Storage chips break through midsummer consolidation! AGI boom combines with new RSI training paradigm, Goldman Sachs senses the start of a new storage bull market

According to English Financial News, Wall Street financial giant Goldman Sachs recently released a research report stating that, after the July AI deleveraging storm and the massive sell-off from the extremely crowded bullish positions leading to a continued sharp decline in the global storage theme, and with the global stock markets spending most of the summer in sideways consolidation and box fluctuations, storage chip stocks and investment targets related to the AI data center construction boom have started to break significantly out of the recent downward trend, heading towards a new bullish rally trajectory.

The core opportunity captured by Goldman Sachs lies in the expectation surge for storage chip demand, driven by the high-performance AI computing power brought by the emergence of OpenAI Astra and the recursive self-improvement (RSI) training paradigm beginning to dominate AI training, intersecting with traditionally low-position allocations by Wall Street asset managers and hedge funds.

Goldman Sachs’ latest compiled data shows that fundamental long-short fund net leverage is at the 4th percentile over the past year, indicating that long exposure is near annual lows; the implied volatility for the global semiconductor theme has fallen from 65 to 36, a decrease of about 44.6%, suggesting that previously extreme risk pricing has significantly eased. Should the robust profit trajectory of storage chips, catalyzed by global large model iterations and new AI training paradigms, persist, institutions increasing their risk exposures could substantially amplify the post-consolidation rally for storage-related stocks.

The storage chip components of AI data center server clusters remain the clearest supply bottleneck in the AI computing industry chain. Market research firm TrendForce predicts that by 2026, server DRAM contract prices will cumulatively rise by about 270%, and enterprise SSD prices by about 235%; even HBM contract prices are expected to rise by 70%—140% in 2027. These data reflect the combined effect of AI computing expansion and storage price surges. TrendForce’s latest estimates show that the combined share of DRAM and NAND in major cloud providers' capital expenditures will rise from 47% in 2026 to 68% in 2027, driven by both increased procurement volume and price hikes.

The Korean stock market has already shown concrete signals of expanding capital participation. On September 7, Samsung Electronics rose by 5.68%, SK Hynix by 8.26%, and the KOSPI index—known as the "AI computing barometer"—surged by 4.61% to close at 6,995.39 points; foreign and institutional investors bought a net of about KRW 2.55 trillion and KRW 2.64 trillion respectively, with buying expanding beyond corporate buybacks. Calculated from the July 30 low of 5,593.56 points, the index had rebounded by about 25.06%, entering the technical bull market range. On September 8, the KOSPI index retreated 0.58% to 6,954.52 points, still up about 24.33% from that low, showing robust storage momentum, though the overall market remains influenced by interest rates and energy risks.

Another Wall Street giant Nomura Securities recently stated that AI training and inference expansion are driving continuous demand growth, while supply expansion remains constrained, and the tight market is expected to last until 2028; additionally, Nomura emphasized that long-term supply agreements (LTAs) through volume locking, price protections, and prepayments continue to improve the long-term earnings predictability for storage chip giants.

Nomura predicts that eventually about 50%—70% of sales will be covered by long-term agreements, thus, the market's continued valuation of these stocks as traditional cyclical names with roughly 3x 2027 forward PE underestimates the shift in business models. This echoes Goldman Sachs' emphasis on the opportunity for low-position replenishment; Nomura values the long-term profit foundation supporting this restocking rally. Based on Samsung Electronics' closing price of approximately KRW 269,500 and SK Hynix’s KRW 1,793,000 as of September 8, 2026, Wall Street giant Nomura’s target price of KRW 670,000 for Samsung Electronics suggests a potential 150% upside in the next 12 months, while SK Hynix’s target of KRW 4,700,000 indicates about 160% potential upside.

Storage-themed stocks break out of midsummer consolidation as Goldman Sachs detects positive breakout signals

According to a report by Lee Koppersmith, Managing Director of Fixed Income, FX, Commodities, and Equities at Goldman Sachs, several critical storage industry investment targets—Micron Technology (MU.US), SanDisk (SNDK.US), iShares MSCI South Korea ETF (EWY.US), and Roundhill Storage Chip ETF (DRAM.US)—are showing similar bullish technical patterns. The latest Goldman Sachs charts show these targets starting to break out of summer consolidation technical indicators, though those moves are still in early or mid-stages.

These potential breakouts come as the broader AI computing theme enjoys a more favorable market environment: investor bullish positions are much lighter than they were at share price peaks, implied volatility is decreasing, and several potential catalysts are approaching, including the upcoming Goldman Sachs Communacopia+ Technology Conference.

Goldman Sachs Prime Brokerage data shows total leverage for fundamental long-short hedge funds in the US stock market is at just the 27th percentile for the past year, with net leverage at only the 4th percentile. This highlights that hedge fund institutional investors are starting to rebuild AI computing positions, but levels are still significantly lower than about two months ago.

The options market has also undergone a dramatic reset. Semiconductor implied volatility, as benchmarked by the Chicago Board Options Exchange Semiconductor ETF Volatility Index (VXSMH), has dropped from around 65 in July to 36, nearly halving and returning close to early 2026 levels.

Meanwhile, Goldman’s compiled US stock market tech long-short momentum basket, despite rebounding nearly 7% last week, remains about 50% below its late-June high. Koppersmith notes this adjustment has narrowed the potential performance range reflected in AI computing-related stock prices, but has not formed a clear bearish market trend.

The Korean market may also offer further sentiment-driven momentum for the computing theme. Goldman Sachs said that over the past four weeks, inflows to the Korean stock market were entirely driven by corporate buybacks, while capital flows from other investors remained net negative—but the pace of selling has narrowed significantly. This leaves room for more investors to actively participate in Korea’s stock market, especially in another round of bullish frenzy for the two largest global storage chip stocks—SK Hynix and Samsung Electronics.

Astra ignites AGI mania + AI begins to participate in AI development, with storage chip re-rating seeing a dual demand expansion curve

The newly launched OpenAI Astra large model continues to expand the scope of professional tasks AI can undertake; Nvidia CEO Jensen Huang went further on social media on Sunday, stating that the launch of GPT-6 Astra means "AGI has arrived," and Nvidia has already confirmed robust revenue guidance and strong shipment outlooks. On top of that, AI large model development has entered a new phase—“recursive self-improvement (RSI)”—opening another steep curve of AI computing demand: Astra is expected to boost commercial application-end AI computing needs, while the trajectory of AI "creating AI" research may increase investment in frontier operator experiments, evaluations, and long-term sustained training, all extending the computing investment cycle.

The investment significance of Astra and RSI lies in the fact that advanced high-performance large AI models, as well as AI R&D itself, are becoming sustained new scenarios for continuous compute consumption. OpenAI revealed on September 6 that it has achieved the “automated research intern” milestone, able to finish some tasks under human supervision that previously required skilled researchers several days to complete; as of mid-August, each human workday equated to about 3.1 agent workdays. This measures run-time, not a 3.1-fold increase in research output. Thus, research automation will simultaneously increase the inference required for code generation and experimental evaluation, as well as candidate model training demand. However, full RSI is not yet an established leading paradigm, as research direction and resource allocation remain determined by humans.

Astra represents the most cutting-edge mechanism for expanding performance demand: With improved large model capabilities, tasks that were previously hard to reliably complete become commercially viable. Additionally, Astra could shift the entire demand curve outward—when large AI models become more intelligent, enterprises can attempt jobs previously unreliable, and competitors need to keep investing in research and training too, providing strong support for the AI spending cycle.

OpenAI’s release of the GPT-6 Astra large model, along with the AI leaders’ focus on RSI technical paths, is expected to be the two driving forces behind exponential AI computing demand expansion. That is, stronger performance in large AI models, more widespread use of AI tools, and greater compute demands for the next generation of AI training paths, are crucial evidence supporting sustained growth in AI computing infrastructure demand.

OpenAI has revealed that Astra scored 98% on the FrontierMath Level 4 test and 99.9% on the ARC-AGI-3 benchmark. On this, Jensen Huang declared “AGI has arrived” and said model training used over 100,000 Nvidia GPUs, with another 400,000 GPUs coming online soon. It’s worth noting that “AGI has arrived” remains a controversial verdict, but the expectation for even larger Nvidia AI GPU clusters to be deployed directly strengthens an outlook of robust ongoing AI compute demand as frontier AI large models continue to ramp up training resources.

On the technical side, storage demand depends on parameter size, context length, concurrency, and experiment density. HBM carries the GPU-side model weights, training intermediates, and active key-value caches; server DRAM handles data processing, runtime environments, and cache offloading; enterprise NAND SSDs store datasets, training checkpoints, and reusable caches. According to a Nvidia example, loading Llama 3 70B weights at FP16 precision requires about 140GB of memory; a single user with a 128,000-token context would need another ~40GB for KV cache. As more powerful models handle longer tasks and more agents operate in parallel, with RSI research flows adding concurrent experiments and checkpoint saving, capacity, bandwidth, and I/O throughput requirements all surge together.

KB Securities from Korea expects storage’s share of AI infrastructure investment will climb from 14% in 2025 to 40% in 2026, then to 57% in 2027. KB’s core bullish logic for SK Hynix and Samsung focuses on the gap between extremely thin inventory buffers and long-term profit valuations being too low.

KB Securities notes that storage inventories at Samsung and SK Hynix are below ten days and forecasts hyperscale cloud AI infrastructure investment will reach $1.3 trillion in 2027, up roughly 60% year-over-year. According to projections based on current share prices and earnings outlooks, the two firms have retreated about 38% from prior highs, equating to just 3x 2027 PE. Therefore, KB bets on demand expansion and price increases driving up earnings expectations and triggering a valuation recovery. For Korea’s KOSPI benchmark, Goldman Sachs has set a target of 12,000 points; as of September 8, the KOSPI retreated 0.58% to 6,954.52 points. Goldman Sachs notes the market systematically underestimates the duration of the AI-driven storage chip demand cycle and has sharply revised upward next year’s capital expenditure forecasts for large US tech companies to $1.2 trillion, asserting the data center expansion-induced "storage crunch" will intensify further in 2027.