Over the past two years, AI investment has focused almost entirely on GPUs, with NVIDIA emerging as the most prominent beneficiary. However, as large model parameter sizes continue to grow, a deeper issue is coming to the forefront: while computing power is increasing, data throughput and storage capabilities are now becoming the new bottlenecks.
In the latest wave of AI infrastructure expansion, the market is gradually recognizing a significant shift: no matter how powerful GPUs become, they still require an ultra-fast "data supply system" to maintain operational efficiency. This realization is at the heart of the recent revaluation of HBM (High Bandwidth Memory).
Micron’s latest earnings report shows the company not only far exceeded market expectations but also secured approximately $22 billion in long-term supply agreements. Management made it clear that AI storage demand remains extremely tight and could persist well beyond 2027. At the same time, SK Hynix, leveraging its leadership in HBM, has overtaken Samsung Electronics to become one of South Korea’s most valuable publicly traded companies.
With leading storage companies from different markets both signaling robust growth, a key question is emerging: Is HBM becoming the most reliable growth track in the AI era?
The Essence of HBM: The GPU’s "High-Speed Memory System"
To understand the value of HBM, it’s important to first grasp the structure of AI computing.
During large model operations, GPUs handle computation, but true efficiency depends on whether data can be delivered rapidly and continuously to the processing units. As model parameters grow, traditional DRAM can no longer meet bandwidth demands, paving the way for HBM.
You can think of an AI chip system as follows:
- GPU = Computing engine
- HBM = High-speed cache and memory system
- Data center storage = External data warehouse
As models move from training to inference, data access frequency increases even further, making HBM even more critical. This is why the market increasingly refers to HBM as the "key infrastructure of AI factories."
From a technology perspective, HBM’s stacked architecture dramatically increases bandwidth density, enabling GPUs to access data more efficiently, reducing latency, and boosting overall throughput. This structural optimization isn’t just an upgrade—it’s a fundamental redesign of traditional storage architecture.
Micron and SK Hynix: Two Main Threads in the AI Storage Cycle
Today’s global HBM market is highly concentrated, dominated by SK Hynix, Samsung Electronics, and Micron. SK Hynix holds the leading market share in HBM and is particularly well positioned with AI customer orders.
SK Hynix’s advantage comes from its early bet on HBM technology, with its products deeply integrated into the NVIDIA AI chip ecosystem. Recent data shows that its HBM business has driven significant profit growth and helped it surpass Samsung in market capitalization in Korea.
Micron, on the other hand, reflects cyclical trends in the US market. Its latest earnings report revealed not only strong revenue and profit beats but also sent a clear supply-demand signal: AI storage orders are now locked in for the long term, with some customers signing multi-year procurement contracts.
This points to a fundamental shift: the storage industry is moving from being a "cyclical commodity" to one driven by "structural demand." In the past, storage pricing was dictated by supply-demand cycles, but now, more and more demand comes from the long-term expansion of AI infrastructure itself.
Is HBM Entering a "Supercycle"?
Debate around HBM centers on two main questions: Is demand sustainable? And will supply quickly catch up?
On the demand side, AI is shifting from training to inference, and inference workloads require always-on, high-frequency access, creating a more stable and long-term need for storage bandwidth. Meanwhile, agents, long-context models, and enterprise AI applications are expanding rapidly, further increasing data access frequency.
On the supply side, HBM manufacturing is complex, with slow yield improvements and a heavy reliance on advanced packaging and high-end manufacturing capabilities. This means capacity expansion is lagging behind demand growth. Industry research also suggests HBM will likely remain in tight supply for several years, with some manufacturers already locking in capacity through 2026.
However, there’s a second layer of concern: if supply expansion accelerates, will prices fall? The storage industry has historically experienced similar cycles, so whether HBM can escape its cyclical nature remains a subject of debate.
Asset Logic Shift: From "GPU-Led" to "Storage Revaluation"
Previously, the logic for AI investment in the market was straightforward:
Whoever controls computing power enjoys the greatest premium.
But now, the structure is changing:
- GPU → Still core, but growth is becoming more concentrated
- HBM → Emerging as the new source of growth flexibility
- Data centers → Entering a phase where infrastructure pricing logic applies
This shift means capital markets are beginning to break down the AI value chain, rather than simply pricing around a single dominant player. Especially after both Micron and SK Hynix signaled strong growth, the market is beginning to embrace a new narrative: the AI bottleneck is shifting from "insufficient computing power" to "insufficient data flow capability."
Gate Stock Trading: 24/7 Access to the AI Storage Value Chain
As AI storage draws global investor attention, demand for cross-market trading is rising. Core companies like Micron, NVIDIA, and SK Hynix are listed in different markets, making it difficult for a single trading session to capture all market movements.
Against this backdrop, Gate’s stock trading platform has upgraded to a 24/7 trading model, supporting US, Hong Kong, and Korean stocks, and covering key players in the AI storage value chain.
Users can trade the following from a single account:
- US stocks: AI infrastructure companies like Micron and NVIDIA
- Korean stocks: Storage leaders such as SK Hynix and Samsung Electronics
- Hong Kong stocks: AI server, optical module, and new economy companies
Gate also supports trading with USDT, reducing the cost of moving funds across markets and making global asset allocation more flexible.
For the highly interconnected and event-driven AI value chain, round-the-clock trading means investors can respond more quickly to earnings reports, supply-demand shifts, and value chain updates.
Conclusion: HBM Isn’t the "Endgame"—It’s the Beginning of an AI Infrastructure Revaluation
Whether HBM will become the most profitable track in the AI era is still an open question. What’s certain is that it’s no longer just a "supporting technology"—it’s becoming an indispensable part of AI infrastructure.
The changes in Micron’s earnings and SK Hynix’s market capitalization both reflect the same trend: AI’s value is shifting from the "application layer" back down to the "infrastructure layer."
In this wave of structural change, the storage industry is likely still in the early to mid-stages of its cycle—not at the end.
FAQs
What’s the difference between HBM and traditional DRAM?
HBM is high-performance memory that uses a stacked architecture to boost bandwidth density, mainly for AI GPUs and high-performance computing, while DRAM is geared more toward general-purpose computing.
Why does Micron’s earnings report impact the entire AI sector?
Because Micron is one of the world’s major storage suppliers, its performance directly reflects real demand for storage chips in AI data centers.
Why is SK Hynix a leader in HBM?
SK Hynix invested early in HBM technology and is deeply embedded in the AI chip customer ecosystem, giving it an edge in the high-end storage market.
Can HBM price increases continue?
In the short term, prices remain supported by tight supply and demand, but long-term sustainability depends on how fast capacity can expand and whether alternative technologies develop.
When is Gate’s 24/7 stock trading most useful?
It’s ideal for tracking AI earnings, chip market trends, and cross-market opportunities, allowing investors to respond quickly to global market changes.
