: first, the NAND-based HBF technology path, which expands the role of flash memory in AI inference storage systems through high-capacity and high-bandwidth designs; second, long-context, multi-turn interactions, and high-concurrency agents are driving tiered KV cache storage, allowing reusable and temporarily inactive cache to be increasingly offloaded to data center SSDs. This reduces GPU/TPU/XPU memory usage and duplicate computation costs, thereby increasing enterprise-grade NAND demand.

According to a recent research report from another Wall Street financial giant, Bernstein, with Astra and the automation of AI training operator research (i.e., Astra and RSI) providing new sources of semiconductor demand for this unprecedented memory chip boom, the firm has even set a $3,000 price target for SanDisk. Bernstein maintains "Outperform" ratings on the stellar-performing global memory chip leaders this year—namely Samsung Electronics, SK Hynix, Micron, and SanDisk—with price targets of ₩440,000, ₩3.3 million, $1,300, and $3,000, respectively, reflecting a fresh wave of positive expectations from Wall Street institutions on the memory chip boom.

Long-Term Agreements Reshape the Profit Base: SanDisk's NAND Business Shifting Gears in Growth Engines

In the meeting notes, the Goldman Sachs analyst team indicated that SanDisk is transitioning from short-term spot pricing to a long-term agreement-dominated business model. Citing SanDisk management, the analysts wrote that 50% of planned sales volume for fiscal 2027 and 67% for fiscal 2028 are already covered by NBM long-term agreements. The floor price mechanism in these agreements can support approximately 80% gross margins for most of the business in downside scenarios, significantly improving earnings visibility. Goldman Sachs projects fiscal 2027 and 2028 revenue of approximately $53.15 billion and $71.04 billion, representing year-over-year growth of about 162.5% and 33.7%, with EPS of $231.84 and $284.98, respectively.

Supply persistently trailing demand, combined with capital allocation, forms the second pillar of the long-term bullish thesis for SanDisk. Goldman Sachs noted that SanDisk management believes the comprehensive proliferation of AI applications across various sectors in the AI inference era will continue to increase demand, while NAND supply growth remains relatively limited in the foreseeable future, with new capacity from Chinese competitors largely absorbed by their domestic market. The company has extended its joint venture partnership with Kioxia through 2034 and emphasized that its proprietary intellectual property, R&D investment, and BiCS technology roadmap can support bit output growth for years to come, while maintaining a low capital expenditure intensity of approximately 5%.

Goldman Sachs indicated these factors mean SanDisk aims to increase sellable capacity through manufacturing efficiency and technology upgrades, improving output per unit of capital investment. Additionally, the company has executed approximately $4.5 billion in buybacks and continues to use repurchases as the primary method of returning excess capital, while maintaining a positive openness toward future dividends.

Nvidia's fiscal 2027 second-quarter revenue reached $96.2 billion, up 106% year-over-year, with data center revenue of $89 billion, up 117% year-over-year. The company also guided next-quarter revenue to $108 billion, plus or minus 2%, and provided a 70% growth outlook for the following fiscal year. Meanwhile, Anthropic, reportedly preparing for an IPO, has signed a computing agreement with Nscale valued at approximately $45 billion over six years, corresponding to 460 megawatts of capacity, as well as a cloud computing deal with Lambda worth approximately $35 billion. These latest signs of global AI computing demand show leading model companies securing long-term compute supply in advance to accommodate continuously growing AI application demand, driving large-scale procurement of data center infrastructure including accelerators, server memory, enterprise SSDs, high-performance networking equipment, and server CPUs.

From a technical foundation perspective, the incremental NAND demand from AI inference stems from more data retention and storage taking on more work in the inference process. Agents need to repeatedly read enterprise knowledge bases, code, documents, and multimodal materials, while preserving task states, tool outputs, and reusable context. Under typical Transformer architectures, longer contexts and higher concurrency also expand KV cache. Active computation prioritizes HBM, CPU-side DRAM handles expanded memory, while suitable reusable and temporarily inactive cache can be offloaded to enterprise SSDs to reduce recomputation costs. Therefore, opportunities for NAND memory leaders—namely SanDisk, Kioxia, Micron, and Samsung—span NAND capacity, read throughput, and product capability to adapt to specific workloads.

SanDisk management also views HBF as a long-term option to alleviate the "DRAM/HBM memory wall" through higher density, and management projects the AI data center storage capacity market to reach approximately 1.2ZB by 2032, with KV cache-related demand accounting for approximately 35%, or roughly 0.42ZB. However, these capacity forecasts from SanDisk management should not be directly equated to revenue expectations, nor do they imply that data center NAND technology can directly replace all HBM use cases.

From Answering Questions to Sustained High-Output Work: Astra Opens New Round of Compute Demand, NAND Storage Absorbs Massive Data Loads

OpenAI's GPT-6 Astra large model, along with the RSI technology path that AI leaders are focusing on, is expected to become two core drivers of exponential expansion in AI compute demand. Stronger AI models, broader use of AI application tools, and next-generation AI training paths with more robust compute requirements are strengthening the evidence base for sustained growth in AI infrastructure demand.

Astra represents the frontier performance demand expansion mechanism: as large model capabilities improve, tasks that were previously difficult to complete reliably enter the commercially viable range. Additionally, Astra may shift the entire demand curve outward—when AI models become smarter, enterprises can attempt work they previously couldn't reliably accomplish, and competitors will also need to continue investing in R&D and training. This provides new strong support for the AI spending cycle.

Over the weekend, Nvidia CEO Jensen Huang even made a significant statement on social media that the arrival of GPT-6 Astra means "AGI is here." Nvidia's confirmed strong revenue ranges and subsequent robust shipment guidance, combined with AI model R&D entering a new phase of "Recursive Self-Improvement (RSI)"—opening yet another curve of surging AI compute demand—suggests Astra may expand commercial application-side AI compute demand, while the R&D trajectory of AI "creating AI" may increase frontier operator experimentation, evaluation, and sustained long-term training investment. Together, these extend the compute investment cycle.

Astra's significance to the demand curve lies in improving the completion rate of complex workflows, making tasks that were previously not worth automating begin to hold commercial value. In OpenAI's latest OSWorld 2.0 testing, Astra scored 72.6%, higher than GPT-5.6 Sol's 65.7%. Its capabilities cover long-horizon tasks such as computer operations, software engineering, and scientific research. From this, it can be deduced that global frontier technology research organizations and internet IT enterprises may deploy more parallel AI agent workflows, executing longer-duration tasks involving more materials, thereby expanding inference calls, cache reuse, and persistent storage demand. This aligns with Morgan Stanley's logic of shifting focus to supply constraints in compute, power, and materials.

OpenAI's product lead's suggestion of "possibly pausing new Pro subscriptions" also reflects the capacity pressure on short-to-medium-term AI compute and data center high-performance storage services amid explosive demand. The research agents surrounding Astra further demonstrate the trajectory of AI compute demand entering a new round of exponential expansion. Multi-agent research continuously generates and reuses context, code, experimental results, and checkpoints, highlighting that high-performance SSDs, potential HBF, and research agents may provide longer-term growth space for NAND leaders like SanDisk.

OpenAI disclosed that its Navier-Stokes research employed approximately 10,000 concurrent agents, producing a solution in about 88 hours. All attempted research tasks generated approximately 300 billion output tokens, with the Navier-Stokes portion accounting for about 130 billion. The internal model used for problem-solving was stronger than Astra, which separately took approximately 17 hours to complete Lean formal verification.

The global semiconductor sector in equity markets has experienced a process of "mad selling during the July deleveraging frenzy, sentiment repair in August, and a new bull market catalyzed by September model developments." The Philadelphia Semiconductor Index fell nearly 29% from June to July, then rebounded approximately 20% from the July 29 low by the August 13 intraday session, touching technical bull market territory. South Korea's KOSPI rebounded from approximately 5,593 points on July 30 to 7,129.34 intraday on September 8, recovering about 27.5%. On September 7, Samsung Electronics and SK Hynix rose 5.7% and 8.1% respectively, showing memory leaders remain a key driver in the Korean market. The Astra release has added new demand expectations to existing earnings and order support—market capital is beginning to reassess just how large the compute and storage services enterprises are willing to purchase can become as models prove capable of completing more complex work.