Wall Street heavyweight Goldman Sachs has released a research note indicating that memory chip stocks and AI data center-related memory product lines are decisively breaking out of their recent weak downtrend and shifting toward a fresh bull run. This comes after the July AI deleveraging storm, a massive selloff driven by extremely crowded bullish positioning, and a global equity market that spent most of the summer in a sideways consolidation pattern.
Goldman identifies a core opportunity at the intersection of expanding demand for cutting-edge high-performance AI compute, driven by the launch of OpenAI's Astra, and the rise of the RSI (Recursive Self-Improvement) training paradigm now dominating AI training. This is colliding with low positioning from traditional Wall Street asset managers and hedge funds.
A Goldman-compiled data point shows fundamental long-short fund net leverage sits at the 4th percentile of the past year, meaning bullish exposure is near annual lows. Meanwhile, implied volatility for the global semiconductor theme has fallen from 65 to 36, a drop of roughly 44.6%, indicating that previously strained risk pricing has eased considerably. If the robust earnings trajectory for memory chips, driven by major model upgrades and the sweeping new AI training paradigm, continues to provide positive catalysts, institutions rebuilding risk exposure could significantly amplify the upside breakout in memory-related stocks.
Memory components in AI data center server clusters remain the clearest supply bottleneck in the AI compute supply chain. Market research firm TrendForce projects server DRAM contract prices will cumulatively rise about 270% by 2026, with enterprise SSD prices up about 235% cumulatively. For 2027, HBM contract prices could still rise 70%-140%, data that reflects the combined force of AI compute expansion and memory price increases. TrendForce's latest estimates show DRAM and NAND combined will account for 47% of major cloud service providers' capital expenditures in 2026, rising to 68% by 2027, driven by both volume growth and price appreciation.
The Korean stock market is showing concrete signs of broadening participation. On September 7, Samsung Electronics rose 5.68% and SK Hynix jumped 8.26%, while the KOSPI index, the Korean benchmark dubbed the "AI compute barometer," surged 4.61% to 6,995.39 points. Foreign investors and institutions recorded net purchases of approximately 2.55 trillion and 2.64 trillion Korean won respectively, with buying expanding beyond corporate buybacks. Calculated from the July 30 low of 5,593.56, the index has rebounded roughly 25.06%, placing it in technical bull market territory. On September 8, the KOSPI slipped 0.58% to 6,954.52, still up about 24.33% from that low, showing the memory rally remains strong even as the broader market is swayed by interest rate and energy risks.
Another Wall Street giant, Nomura, recently stated that AI training and inference expansion continue to drive demand growth while supply expansion remains constrained, with shortages expected to persist through 2028. Nomura also emphasizes that long-term supply agreements (LTAs), through volume locks, price protection, and prepayments, continue to enhance the predictability of future earnings for memory chip giants. Nomura projects that ultimately 50%-70% of sales will be covered by long-term contracts, meaning the market's continued pricing of these stocks as traditional cyclical names at roughly 3 times 2027 expected earnings underestimates the business model shift. This aligns with Goldman's emphasis on a low-position re-entry opportunity: Nomura focuses on the long-term earnings foundation supporting a catch-up trade.
Based on September 8, 2026 closing prices of approximately 269,500 Korean won for Samsung Electronics and 1.793 million won for SK Hynix, Nomura's target prices of 670,000 won for Samsung imply roughly 150% upside over the next 12 months, while the 4.7 million won target for SK Hynix implies about 160% upside.
Memory theme stocks emerging from mid-summer consolidation, Goldman sees breakout signals in a report from Lee Coppersmith, Managing Director in Goldman's Fixed Income, FX, Commodities, and Equities division. He notes that several critical memory supply chain investment targets, including Micron Technology (MU.US), SanDisk (SNDK.US), the iShares MSCI South Korea ETF (EWY.US), and the Roundhill Memory & AI ETF (DRAM.US), are displaying similar bullish technical patterns. Goldman's latest charts indicate these names are beginning to exit their summer consolidation technical indicators, though the moves remain in early-to-mid stages.
These potential breakouts come as the broader AI compute trade enjoys a more favorable market environment: investor bullish positioning is much lighter than at price peaks, implied volatility has fallen, and a slate of potential catalysts is approaching, including the Goldman Communacopia + Technology Conference. Goldman's prime brokerage data shows fundamental long-short hedge fund gross leverage in US equities sits at just the 27th percentile of the past year, with net leverage at only the 4th percentile. This highlights that hedge fund institutions have begun rebuilding AI compute positions, but levels remain significantly below those of roughly two months ago.
The options market has also undergone a major adjustment. Semiconductor implied volatility, benchmarked by the Cboe Semiconductor ETF Volatility Index (VXSMH), has fallen from roughly 65 in July to 36, nearly halved, and is back near levels seen in early 2026. Meanwhile, even after Friday's nearly 7% rebound, Goldman's basket of tech long-short momentum trades remains about 50% off its late-June highs. Coppersmith says this adjustment has narrowed the range of potential outcomes priced into AI compute-related stocks but has not created a distinctly bearish bias.
The Korean market may also provide further upside momentum for the compute theme. Goldman notes that over the past four weeks, capital inflows into Korean equities were entirely supported by corporate buybacks, with other investor flows in net selling territory overall, though the selling pace has narrowed significantly. This leaves room for more investors to deeply participate in the Korean market, especially in the renewed bullish wave for the world's two largest memory chip stocks, SK Hynix and Samsung Electronics.
Astra Ignites the AGI Wave and AI Begins Helping Develop AI, Memory Chip Re-rating Gains a Dual Demand Expansion Curve. OpenAI's newly launched Astra model continues to expand the range of professional tasks AI can undertake. Nvidia CEO Jensen Huang took to social media on Sunday with the heavyweight statement that the arrival of GPT-6 Astra means "AGI is here." Combined with Nvidia's confirmed strong revenue range and subsequent robust shipment guidance, and the AI model development shift into the "Recursive Self-Improvement (RSI)" phase, which opens another curve of surging AI compute demand, the investment case is strengthened. Astra could expand commercial application demand for AI compute, while the AI-building-AI R&D trajectory may increase frontier experimentation, evaluation, and continuous long-term training investment, together extending the compute investment cycle.
The investment significance of Astra and RSI lies in frontier high-performance AI models, and AI R&D itself, becoming new scenarios that continuously consume compute power. OpenAI disclosed on September 6 that it has achieved the "automated research intern" goal, capable of completing tasks under human guidance that would otherwise take skilled researchers several days. As of mid-August, each human workday corresponded to roughly 3.1 agent workdays running. This measures runtime, not a 3.1x increase in research output. From this, it can be inferred that research automation will simultaneously increase code generation, inference needed for experimental evaluation, and candidate model training demand. However, full RSI is not yet an established dominant paradigm; research direction and resource allocation are still determined by humans.
Astra represents a mechanism for expanding cutting-edge performance demand: as model capabilities improve, tasks previously too unreliable to perform enter commercial viability. Additionally, Astra could shift the demand curve outward overall, meaning that as AI models get smarter, companies can attempt work they previously couldn't reliably execute, while competitors must continue investing in R&D and training. This provides powerful new support for the AI spending cycle.
OpenAI's GPT-6 Astra model and the RSI technical path embraced by AI leaders are poised to become two core drivers of exponential AI compute demand growth. More capable AI models, broader use of AI application tools, and the next-generation AI training path requiring more compute are strengthening the case for sustained growth in AI infrastructure demand. OpenAI disclosed that Astra scored 98% on the FrontierMath Level 4 test and 99.9% on ARC-AGI-3. Huang used this to declare "AGI has arrived" and stated that model training used over 100,000 Nvidia GPUs, with another 400,000 GPUs coming online soon. Notably, "AGI has arrived" remains a contested claim, but the announcement of larger Nvidia AI GPU cluster deployments directly reinforces expectations for strong AI compute demand as frontier models continue to expand training resource investment.
At a technical level, memory demand depends on parameter scale, context length, concurrency, and experimental density. HBM handles model weights, intermediate training states, and active KV caches on the GPU side. Server DRAM handles data processing, runtime environments, and cache offloading. NAND enterprise SSDs store datasets, training checkpoints, and reusable caches. An Nvidia example shows Llama 3 70B loaded at FP16 precision requires about 140GB of memory, with a single-user 128K token context requiring an additional ~40GB for KV cache. When stronger models handle longer tasks, more agents run concurrently, and RSI research processes add parallel experiments and checkpoint saving, capacity, bandwidth, and read/write throughput requirements expand dramatically in tandem.
KB Securities from South Korea projects memory's share of AI infrastructure investment will rise from 14% in 2025 to 40% in 2026, then to 57% in 2027. KB's core bullish thesis for SK Hynix and Samsung centers on the gap between razor-thin inventory buffers and low forward earnings valuations. KB states that Samsung and SK Hynix memory inventories are below 10 days, and projects hyperscaler AI infrastructure investment will reach $1.3 trillion in 2027, up about 60% year over year. Based on the stock prices and earnings forecasts in its report, both companies have retraced roughly 38% from prior highs, corresponding to a 2027 P/E of just about 3 times. Therefore, KB is betting on demand expansion and price increases driving upward earnings revisions and subsequent valuation recovery.
For the KOSPI benchmark, Goldman has even set a 12,000-point target. As of September 8, the KOSPI had slipped 0.58% to 6,954.52. Goldman directly asserts that the market systematically underestimates the duration of the AI-driven memory chip demand cycle, and has significantly raised its forecast for US big tech capital expenditures next year to $1.2 trillion, arguing the "memory shortage" triggered by data center expansion will intensify further in 2027.
Comments