As "AI slowdown" impacts the semiconductor sector, Goldman Sachs issues a bullish report! Target prices for the "Korean memory chip giants" indicate nearly 90% upside potential.

According to Zhitong Finance APP, as global AI leaders such as Anthropic and OpenAI jointly call for a slowdown in the development of cutting-edge AI models, the global stock market is simultaneously reassessing investment growth expectations in the AI computing power supply chain, cooling down due to rising oil prices and the risk of a new round of rate hikes by the Federal Reserve. AI computing power-related stocks were generally weak, with SK Hynix dropping over 5% in early trading on the Korean stock market, Samsung Electronics falling over 3%, and the Korean KOSPI index—dubbed the "AI computing power barometer"—down more than 3%. However, many veteran Wall Street analysts stated that the latest developments will not have a lasting impact on the industry and are unlikely to undermine the long-term AI computing power bull market narrative.

The latest “Korea Technology Industry Investor Feedback Report” released by Wall Street financial giant Goldman Sachs shows that North American investors are very positive about memory chip stocks, much more so than their Asian counterparts. The institution reiterated its “buy” ratings on the world’s two largest memory chip giants—Samsung Electronics and SK Hynix—while Samsung Electronics continues to be included in Goldman Sachs’ “Conviction List.”

According to the research report, Goldman Sachs analysts set a target price of KRW 490,000 for Samsung common shares, KRW 360,000 for preferred shares, and KRW 3,500,000 for SK Hynix; based on the closing price on September 11, the potential stock price increases are approximately 88.8%, 86.2%, and 93.2%, respectively. Goldman Sachs continues to expect the average selling price of HBM to rise about 100% year-on-year in 2027, with Samsung set to benefit from improved product and customer mix, and both companies also possessing potential catalysts in shareholder returns.

Some analysts noted that calls to slow the development of cutting-edge models do not directly equate to reduced computing power capital expenditures or memory chip orders; demand for inference and applications for existing models may also continue to grow. “This may bring some short-term pressure, but it is unlikely to derail the long-term AI trade. AI development is still at a relatively early stage, and I am not sure whether other participants in the AI ecosystem are willing to accept the current industry rankings and slow down at a time when technology is evolving so rapidly,” said Gary Tan, portfolio manager at Allspring Global Investments in Singapore.

“The three CEOs' agreement to control the pace will not truly change the funding going into chips, power, and infrastructure. In fact, it extends the development timeline,” said Billy Leung, investment strategist at Global X Management in Sydney. “If commercialization and adoption continue to grow while the rollout of new capabilities slows somewhat, this actually helps the industry shift from spending to build towards accelerating monetization of what’s already built—that is, AI monetization.”

For investors, key indicators for memory chip demand and price-volume increases will undoubtedly be actual bit shipments by memory chip manufacturers, customer certifications, contract prices, profit margins, free cash flow data, and stage guidance intervals; as for the optimistic target prices from institutions like Goldman Sachs, these ultimately still need to be substantiated by these metrics.

Price hike expectations have cooled compared to peak levels, but the profit-driven logic of AI computing power demand remains intact

After communicating with investors in Toronto, Boston, New York, and San Francisco, Goldman Sachs found that North American investors are generally more bullish on memory than Asian investors, but optimism has not fully translated into active buying due to a lack of major short-term catalysts and portfolio allocation concerns.

Goldman Sachs noted a divergence between short-term and long-term expectations: Most respondents expect DRAM and NAND average selling prices to rise about 20% quarter-on-quarter in the third quarter of 2026, but price increase and profit expectations have been lowered compared to before, and won appreciation may also depress the companies’ won-denominated profits.

As for 2027 HBM pricing, conservative investors forecast a year-on-year increase of around 50%; more optimistic ones see a need for increases of more than 100% to bring profit margins close to traditional DRAM, whereas some previous expectations reached as high as 200%. Goldman Sachs itself still expects the average selling price of HBM to rise about 100% year-on-year in 2027, and believes Samsung’s product and customer mix improvement could provide greater upside. Goldman Sachs said these latest signals indicate that North American memory-focused investors are reducing their expectations for extreme price hikes while maintaining their judgement on supply-demand tightness and profit resilience.

Whether Long-Term Agreements (LTAs) can make memory profits more stable is a core point of debate revealed in the report. Optimists favor rolling contracts with broader coverage, and arrangements like deposits and prepayments to improve order visibility; cautious voices want to see if these agreements can truly bind buyers and sellers and endure downward price cycles. Regarding customers reducing memory allocation and optimizing memory usage, most surveyed investors believe the main reason is insufficient supply, not a sudden weakening of terminal demand; but the report also acknowledges that reducing memory per consumer electronic device could offset, in the short term, some bit shipment growth from data center HBM or high-performance enterprise SSDs.

Many investors have also incorporated the large-scale supply expansion of Chinese memory chip manufacturers into their models, including a scenario where Chinese DRAM suppliers capture more than 10% market share by 2028. Thus, new supply is no longer widely seen as a sudden shock; respondents still believe that the technology gap and the lack of EUV (extreme ultraviolet) lithography equipment limit China’s pace in updating data center server DRAM and actual capacity catch-up.

Goldman Sachs emphasized that each company’s investment appeal has its own focus. Samsung’s larger traditional memory exposure, HBM4 advancement, and synergy between memory and foundry businesses provide potential for fundamental improvement; investors discussed when the foundry business would break even, contributions from HBM base chips, and capex, but there is still disagreement on how much value should be assigned to the foundry segment. SK Hynix may be more attractive to investors focused on shareholder returns, higher stock price elasticity, maintaining the largest market share in HBM far ahead of Samsung and Micron, securing Nvidia’s largest HBM orders, and the opportunity for the discount of Korean local shares to their US ADRs to narrow. Most respondents prefer share buybacks over cash dividends.

Among North American investors, valuation discussions of Samsung and SK Hynix have shifted more towards P/E ratios, but those expecting double-digit P/E multiples have decreased compared to the first half of the year, evidencing that the market still discounts for cycle sustainability. Goldman Sachs uses the sum-of-the-parts valuation (SOTP) for Samsung, with the target price for preference shares about 27% lower than common shares; SK Hynix’s target price is based only on 9x average earnings for 2026–2027, representing a significant valuation discount compared to Micron.

From AI Agent Demand to Sustained Memory Profit Growth: The More Intelligent AI Becomes, the More Memory Must Expand

Prior to last Thursday’s US PPI-driven market selloff, the market performance already reflected full sentiment recovery among global investors toward the memory theme. The KOSPI index rebounded about 22% from the July 30 closing low to August 13, entering what is usually called a technical bull market, and then remained sideways until the index surged 4.61% on September 7, with Samsung and Hynix rising 5.68% and 8.26%, respectively. Year-to-date, the Korean KOSPI index has soared 60%.

The strong AI computing demand tied to the AI computing supply chain has already been notably reflected in the robust performance of industry leaders and long-term capacity contracts. Nvidia’s fiscal Q2 2027 revenue was $96.2 billion, up 106% year-on-year, with data center revenue at $89 billion, up 117% year-on-year; recent media reports said that Anthropic has secured $45 billion in Nscale computing power leasing arrangements and a $35 billion Lambda cloud computing deal, involving about 460MW and 350MW of capacity. These multi-year commitments have undoubtedly strengthened the visibility of AI computing power demand for core hardware systems of AI chips and memory chips.

From an engineering perspective, increased inference demand simultaneously enhances the importance of memory bandwidth, operating capacity, and persistent capacity, but each benefits differently. High Bandwidth Memory (HBM, a type of DRAM) sits close to accelerators, hosting model weights and active KV caches; many decoding scenarios’ performance hinges on whether data can be swiftly delivered to the compute unit. DDR5, LPDDR, and other server memory undertake CPU workloads, data processing, and some cache layers; NAND-based enterprise SSDs store model files, knowledge bases, task results, and can host historical KV caches suitable for offloading and reuse.

Given model architectures and cache precision, longer contexts and more concurrent sessions expand cache requirements, and continuously running agents further boost the need for state saving and data reading. Thus, Samsung, Hynix, and Micron collectively benefit from expanding the entire memory stack; though SSDs help ease capacity pressure, their latency and bandwidth still mean they cannot generally replace HBM. Official Micron technical papers also illustrate this tiering trend using HBM, main memory, extended memory, context SSDs, and network data lakes.

Meanwhile, the widespread adoption of high-performance AI inference led by Astra and agent-centric AI workflow technologies is continuously driving explosive growth in data center demand for HBM/high-performance DRAM and NAND storage components for AI computation.

Another Wall Street financial giant, Bernstein, recently released a research report stating that the semiconductor industry has seen a seasonal dip as expected, but semiconductor demand related to AI computing infrastructure—especially for next-generation HBM memory systems and data center server-level DRAM/NAND memory pricing and demand—remains extremely strong. Bernstein points out that July is usually an off-season for chip sales, but sales were still up 131.4% year-on-year; global memory chip sales in July jumped an astonishing 451.7% year-on-year, and excluding memory chips, the global semiconductor industry grew by about 35% year-on-year.

OpenAI’s recent launch of the GPT-6 Astra large model and the RSI technology path pursued by leading AI companies are expected to become the two core drivers of exponential demand for AI computing power—meaning next-generation AI training routes that demand stronger performance and broader AI application tools, all continuing to boost core infrastructure demand growth in AI computing.

The significance of Astra for investors lies in increasing the success rate and economic feasibility of complex tasks, encouraging enterprises to deploy more agents and process more specialized tasks. Morgan Stanley, another Wall Street giant, recently emphasized the “shift from demand debates to physical supply constraints of the AI theme,” which is exactly the new mechanism by which Astra, the most advanced large model, is expanding AI computing resource demand. The statement by OpenAI's Head of Product about unprecedented demand to the extent that the company may suspend new Pro subscriptions is a crucial recent signal of the pressure on AI computing service capacity.