Astra's true strengthening lies not simply in selling more of a particular chip, but in bolstering market confidence that frontier models will continue evolving toward complex reasoning, long context lengths, and agent-based tasks—workloads that simultaneously increase demand for computing power, memory bandwidth, server DRAM, and high-speed storage.

Article author, source: ME News



TL;DR

• Astra's true strengthening lies not simply in selling more of a particular chip, but in bolstering market confidence that frontier models will continue evolving toward complex reasoning, long context lengths, and agent-based tasks—workloads that simultaneously increase demand for computing power, memory bandwidth, server DRAM, and high-speed storage.
• On September 4, U.S. stock prices for storage and memory companies led the rally, with SanDisk rising 11.9% and Micron up 6.1%. On September 7, South Korean markets followed suit, with SK Hynix gaining 8.26% and Samsung Electronics rising 5.68%. This indicates that capital is shifting from solely buying GPUs to identifying still-profitable bottlenecks within AI infrastructure.
• This round of storage market momentum was not entirely created by Astra. According to TrendForce data, global DRAM industry revenue increased by 59.5% quarter-over-quarter in the second quarter of 2026 to approximately $154.7 billion, with supply expansion continuing to lag behind demand; Astra merely brought the pre-existing narratives of scarcity, price increases, and upward profit revisions back into the spotlight.
• "Storage transactions" cannot be generalized. HBM is most tightly coupled with GPUs and AI ASICs, server DRAM is more directly driven by AI inference and agent workloads, while NAND is primarily influenced by datasets, vector databases, checkpoints, and data tiering—each differs in terms of market sensitivity, technological barriers, and cyclical risks.
• What truly matters to track are not daily or weekly stock prices, but three key variables: whether AI capital expenditures continue to be revised upward, whether memory contract prices and supply-demand gaps can be sustained, and whether the mass production and customer certification of HBM4 and next-generation products proceed smoothly.

Why is Astra redirecting funds toward memory chips?

On September 3, OpenAI officially released GPT-6 Astra, positioning it as a flagship model designed for complex, multi-step professional tasks, with key capabilities in reasoning, programming, research, computer operations, and end-to-end task execution across tools. While the product launch itself does not equate to OpenAI announcing plans to purchase more memory, it sends a significant signal to capital markets: the competition in frontier models has not entered a phase focused solely on cost reduction; the industry continues to pursue longer task chains, more complex reasoning, and higher levels of autonomous execution. For infrastructure, this means bottlenecks will no longer be confined to GPU matrix computation capabilities—how data rapidly enters computational units, how state is maintained during long contexts and multi-turn reasoning, and how numerous agents simultaneously read and write data will become increasingly critical.

This is also the key distinction between this market cycle and the 2023–2024 era of “buy GPUs to buy AI.” In the early stages, generative AI investments were easier to measure by GPU count, with market focus on how many accelerator cards NVIDIA could deliver and how many training clusters cloud providers could deploy. By 2026, model training remains important, but the weight of inference, agents, and enterprise workflows continues to rise, making system efficiency increasingly dependent on whether compute, memory, network, and storage are properly balanced. Even the fastest GPU will idle if memory bandwidth is insufficient; longer context windows and higher concurrency require servers to store and manage more KV cache, parameters, and intermediate states. As a result, the investment logic for AI infrastructure is shifting from “scarce compute chips” to “the entire data pipeline may be scarce.”

The market's reaction was immediate. On September 4, SanDisk rose 11.9%, Micron increased 6.1%, while the S&P 500 index as a whole declined; on September 7, Asian markets continued the momentum, with Samsung Electronics closing up 5.68%, SK Hynix up 8.26%, and Kioxia also showing strong gains. While Astra was undoubtedly the catalyst, for memory stocks across multiple markets to rise in tandem based solely on a single model release, there must already have been underlying fundamentals supporting the move.

Storage is not an "afterthought for GPUs"—it's a new bottleneck being exposed by AI systems.

To understand this, it’s necessary to first distinguish between HBM, DRAM, and NAND. HBM fundamentally addresses the need for being “close enough to the GPU with sufficient bandwidth”; it works in tandem with the GPU or AI ASIC through advanced packaging, directly impacting whether high-end accelerators can sustain full computational throughput. Server DRAM serves as a broader memory pool, supporting both the CPU and caching/scheduling demands during inference. NAND and enterprise SSDs handle larger-scale persistent data, including training datasets, vector databases, model checkpoints, and tiered storage. As AI evolves from one-off question-answering to continuous, autonomous agents, the coordinated demand for these three storage layers becomes increasingly evident.

TrendForce’s data released on September 7 is highly representative: Global DRAM industry revenue is projected to increase by 59.5% quarter-over-quarter in the second quarter of 2026, reaching approximately $154.73 billion, driven by significant increases in traditional DRAM contract prices as well as rising demand for HBM3E, LPDDR5X, and high-capacity RDIMMs. The firm also forecasts that traditional DRAM contract prices may rise another 13% to 18% quarter-over-quarter in the third quarter. More notably, supplier inventories remain at historically low levels, with new supply clearly favoring the server market. This indicates that the current improvement in memory industry profitability is not merely based on “AI-themed valuation,” but is being realized through price increases, revenue growth, and product mix adjustments that reflect actual supply-demand imbalances.

This is precisely what distinguishes memory chip companies from general AI-themed stocks. As manufacturers shift more advanced DRAM capacity toward HBM and server products, the supply available for traditional DRAM is also constrained; when AI servers demand even greater memory capacity, HBM and standard server DRAM effectively compete for the same wafers, advanced processes, and downstream manufacturing resources. As a result, AI-driven demand is no longer just about increased demand for a single high-end product—it may trigger a chain reaction: “rising high-end demand → reallocation of production capacity → simultaneous tightening of other server memory supplies.” For investors, this dynamic is more significant than simply tracking which company secures orders for a specific generation of HBM, because the former determines whether the entire industry’s profit pool can sustainably expand.

Why are "storage transactions" now more resilient than simple "computing power transactions"?

First, the market is seeking segments within AI capital expenditures that offer greater profit elasticity. While GPUs remain core assets, the growth trajectory of leading GPU manufacturers has been thoroughly analyzed, and marginal surprises now increasingly depend on higher shipment volumes and stronger product cycles. In contrast, the memory industry has undergone multiple rounds of oversupply, causing major manufacturers to naturally be more cautious about large-scale capacity expansion. Should demand surge unexpectedly, short-term supply elasticity becomes even more constrained, making prices and profits more sensitive to changes in demand. The significant rise in DRAM prices and industry revenue since 2026 exemplifies this cyclical elasticity.

Second, the "compute-to-memory demand" for AI workloads has not decreased linearly due to improvements in model efficiency. While more efficient models may reduce the computation required for a single task, if lower costs lead to more frequent calls, longer contexts, higher concurrency, and greater numbers of agents, the total memory consumption and data movement may still increase. This is why we cannot simply equate "models becoming more efficient" with "hardware demand reaching a peak." During the AI infrastructure development phase, efficiency gains often make previously uneconomical applications viable; ultimately, demand depends on which grows faster—the decline in unit cost or the increase in usage volume.

Third, concentration on the supply side makes it easier to establish price discipline in high-end storage. The global advanced DRAM market has long been dominated by Samsung, SK Hynix, and Micron, while HBM adds additional barriers such as advanced packaging, yield rates, thermal design, and customer certification. Samsung has already announced that HBM4 has entered mass production and begun commercial shipments as of February 2026, and SK Hynix also confirmed in its second-quarter earnings report that HBM4 is now being shipped at scale. Although technological competition will continue to intensify, the number of suppliers capable of simultaneously meeting performance, yield, and major customer certification requirements remains limited in the short term—making the supply dynamics of HBM fundamentally different from those of conventional mature-node chips.

There’s another often-underestimated shift: AI demand is no longer defined solely by NVIDIA GPUs. As Google TPUs and custom ASICs from major cloud providers scale up, high-bandwidth memory (HBM) is evolving from a “NVIDIA配套 product” into a common foundational component for advanced AI accelerators. In other words, even if the GPU market share changes in the future, as long as total AI compute demand continues to grow and the memory capacity per AI chip increases, HBM demand can still be supported by a broader range of accelerator types. For memory manufacturers, this means the sources of demand may become even more diversified than in the previous phase.

However, the claim that "more than half of AI capital expenditures in 2027 will be directed toward memory" needs to be viewed with caution.

A striking figure has been widely circulated in the market: Goldman Sachs and Morgan Stanley project that global AI capital expenditures will reach $1.3 trillion to $1.5 trillion by 2027, with more than half allocated to memory infrastructure. This figure effectively captures current market sentiment, and some market reports have indeed adopted this wording. However, upon strict verification against the original public sources, it should not be accepted as an established fact without reservation.

Publicly available materials from Morgan Stanley show that its research team has repeatedly and significantly raised its capital expenditure forecasts for AI and hyperscale cloud providers as early as 2026. In an official podcast released in August, Morgan Stanley noted that estimated capital expenditures for 2027 had risen to approximately $1.2 trillion to $1.3 trillion, substantially higher than previous projections; its core assessment is that computing demand continues to outstrip existing supply, and the scale and pace of the capital expenditure cycle remain consistently above expectations. While this trend itself is highly significant, it does not directly imply that more than half of it is attributable to memory.

The concepts of “computers and peripheral devices,” “server hardware,” and “memory infrastructure” must be separated. GPU, CPU, networking equipment, complete servers, storage systems, and memory all constitute part of data center capital expenditures. Equating the entire equipment budget directly with storage spending would significantly overstate the proportion attributable to memory. Therefore, evaluating whether the storage thesis holds does not require reliance on a ratio that may be inflated during secondary dissemination. More robust evidence is already evident in industry data: rapid growth in DRAM revenue, continued price increases in contracts, low inventory levels, priority allocation of servers, HBM4 entering mass production, and ongoing upward revisions in cloud provider capital expenditures. Building a logical case using these variables is far more reliable than citing a seemingly compelling but ambiguously defined figure.

The real main theme is not "all stored stocks rising together."

The most common misconception in this market cycle is grouping HBM, DRAM, NAND, and hard drive manufacturers all under a single “AI storage” narrative. In reality, HBM is most directly driven by GPU and AI ASIC shipments, with its market leadership determined by technological barriers and customer certifications; server DRAM is more influenced by data center expansion, CPU count, and memory consumption for inference workloads; while NAND, though also supported by growing AI data volumes, faces stronger volatility due to supply expansion rates, enterprise SSD adoption levels, and price cycles. The significant rallies in SanDisk and Kioxia indicate that the market is extending the AI narrative from HBM further into NAND, but their ultimate paths to earnings realization are not equivalent to HBM’s.

This is why the phrase “storage becoming the new main theme” should not be interpreted as a mass migration of capital away from GPUs, but rather as a shift in how AI infrastructure is priced. In the past, the market only needed to ask one question: How many GPUs are there? Now, at least several follow-up questions must be asked: How much HBM is paired alongside each GPU? How much DRAM is required per server? What scale of SSDs is needed for the data center? Can the network connect all these nodes? And can the power supply sustain the entire facility? As AI model complexity continues to rise, any shortage in one component can prevent other expensive equipment from operating at full capacity, and the bottleneck component naturally gains stronger pricing power.

Memory is also one of the most typical cyclical products in the semiconductor industry. Today’s shortages will inevitably stimulate tomorrow’s capital expenditures. Once new capacity comes online en masse between 2027 and 2028, while AI-related capex growth begins to slow, price elasticity could reverse rapidly. Meanwhile, Chinese memory manufacturers continue to expand; although they face clear technological constraints in advanced equipment and high-end HBM, competition in traditional DRAM and NAND markets will intensify. After the significant price surge in 2026, some overseas memory stocks have already priced in highly optimistic assumptions regarding prices and profits. Any shortfall in contract price increases, delays in HBM customer certifications, or reductions in cloud provider capex could trigger sharp corrections.

Therefore, I prefer to define the market conditions following Astra as the "AI infrastructure entering a phase of catching up on gaps," rather than simply declaring that "GPU demand is fading and storage is taking over." GPUs remain the core of computation, networks determine cluster efficiency, and power and data centers continue to be physical constraints; it’s just that as models evolve from chat tools to complex agents, the market is beginning to realize that computing power is not just a story about a single chip, but an entire chain spanning computation, memory, interconnectivity, and storage. Whichever link is most constrained will attract capital and profits first.

If this assessment holds, what truly matters going forward is not how many more days Astra’s stock will rise, but whether the storage industry can consistently demonstrate over several quarters three key points: that demand growth continues to outpace supply expansion, that price increases can sustainably translate into cash flow and capital returns, and that the supply of next-generation HBM and high-capacity server memory remains concentrated among a few firms with technological and customer barriers. Only if all three of these conditions persist will the “storage theme” evolve from a trade ignited by new product launches into a more enduring reallocation of profits within the AI capital expenditure cycle.

Reference materials

1. OpenAI, "Release Notes: Introducing GPT-6 Astra", September 3, 2026.
2. Barron’s, “Micron Stock Closes Above $1,000. Why It’s Not What It Seems,” September 4, 2026.
3. Yonhap News Agency, “OpenAI craze drives Samsung Electronics and SK Hynix higher,” September 7, 2026.
4. MarketWatch, “Why the launch of OpenAI’s latest Astra model reignited the memory-chip trade,” September 8, 2026.
5. TrendForce, “DRAM Industry Revenue Rises 59.5% QoQ in Q2 2026 as Supply Expansion Continues to Lag Demand Growth”, September 7, 2026.
6. Samsung Global Newsroom, “Samsung Ships Industry-First Commercial HBM4 with Ultimate Performance for AI Computing,” February 12, 2026.
7. SK hynix Newsroom, “SK hynix Announces 2Q26 Financial Results,” July 29, 2026.
8. Morgan Stanley, "Thoughts on the Market: AI Financing and the Evolution of Credit Markets," August 21, 2026.