Amid the "AI slowdown" hitting semiconductors, Goldman Sachs delivers a bullish report! "Korea's memory duopoly" price targets point to nearly 90% upside

Title context: Amid the "AI slowdown" hitting semiconductors, Goldman Sachs delivers a bullish report! "Korea's memory duopoly" price targets point to nearly 90% upside Text: As global AI leaders such as Anthropic and OpenAI unanimously call for slowing the pace of frontier AI large-model development, global stock markets are simultaneously repricing expectations for cooling investment growth across the AI computing power supply chain, along with rising oil prices and a new round of Federal Reserve rate-hike risks. AI computing power-themed stocks have broadly weakened. SK Hynix fell more than 5% at one point in early trading on the Korean stock market, Samsung Electronics fell more than 3%, and South Korea's KOSPI index, known as the "AI computing power bellwether," dropped more than 3%. However, many senior Wall Street analysts said the latest developments will not have a lasting impact on the industry and are unlikely to undermine the long-term bull-case logic of the AI computing power trade. A newly released report from Wall Street financial giant Goldman Sachs, titled "Investor Feedback Report on the Korean Technology Industry," shows that North American investors hold very positive views on memory chip stocks, especially much more positive than Asian investors. The firm reiterated its "Buy" ratings on the world's two largest memory chip giants, Samsung Electronics and SK Hynix, with Samsung Electronics remaining on Goldman Sachs' "Conviction List." According to the research report, Goldman Sachs analysts set a target price of as high as KRW 490,000 for Samsung's common stock and KRW 360,000 for its preferred stock, and a target price of as high as KRW 3.5 million for SK Hynix. Based on the closing prices on September 11, the potential upside would be approximately 88.8%, 86.2%, and 93.2%, respectively. Goldman Sachs continues to expect HBM blended average selling prices to rise about 100% year over year in 2027. Samsung is expected to benefit from an improved product and customer mix, and both companies also have potential catalysts in shareholder returns. Some analysts said that calls to slow frontier model development cannot be directly equated with cuts to computing power capital expenditure or memory orders, and inference and application demand for existing models may continue to grow. "This could create some short-term pressure, but it is unlikely to break 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 while technology is still evolving so rapidly," said Gary Tan, a portfolio manager at Allspring Global Investments in Singapore. "The three CEOs agreeing to control the pace does not really change the money going into chips, power, and infrastructure. In fact, it extends the development timeline," said Billy Leung, an investment strategist at Global X Management in Sydney. "If commercialization and adoption continue to grow while the rollout of new capabilities slows slightly, that actually helps the industry shift from spending money to build toward accelerating profits from what has already been built, in other words, AI monetization." For investors, the key indicators regarding memory chip demand and simultaneous growth in volume and price will undoubtedly be memory chip makers' actual bit shipments, customer certifications, contract prices, margin and free cash flow data, and phased outlook ranges. Optimistic target prices from Wall Street institutions such as Goldman Sachs will ultimately still need these indicators to materialize. Price-hike expectations cool from their peak, but the profit logic driven by AI computing power demand has not exited the stage After speaking with investors in Toronto, Boston, New York, and San Francisco, Goldman Sachs found that North American investors are generally more positive on memory than Asian investors, but the optimism has not fully translated into aggressive position adding, due to reasons including a lack of major short-term catalysts and position-allocation concerns. Goldman Sachs said short-term and long-term expectations have also diverged: most respondents expect DRAM and NAND average selling prices to rise about 20% quarter over quarter in the third quarter of 2026, but price-hike and profit expectations have been revised down from before, and won appreciation could also both companies' won-denominated earnings. For 2027 HBM prices, conservative investors expect an increase of about 50% year over year, while more optimistic investors believe an increase of more than 100% is needed to bring margins close to those of conventional DRAM. Some earlier expectations had been as high as 200%. Goldman Sachs itself still expects HBM blended average selling prices to rise about 100% year over year in 2027 and believes Samsung's improved product and customer mix could bring greater upside. Goldman Sachs said these latest signs mean that North American investors focused on the memory chip sector are lowering expectations for extreme price increases while retaining their judgment on tight supply-demand and earnings resilience. Whether Long-Term Agreements (LTAs) can make memory profits more stable is the core disagreement revealed by the report. Optimists are bullish on rolling contracts, broader coverage, and arrangements such as deposits and prepayments improving order visibility. Cautious investors want to observe whether these agreements can truly bind both buyers and sellers and survive a downcycle in prices. Regarding customers cutting memory configurations and optimizing memory usage, most surveyed investors believe the main reason is insufficient supply rather than a sudden weakening in end demand. But the report also acknowledges that reducing memory capacity per consumer electronics device could in the short term offset some of the growth in data center HBM or high-performance enterprise SSD shipments. Large-scale supplier expansion of memory chip capacity from China has also been incorporated into many investors' models, including a scenario in which Chinese DRAM suppliers hold more than 10% share in 2028. Therefore, new supply is no longer widely viewed as a sudden shock. Respondents still believe the technology gap and the lack of extreme ultraviolet (EUV) lithography equipment limit the pace at which China can catch up in performance and actual capacity for data center server DRAM upgrades. Goldman Sachs emphasized that the two companies' investment appeal has different emphases. Samsung's greater exposure to conventional memory, progress on HBM4, and synergies between memory and foundry give it room for fundamental improvement. Investors also discussed when the foundry business will break even, the new contribution from HBM base dies, and capital expenditure, but there is still disagreement over how much valuation the foundry business should receive. SK Hynix may attract more investor attention than Samsung on shareholder returns, higher stock price elasticity, its continued position as the largest market share holder in HBM ahead of Samsung and Micron, its capture of Nvidia's largest HBM orders, and the opportunity to repair the discount of Korean local shares relative to their U.S. depositary receipts. Respondents generally prefer buybacks over cash dividends. North American investors' valuation discussions for Samsung and SK Hynix have shifted more toward price-to-earnings (P/E), but fewer investors expect double-digit P/E ratios than in the first half, which is enough to show that the market still applies a discount to cycle durability. Goldman Sachs uses a sum-of-the-parts (SOTP) valuation for Samsung, with the preferred stock target price at about a 27% discount to the common stock. For Hynix, the target price is based on only 9x target P/E on average earnings for 2026-2027, a significant valuation discount compared with Micron. From AI agent demand to continued expansion of memory profits: the better AI works, the more memory must scale The market performance before last Thursday's selloff triggered by strong U.S. PPI had already reflected a full recovery in global capital sentiment toward the memory investment theme. The KOSPI index had rebounded about 22% by August 13 from its closing low on July 30, entering what is commonly called a technical bull market, and then largely moved sideways until September 7, when the KOSPI surged 4.61%, with Samsung and Hynix rising 5.68% and 8.26%, respectively. So far this year, South Korea's KOSPI index has gained as much as 60%. Strong AI computing power demand support linked to the AI computing power supply chain has already been significantly reflected in industry leaders' strong results and long-term capacity agreements. Nvidia's revenue for the second quarter of fiscal 2027 was $96.2 billion, up 106% year over year, with data center revenue of $89 billion, up 117% year over year. Recent media reports said Anthropic reached a $45 billion computing power leasing arrangement with Nscale and a $35 billion Lambda cloud computing deal, involving capacity of about 460 megawatts and 350 megawatts. These multi-year commitments undoubtedly greatly strengthen the visibility of AI computing power resource demand around the two core AI hardware systems: AI chips and memory chips. From an engineering perspective, increased inference demand simultaneously strengthens the importance of memory bandwidth, runtime capacity, and persistent capacity, but the three benefit through different paths. High Bandwidth Memory (HBM, itself a type of DRAM) sits next to accelerators and carries model weights and active key-value cache (KV Cache). Performance in many decoding scenarios depends on whether data can be delivered to compute units in time. Server memory such as DDR5 and LPDDR handles CPU workloads, data processing, and some cache tiering. Enterprise solid-state drives (SSDs) made from NAND store model files, knowledge bases, and task results, and can also handle historical KV cache suitable for offloading and reuse. Given a model architecture and cache precision, longer context and more concurrent sessions expand cache demand, and continuously running agents also increase state preservation and data reads. Therefore, the common opportunity for Samsung, Hynix, and Micron lies in scaling across the entire memory hierarchy. Although SSDs can relieve capacity pressure, their latency and bandwidth still mean they cannot generally replace HBM. Micron's official technical articles also clearly explain this tiering trend through HBM, main memory, expanded memory, context SSDs, and network data lakes. At the same time, high-performance AI inference led by Astra and the large-scale adoption of AI agent technologies focused on agentic AI workflows are continuing to explosively drive up demand for HBM/high-performance DRAM capacity at the AI computing layer and demand for data center NAND storage components. Another Wall Street financial giant, Bernstein, recently released a research report showing that as Astra and AI training operator research automation, namely Astra and RSI, can be said to provide a new source of semiconductor demand for this unprecedented memory chip boom-driven semiconductor upcycle, the institution even called for a $3,000 target price on SanDisk. Bernstein maintained "Outperform" ratings on the global memory chip leaders that have surged this year, namely Samsung Electronics, SK Hynix, Micron, and SanDisk, with target prices of KRW 440,000, KRW 3.3 million, $1,300, and $3,000, respectively, reflecting a new round of positive Wall Street institutional forecasts for memory chip prosperity. Bernstein recently released a research report saying that the semiconductor industry has shown a seasonal pullback as the market expected, but semiconductor demand related to AI computing power infrastructure construction, especially pricing and demand for next-generation HBM memory systems and data center server-grade DRAM/NAND memory chips closely tied to AI infrastructure, remains extremely strong. Bernstein said July is traditionally a slow season for semiconductor sales but still grew 131.4% year over year. In July, global memory chip sales surged 451.7% year over year. Excluding memory, global semiconductor industry sales grew about 35% year over year. OpenAI's recently launched GPT-6 Astra large model and the RSI technology path focused on by AI leaders are expected to become the two core DRIVE factors driving exponential expansion of AI computing power demand, namely stronger AI large models and broader use of AI application tools, and a next-generation AI training path with even stronger computing power demand. These are important evidence strengthening the continued growth of AI computing power infrastructure demand. The investment significance of Astra lies in improving the success rate and economic feasibility of complex tasks, making enterprises willing to deploy more agents and handle more specialized tasks. The "shift from the demand debate back to physical supply constraints in the AI theme" recently emphasized by Wall Street financial giant Morgan Stanley is precisely the new round of AI computing power resource demand expansion mechanism brought by the most frontier large model Astra. OpenAI's product chief's statement that demand is unprecedented to the point that the company may pause new Pro subscriptions can be described as an important signal of recent pressure on AI computing power service capacity.