The off-season for semiconductor sales can't stop the AI wave! Bernstein reveals that both storage volume and price are on the rise, adding nuclear fuel to the AI computing power bull market.

One of the major investment banks on Wall Street, Bernstein, recently released a research report stating that the semiconductor industry is experiencing a seasonal decline as anticipated by the market. However, demand for semiconductors related to AI infrastructure constructionespecially the next-generation HBM storage systems closely linked to AI infrastructure and the pricing and demand for data center server-class DRAM/NAND storage chipsremains incredibly strong. According to SIA data cited by Bernstein, July marks a traditional off-season for semiconductor sales, with global semiconductor sales decreasing by 9.5% month-on-month, slightly weaker than the historical average seasonal decline of 8.5%. Nevertheless, year-on-year, the sales still grew by 131.4%. Bernstein noted that sales of storage chips skyrocketed by 451.7% year-on-year, and when excluding storage, global semiconductor sales grew by about 35% year-on-year. It's particularly noteworthy that sales of storage chips decreased by 16.3% month-on-month compared to the record sales in June; however, this was significantly better than the historical average seasonal drop of 26.1% for July, indicating that the seasonal performance of the storage sector is considerably stronger than the markets consensus expectations. Indicators for July 2026 further showed that DRAM sales increased by 427.8% year-on-year, with bit shipments achieving a year-on-year growth of 47.5%, while NAND sales also saw a year-on-year growth of 427.8%. The average selling price per bit for DRAM increased by 257.9% year-on-year, and for NAND, it rose by 344.1%. These figures related to storage chip demand and semiconductor sales data suggest that, despite a seasonal decrease in shipments during the quarter, the average selling price per bit for DRAM and NAND continues to rise. A year-on-year growth of over 40% in bits provides evidence of expanded physical demand, while the larger increase in average selling prices strengthens the overall revenue elasticity for the major storage chip manufacturers and NAND storage giants. A deeper change is that storage is becoming the main source of revenue growth for the semiconductor industry. The report indicates that global semiconductor sales totaled approximately $861 billion in the first seven months of this year, up from $408 billion in the same period last year, marking a year-on-year growth of about 111%. Storage contributed roughly $355 billion in new sales. Of this, the contribution from changes in storage prices and product mix was approximately $306 billion, accounting for about 68% of the industrys new sales, which the report describes as close to 70%. OpenAI's newly launched Astra large model continues to actively expand the professional tasks that AI can undertake. NVIDIA CEO Jensen Huang stated on social media on Sunday that the advent of GPT-6 Astra signifies that AGI has arrived. NVIDIA has already confirmed a robust revenue range and strong shipment guidance moving forward. Additionally, the ongoing development of large AI models is entering a new phase called "Recursive Self-Improvement (RSI)," which opens up another wave of rising demand for AI computing powersuggesting that Astra is likely to expand demand for AI computing power on the commercial application side, and the trajectory of AI beginning to "create AI" may lead to increased investments in cutting-edge operator experiments, evaluations, and long-term continuous training, thereby extending the investment cycle for computing power. Astra and the automation of AI training operator research (i.e., Astra and RSI) provide a new source of semiconductor demand driven by this unprecedented storage chip frenzy within the semiconductor boom cycle. The investment significance of Astra and recursive self-improvement (RSI) lies in the fact that cutting-edge high-performance AI models and AI research itself are becoming persistent consumers of computing power in new scenarios. The launch of Astra has heightened market expectations for AGI progress, with observable advancements in completing more complex workflows: The Legora case published by OpenAI demonstrated that Astra verified 41 documents in a single agent run, with the benchmark performance of this financial statement workflow improving nearly 40% over previous generations. At the same time, OpenAI disclosed on September 6 that it had reached a new operational stage called Automated Research Intern; as of mid-August, the research team required about 3.1 smart agent workdays for every one human workday invested. This latest model indicates that "AI creating AI" (i.e., RSI training paradigm) is itself becoming a sustained consumer of reasoning, training, and evaluation resources, thereby adding a strong demand curve beyond external commercial applications. In terms of global capital pricing, the Korean stock market and the U.S. semiconductor sector have shown positive signals of renewed bullish sentiment toward storage chips and the entire semiconductor market. On September 7, Samsung Electronics rose by 5.68%, SK Hynix rose by 8.26%, and the benchmark KOSPI indexknown as the "barometer" for AI computing powerrose sharply by 4.61% to 6995.39 points, rebounding approximately 25.06% from 5593.56 points on July 30, surpassing the typically used technical bull market threshold. Astra ignites the AGI wave + AI begins to engage in AI research and development, ushering in a dual wave of demand for semiconductors. Astra represents the mechanism of expanding demand for cutting-edge performance: the capabilities of large models allow tasks that were previously difficult to complete reliably to become commercialized. Additionally, Astra may shift the overall demand curve outwardwhen AI large models become smarter, enterprises can attempt to undertake tasks that were previously impossible to execute reliably, and competitors will also need to continue investing in research and training, thus providing new strong support for the AI spending cycle. The GPT-6 Astra large model launched by OpenAI, along with the RSI technology path focused on by leading AI companies, is expected to drive exponential growth in demand for AI computing powerspecifically through the core drives of more powerful AI large models and broader utilization of AI application tools, as well as the next generation of AI training paths with significantly stronger computing power demands, which are enhancing the ongoing growth in demand for AI infrastructure. OpenAI disclosed that Astra achieved a score of 98% in the FrontierMath Level 4 test and a score of 99.9% in ARC-AGI-3. Huang, based on this, expressed the judgment that "AGI has arrived," stating that the model training utilized over 100,000 NVIDIA GPUs, with an additional 400,000 GPUs set to come online. Notably, the assertion that "AGI has arrived" remains a contentious judgment, yet the forecast of deploying larger-scale NVIDIA AI GPU clusters directly strengthens the expectation for ongoing strong demand for computing power resources for training in cutting-edge AI large models. From a technical standpoint, storage benefits from changes in model operation. Training requires the preservation of model weights, activation values, gradients, and optimizer states; long context inference and parallel agents expand KV cache and working state demands; automated research increases experiments, evaluations, training checkpoints, and data reads and writes. These tasks consume GPU-side HBM, server DRAM, and enterprise-level SSDs respectively. Storage demand depends on parameter scale, context length, concurrency count, and experimental density. HBM handles model weights, intermediate training states, and active key-value caches on the GPU side; server DRAM supports data processing, running environments, and offloading cache; NAND enterprise SSDs store datasets, training checkpoints, and reusable caches. As stronger models handle longer tasks and more agents run concurrently, while the RSI research flow increases parallel experiments and checkpoint saving, capacity, bandwidth, and read/write throughput demands will simultaneously expand significantly. NVIDIA, dubbed the "super master of AI chips," has already implemented this strong storage demand at the system architecture level for servers. The Rubin platform features up to 288GB HBM4 per GPU, a maximum memory bandwidth of 22TB/s, and introduces a flash-based shared context storage layer to accommodate reusable KV cache. It can be inferred that an increase in model capabilities leads to more concurrent tasks, longer runtimes, and denser experiments, thus expanding storage demand across all levels; the evaluation of AI economics will also further shift towards the total cost of each successful task rather than just the price per million tokens. The storage chip components in AI data center server clusters remain the clearest supply bottleneck in the AI computing power industrial chain. Market research firm TrendForce estimates that by 2026, server DRAM contract prices could rise approximately 270%, while enterprise-level SSD prices could see an increase of about 235%; by 2027, HBM contract prices could still rise by 70% to 140%. These figures reflect the combined effect of AI computing power expansion and rising storage prices. TrendForce's latest estimates predict that by 2026, the combined share of DRAM and NAND in the capital expenditures of major cloud service providers will rise from 47% to 68% by 2027, driven by increases in procurement volumes and prices. Bernstein is extremely optimistic about the trajectory of the storage super bull market! Forecasting that SanDisk will soar to $3,000, and NVIDIA will sprint to $400. The shock of the storage chip super cycle can be directly seen from the market share and growth contributions reported by WSTS as cited by Bernstein. According to WSTSs spring forecast, the storage market is projected to grow from $230.04 billion in 2025 to $803.94 billion in 2026, reflecting a year-on-year growth of 249.5%. By 2027, this is expected to further reach $1,062.085 billion, marking a year-on-year growth of 32.1%. Consequently, the proportion of storage in global semiconductor sales is expected to rise from 28.9% to 53.2%, and then to 55.5%; its contributions to new sales in the industry for 2026 and 2027 are projected to be approximately 80.2% and 64.1%, respectively. This means that the preliminary forecast scale for storage chips as a single category for 2026 is already quite astonishingslightly higher than the entire semiconductor market in 2025. The data from WSTS aligns with Bernstein's bullish logic regarding storageexpanding demand for AI computing power provides a sales foundation, limited supply enhances pricing power, and product upgrades along with long-term agreements improve profit structure. AMD, the most powerful competitor to NVIDIA's GPU ecosystem, reiterated in a Citigroup tech conference on September 8 that by 2030, the market size related to AI data center accelerated computing has expanded to $2 trillion. It also pointed out that AI inference demand has become the primary growth source for AI computing resource demand, pulling along both GPUs and server CPUs with AI agents focusing on agent-based AI workflows. According to AMD at the conference, three Helios core customersMeta and two other AI laboratorieshave projected future procurement needs surpassing expectations set when their strategic collaborations were initially established. The company expects server CPU business growth to exceed 80% year-on-year in the second half of this year and over 70% next year. This latest expectation set undoubtedly strongly supports the ongoing expansion of computing power demand; however, it should not be wholly interpreted as irrevocable orders for computing infrastructure already placed, as it is concentrating on significantly upward revised customer demand expectations. Helios is AMD's rack-level AI computing system, with Facebook's parent company Meta being one of the core clients purchasing this system. The upcoming long-term power procurement by Anthropic further increases visibility for future AI infrastructure demands surrounding storage chips. According to media reports, Anthropic has signed approximately $35 billion in cloud computing agreements with Lambda and secured about $45 billion in a six-year computational leasing deal with Nscale, totaling around $80 billion across both agreements. These figures reflect cross-period contract amounts, but their overall direction is quite clear: cutting-edge laboratories are preemptively locking in needed infrastructures for future training and inference. The WSTS data from July provides evidence of already occurring volume and price growth, while capacity procurement and model advancements from August to September enhance the judgment of subsequent demand continuity. In this report, Bernstein maintains its "outperform" rating for Samsung Electronics, SK Hynix, Micron, and SanDisk, which have seen remarkable increases since the start of the year, with target prices of 440,000 KRW, 3.3 million KRW, $1,300, and $3,000, respectively, reflecting its positive judgment on the storage market. Bernstein indicates that the latest data on volume and price collected from research and downstream procurement supports that the advantages of storage leaders are sustained by bit growth, high-value product upgrades, and pricing power; however, rising storage prices will also increase the material costs for GPU and server manufacturers, so industry profits will not grow uniformly. Bernstein commented that rising prices for ordinary DRAM have widened the profit margin between it and HBM, prompting negotiations for next years HBM contract prices. Market profit forecasts still have room for upward adjustment. However, Bernstein added that the most distinctive investment indicators moving forward will be the actual selling prices, delivery volumes, and free cash flow of storage manufacturers, as well as the ability of downstream clients to maintain stronger returns based on AI computing deployments amidst higher hardware and financing costs. Beyond the storage chip giants, Bernstein's favorable semiconductor targets also encompass AI chip leaders, foundry services, semiconductor equipment, advanced packaging, and high-end semiconductor testing chains. According to Bernsteins latest target prices, NVIDIA, SK Hynix, SanDisk, and Samsung Electronics have potential upside of approximately 77.20%, 77.80%, 72.61%, and 62.66%, respectively. Bernstein's target price for NVIDIA, the company with the highest market capitalization globally, reaches as high as $400, making it one of Wall Street's most optimistic target prices.