As the “AI deceleration” hits semiconductors, Goldman Sachs reports a bullish report! The target price for “Korea Storage Duo” points to nearly 90% upward space

The Zhitong Finance App learned that as global AI leaders such as Anthropic and OpenAI unanimously called for a slowdown in the development of cutting-edge AI models, the global stock market also re-estimated the cooling of investment growth expectations related to the AI computing power industry chain and the risk of a new round of interest rate hikes by the Federal Reserve. AI computing-power-themed stocks in the stock market were generally weak. SK Hynix once fell more than 5% in early trading in the Korean stock market, and Samsung Electronics fell more than 3%. However, many senior Wall Street analysts said that the latest developments will not have a lasting impact on the industry and are unlikely to disrupt the long-term AI computing power trading theme bull market logic.

Wall Street financial giant Goldman Sachs's latest “Investor Feedback Report on Korea's Technology Industry” shows that North American investors are very positive about memory chip stocks, especially far more positive than Asian investors. The agency reiterated the “buy” ratings for the world's two largest memory chip giants, Samsung Electronics and SK Hynix. Samsung Electronics continues to rank on Goldman Sachs's “Conviction List” (Conviction List)

According to information, in this research report, Goldman Sachs analysts target prices as high as 490,000 won for Samsung's common stock and 360,000 won for preferred shares, and as high as 3.5 million won for SK Hynix; based on the unified calculation of the closing price on September 11, the potential share price increases are about 88.8%, 86.2%, and 93.2%, respectively. Goldman Sachs continues to expect the average sales price of HBM to rise about 100% year on year in 2027. Samsung is expected to benefit from improved product and customer portfolios, and the two companies also have potential catalysts in terms of shareholder returns.

Some analysts said that calling for a slowdown in cutting-edge model development is not directly equivalent to cutting computing power capital expenses or storage orders, and demand for inference and application of existing models may also continue to grow. “This may cause some short-term pressure, but it's unlikely to disrupt long-term AI deals. AI development is still in a relatively early stage, and I'm not sure if other players in the AI ecosystem are willing to accept the current industry rankings and slow down while technology is still evolving so fast.” Gary Tan, Allspring Global Investments' Singapore-based portfolio manager, said.

“The three CEOs agreed to control the pace and won't really change the money invested in chips, power, and infrastructure. In fact, it lengthens the development timeline.” Billy Leung, an investment strategist at Global X Management based in Sydney, said, “If commercialization and adoption continue to grow while the rate of introduction of new capabilities slows slightly, this will actually help the industry shift from spending money to accelerate profits from what has already been built — that is, AI monetization.”

For investors, the most important indicators of demand for memory chips and the sharp increase in volume and price will undoubtedly be memory chip manufacturers' actual shipments, customer certification, contract prices, profit margins and free cash flow data, and phased outlook ranges; in the end, the optimistic target prices of Wall Street institutions such as Goldman Sachs still need to be fulfilled.

Price increases are expected to cool down compared to peak cooling, and the profit logic driven by AI computing power demand has not left the market

After communicating with investors in Toronto, Boston, New York, and San Francisco, Goldman Sachs found that North American investors generally have a more positive attitude towards storage than Asian investors, but their optimism did not fully translate into positive positions. The reasons include lack of significant short-term catalysts and position allocation concerns.

Goldman Sachs said that short-term and long-term expectations are also divided: most respondents expect the average sales price of DRAM and NAND to rise by about 20% month-on-month in the third quarter of 2026, but price increases and profit expectations are lower than before, and the appreciation of the won may also depress the two companies' profits in Korean won.

As for the 2027 HBM price, conservative investors expect a year-on-year increase of about 50%, while more optimistic believe that an increase of more than 100% is needed to bring the profit margin close to traditional DRAM; previously, some expectations were as high as 200%. Goldman Sachs itself still expects the average sales price of HBM to rise by about 100% year on year in 2027, and believes that the improvement of Samsung's product and customer portfolio may bring more room for growth. Goldman Sachs said that these latest signs mean that investors in North America focusing on the memory chip sector are lowering their expectations for extreme price increases while retaining judgment on tight supply and demand and profit resilience.

Whether long-term agreements (LTAs) can make storage profits more stable is the core disagreement revealed in the report. Optimists are optimistic about rolling contracts, greater coverage, and arrangements such as deposits and advance payments to improve order visibility; cautious people want to observe whether these agreements can truly bind buyers and sellers and cross the downward price cycle. As for customers to cut memory configurations and optimize storage usage, most investors surveyed believe that the main reason is insufficient supply rather than a sudden weakening in terminal demand; however, the report also acknowledged that reducing the carrying capacity of each consumer electronics device may offset the increase in the number of HBM or high-performance enterprise-grade SSD products out of data centers in the short term.

The large-scale expansion of memory chip production capacity suppliers from China has also been modeled by many investors, including the scenario where China's DRAM supplier share exceeds 10% in 2028, so the new supply is no longer generally viewed as a sudden impact; respondents still believe that the technical gap and lack of extreme ultraviolet lithography (EUV) equipment limit the iterative performance of DRAM updates and actual production capacity catch-up speed for Chinese data centers.

Goldman Sachs emphasized that the investment appeal of the two companies has its own focus. Samsung's higher exposure to the traditional storage business, promotion of HBM4, and collaboration between storage and foundry makes it possible to improve fundamentals; investors also discussed when the foundry business will balance profit and loss, new contributions and capital expenses of HBM base chips, but there are still differences about how much valuation the foundry business should receive. SK Hynix, on the other hand, is probably more concerned by investors than Samsung on shareholder returns, high stock price flexibility, and the fact that it continues to lead Samsung and Micron's largest market share position in the HBM field, and the opportunity to recover the discount of local Korean stocks compared to its American Depositary Receipts. Respondents generally prefer the latter between cash dividends and repurchases.

North American investors' valuation discussions on Samsung and SK Hynix have moved more towards price-earnings ratio (P/E), but the number of investors expecting a double-digit price-earnings ratio decreased compared to the first half of the year, which is enough to show that the market is still discounting the continuation of the cycle. Goldman Sachs used the segmented valuation method (SOTP) for Samsung, and the target price of preferred shares was discounted by about 27% compared to common shares; Hynix's target price only gave a target price-earnings ratio of 9 times based on average profit from 2026-2027, which can be described as a significant valuation discount compared to Micron.

From the demand of AI agents to the continued expansion of storage profits: the more AI works, the more storage needs to be expanded

Market performance before the market sell-off due to strong US PPI last Thursday already reflected a complete restoration of global capital sentiment on the investment theme of storage. The KOSPI index rebounded about 22% from the July 30 closing low on August 13, entering what is commonly known as a technical bull market. Since then, the KOSPI index has basically fluctuated sideways until September 7, when the KOSPI index rose 4.61% sharply. Samsung and Hynix rose 5.68% and 8.26% respectively. Since this year, Korea's KOSPI index has risen as much as 60%.

Strong AI computing power demand support linked to the AI computing power industry chain level has been clearly reflected in the strong performance and long-term capacity agreement arrangements of industry chain leaders. Nvidia's revenue for the second quarter of fiscal year 2027 was US$96.2 billion, up 106% year on year, and data center revenue was US$89 billion, up 117% year on year. Recently, media reported that Anthropic reached a US$45 billion Nscale computing power lease arrangement and a US$35 billion Lambda cloud computing deal involving about 460 megawatts and 350 megawatts of capacity. These multi-year promises have undoubtedly greatly strengthened the visibility of AI computing power resource requirements around the two core AI hardware systems, AI chips and memory chips.

Judging from engineering principles, the increase in inference demand will simultaneously reinforce the importance of storage bandwidth, operating capacity, and persistent capacity, but the three benefit paths are different. High-bandwidth memory (HBM, itself a DRAM) is close to the accelerator and carries model weight and active key value cache (KV Cache). The performance of many decoding scenarios depends on whether data can be sent to the computing unit in a timely manner; server memory such as DDR5 and LPDDR undertakes CPU workload, data processing, and partial cache stratification; an enterprise-grade solid state drive (SSD) composed of NAND stores model files, knowledge base, task results, and can also host historical KV caches suitable for unloading and reuse.

When the model architecture and cache accuracy are given, longer contexts and more concurrent sessions will increase cache requirements, and continuously running agents will also increase state saving and data reading. Therefore, the common opportunity for Samsung, Hynix, and Micron is to expand the entire storage tier; although SSD can relieve capacity pressure, its latency and bandwidth still determine that it cannot generally replace HBM. Micron's official technical article also clearly explains this hierarchical trend using HBM, main memory, extended memory, context SSD, and network data lake.

At the same time, high-performance AI inference led by Astra and the large-scale spread of AI agent technology focusing on agent-based AI workflows are continuing to blowout HBM/high-performance DRAM capacity requirements at the AI computing level and demand for NAND storage components in data centers.

Another Wall Street financial giant Bernstein recently released a research report showing that as Astra and AI training operator research automation (that is, 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 agency even called out SanDisk's target price of 3,000 US dollars. Bernstein maintained a “outperforming the market” rating for the global memory chip leaders that 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 a new round of positive predictions by Wall Street agencies about the memory chip boom.

Bernstein 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, the pricing and demand for next-generation HBM storage systems and data center server-level DRAM/NAND memory chips closely linked to AI infrastructure is still extremely strong. Bernstein said that July was a low season for traditional semiconductor sales, but the year-on-year increase was 131.4%. In July, global memory chip sales surged 451.7% year on year. Excluding storage, sales in the global semiconductor industry increased by about 35% year on year.

The GPT-6 Astra model recently 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 the 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.

The investment significance brought by Astra is to improve the success rate and economic viability of complex tasks, so that companies are willing to deploy more agents and handle more professional tasks; Wall Street financial giant Morgan Stanley's recent emphasis on “shifting from demand debates to physical supply constraints on AI themes” is a new round of AI computing power resource demand expansion mechanism brought about by Astra, the most advanced model. The statement by the OpenAI product manager that demand is unprecedented and that the company may suspend new Pro subscriptions can be described as an important sign that AI computing power service capacity is under pressure recently.