The off-season semiconductor sales cannot stop the AI heatwave! Bernstein reveals that storage volume and price have risen sharply, and the AI computing power bull market has added more nuclear fuel

The Zhitong Finance App learned that Bernstein, one of the largest investment banks on Wall Street, recently released a research report saying that although the semiconductor industry is showing a seasonal decline as expected by the market, demand for semiconductors related to AI computing power infrastructure construction — in particular, next-generation HBM storage systems closely linked to AI infrastructure and data center server-level DRAM/NAND memory chips is still extremely strong. According to SIA data quoted by Bernstein, July is a low season for traditional semiconductor sales. Global semiconductor sales fell 9.5% month-on-month, slightly weaker than the seasonal average decline of 8.5% in the same period in history, but still increased by 131.4% year on year.

Bernstein said that sales of memory chips increased by a whopping 451.7% year over year, and after excluding storage, the global semiconductor industry's sales increased by about 35% year over year. It is particularly noteworthy that memory chip sales fell 16.3% month-on-month compared to record sales in June, but the average storage sales volume in July fell sharply by 26.1%, which is clearly better than the data for the same period in history, which is enough to show that the seasonal performance of the storage sector is much stronger than the market's unanimous expectations.

The July 2026 index also shows that DRAM sales increased 427.8% year on year, bit shipments achieved a 47.5% year-on-year increase, and NAND sales achieved a 427.8% year-on-year increase; the average sales price of DRAM per bit increased 257.9% year on year, and the average price of NAND per bit increased 344.1% year on year. These data on demand for memory chips and semiconductor sales all mean that despite a month-on-month decline in shipments during the quarterly switch, the average price per bit of DRAM and NAND continues to rise. A bit increase of more than 40% over the previous year provided evidence of an expansion in physical demand, and a larger increase in average sales prices strengthened the overall revenue elasticity of the three major memory chip manufacturers and NAND memory chip giants.

Further, the change is that storage is becoming a major source of semiconductor revenue growth. According to the report, global semiconductor sales in the first seven months of this year were about 861 billion US dollars, up from 408 billion US dollars in the same period last year, an increase of about 111% over the previous year; storage contributed about 351 billion US dollars in new sales. Among them, the contribution of changes in storage prices and product portfolios included in the report was about US$306 billion, which is equivalent to about 68% of the new sales of the entire industry, or “close to 70%” as described in the report.

The Astra model just launched by OpenAI continues to actively expand the professional tasks that AI can undertake. On Sunday, Nvidia CEO Huang Renxun made a big statement on social media that the launch of GPT-6 Astra means “AGI has arrived,” and Nvidia's confirmed strong revenue range and subsequent strong shipping guidelines, combined with the strong AI model development is entering a new stage of “recursive self-improvement (RSI)”, which opens up another stage of the AI computing power demand surge curve — that is, Astra is expected to expand AI computing power demand for commercial applications. Beginning the R&D trajectory of “building AI” may increase investment in cutting-edge operator experiments, evaluations, and long-term continuous training, and jointly extend the computing power investment cycle.

Astra and AI training operator research automation (i.e. Astra and RSI) can be described as providing a new source of semiconductor demand for this unprecedented semiconductor boom cycle driven by the memory chip frenzy. The investment significance of Astra and recursive self-improvement (RSI) is that cutting-edge high-performance AI models, and AI research and development itself are becoming new scenarios that continue to consume computing power.

The Astra release strengthened the market's expectations for the AGI process, and its observable progress was the ability to complete more complex workflows: the Legora case published by OpenAI showed that Astra verified 41 documents during an intelligent operation, and the benchmark performance of this financial reporting workflow increased by nearly 40% compared to the previous generation. At the same time, OpenAI revealed on September 6 that it had reached a new stage of operation for “automation research interns”; as of mid-August, the research team spent every manual work day corresponding to the use of about 3.1 smart devices. This latest model means that “AI R&D AI” (that is, the RSI training paradigm) itself is also becoming a customer that continues to consume inference, 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 US semiconductor sector have shown specific positive signs that capital is once again excited and bullish about memory chips and semiconductors as a whole. On September 7, Samsung Electronics rose 5.68%, and SK Hynix rose 8.26%. The Korean stock market benchmark index, the KOSPI Index, which has the title of “AI computing power weather vane,” rose 4.61% to 6995.39 points on the same day, rebounding a cumulative rebound of about 25.06% from 5593.56 points on July 30, exceeding the commonly used technical bull market threshold.

Astra set off an AGI frenzy and AI began participating in AI research and development, and the semiconductor sector ushered in a wave of double demand

Astra represents the most advanced performance demand expansion mechanism: the ability of large models has been increased, making tasks that were previously difficult to complete reliably into the scope of commercialization. Furthermore, Astra may shift the overall demand curve — that is, when the AI model is smarter, companies can try jobs that could not be done reliably before, and competitors also need to continue to invest in R&D and training, which provides new strong support for the AI spending cycle.

The GPT-6 Astra model launched by OpenAI and the RSI technology path focused on by AI leaders are expected to become the two core driving forces driving the exponential expansion of AI computing power demand, namely the AI big model with better performance, the use of a wider range of AI application tools, and a next-generation AI training path with stronger computing power requirements, which are an important basis for the continued growth in AI computing power infrastructure demand.

OpenAI revealed that Astra achieved 98% in the FrontierMath Level 4 test and 99.9% in ARC-AGI-3. Based on this, Hwang In-hoon expressed the judgment that “AGI has arrived” and stated that the model used more than 100,000 Nvidia GPUs for training, and 400,000 GPUs will soon be launched. It is worth noting that “AGI has arrived” is still a controversial judgment, and the announcement of the deployment of the larger Nvidia AI GPU cluster directly reinforces the strong AI computing power demand expectations that the large front-end AI model will continue to expand investment in training resources.

At the bottom of the technology, storage benefits from changes in how models operate. Training requires preserving model weights, activation values, gradients, and optimizer states; long context reasoning and parallel agents expand KV cache and operating state requirements; and automated research increases experiments, evaluation, training checkpoints, and data reading and writing. These tasks consume GPU-side HBM, server DRAM, and enterprise-grade SSDs, respectively.

Storage requirements depend on parameter size, context length, number of concurrency, and experiment density. HBM is responsible for model weight, training intermediate state, and active key value caching on the GPU side; server DRAM undertakes data processing, operating environment and cache offloading; NAND enterprise-grade SSD stores data sets, training checkpoints, and reusable caches. As stronger models handle longer tasks, more agents run simultaneously, and the RSI research process adds parallel experiments and checkpoint storage, the demand for capacity, bandwidth, and read/write throughput will expand dramatically at the same time.

Nvidia, the “superpower of AI chips,” has implemented this strong storage requirement into a system-level server architecture. The Rubin platform is configured with a maximum of 288GB HBM4 and a maximum memory bandwidth of 22Tb/s per GPU, and introduces a shared context storage layer based on flash memory to handle reusable KV caches. As deduced from this, if improved model capabilities drive more concurrent tasks, longer run times, and more intensive experiments, storage requirements will expand along the entire hierarchy; the focus on measuring AI economics will also further shift to the total cost of each successful task, not just the price per million tokens.

The memory chip components of AI data center server clusters are still the clearest supply bottleneck at the AI computing power industry chain level. Market research agency TrendForce predicts that in 2026, server DRAM contract prices will increase by about 270%, and enterprise-grade SSD prices will increase cumulatively by about 235%; HBM contract prices may still rise 70% to 140% in 2027. These data show the combined effects of AI computing power expansion and storage price increases. According to TrendForce's latest estimates, the combined share of DRAM and NAND in capital expenditure of major cloud service providers will rise from 47% in 2026 to 68% in 2027, behind which there is a simultaneous increase in procurement volume and price increases.

Bernstein is extremely optimistic about storing the super bull market trajectory! Shout that SanDisk skyrocketed to $3,000 and Nvidia rushed to $400

The shocking aspect of the memory chip supercycle can be seen directly from the WSTS report quoted by Bernstein as shown in the market share and growth contribution. According to WSTS's official spring forecast, the storage market will grow from US$230.42 billion in 2025 to US$803.941 billion in 2026, an increase of 249.5% year-on-year; in 2027, it will further reach US$1062,085 billion, an increase of 32.1% year-on-year. According to this calculation, the share of storage in global semiconductor sales will rise from 28.9% to 53.2% and then to 55.5%; its contribution to the new sales volume of the entire industry in 2026 and 2027 will be about 80.2% and 64.1%, respectively.

In other words, the initial forecast scale for a single category of memory chips in 2026 is already astonishing — slightly higher than the entire semiconductor market in 2025. WSTS data echoes Bernstein's logic of looking more at storage — the expansion of AI computing power demand provides a sales base, limited supply strengthens pricing capabilities, and product upgrades and long-term agreements improve profit structures.

AMD, the strongest competitor of Nvidia's GPU system, reiterated at the Cititech conference on September 8, local time, that the market size related to accelerated computing in AI data centers had expanded to 2 trillion US dollars by 2030, and pointed out that AI inference demand has become the main source of growth in AI computing power resource requirements. AI agents focusing on proxy AI workflows simultaneously drive GPUs and server CPUs.

AMD said at the conference that Meta and two other AI labs, the three Helios core customers, all gave higher future procurement demand forecasts than when the two sides first established strategic cooperation; the company expects the server CPU business to grow by more than 80% year-on-year in the second half of this year and 70% next year. This latest combination of expectations undoubtedly strongly supports the continued expansion of computing power demand, but the sharp increase in customer demand expectations cannot all be regarded as irrevocable computing power infrastructure orders that have already been placed. Helios is AMD's rack-scale AI computing system, and Facebook's parent company Meta is one of the core customers that purchased the system.

Anthropic's long-term computing power procurement, which is about to go public, has further increased the visibility of future AI computing power infrastructure requirements surrounding memory chips. According to media reports, Anthropic signed a cloud computing agreement of about 35 billion US dollars with Lambda and reached a computing power lease transaction of about 45 billion US dollars with Nscale for a period of 6 years. The two agreements total about 80 billion US dollars. These are cross-term contract amounts, but the common direction is clear: Frontier Labs are locking down the infrastructure needed for future training and reasoning ahead of time. The July WSTS data provides evidence of the volume and price increases that have occurred, and capacity procurement and model progress from August to September strengthened the judgment on the continuity of subsequent demand.

In this report, Bernstein maintained the “outperforming the market” ratings for Samsung Electronics, SK Hynix, Micron, and SanDisk, which have been rising rapidly since this year. The target prices are 440,000 won, 3.3 million won, 1,300 US dollars, and 3,000 US dollars, respectively, reflecting his positive judgment on the storage boom. Bernstein said that in combination with the volume and price data shown in the latest research and downstream procurement, the advantage of storage leaders is that bit growth, high-value product upgrades, and pricing capabilities all support profits; however, storage price increases will also increase material costs for GPU and server manufacturers, and industry profits will not grow evenly.

Bernstein said that the increase in the price of ordinary DRAM widens the wafer profit gap between it and HBM and pushes the HBM contract price renegotiation for the next year. There is still room for an increase in market profit forecasts. However, Bernstein added that the most investment-differentiated indicators in the future are the actual sales price, delivery volume, and free cash flow of storage vendors, and whether downstream customers can maintain stronger returns on AI-based computing power deployments under higher hardware and financing costs.

In addition to memory chip giants, Bernstein's optimistic semiconductor targets also cover AI chip leaders, foundry and semiconductor equipment, advanced packaging, and high-end semiconductor test chains. According to Bernstein's latest target price, the target prices for Nvidia, SK Hynix, SanDisk, and Samsung Electronics correspond to an increase of about 77.20%, 77.80%, 72.61%, and 62.66%, respectively. Bernstein's target price for Nvidia, which has the highest market capitalization in the world, is as high as $400, making it one of Wall Street's most optimistic targets.