AI Memory is becoming the new core of AI infrastructure: HBM super cycle and the revaluation logic of memory chips.

On June 30, 2026, Bitcoin was trading in a narrow range around the $60,000 mark, while Ethereum remained in the $1,600 range. The consolidation in the crypto market did not mask another more defined structural trend—the underlying hardware of AI infrastructure is undergoing a profound power shift.

Over the past decade, memory chips have been viewed as "cyclical products" in the semiconductor industry—demand fluctuating with the inventory cycles of PCs and smartphones, and prices swinging sharply with supply-demand dynamics. But this cognitive framework is being completely shattered by the exponential growth of AI computing power. From large language model training to AI inference, from agentic workflows to autonomous driving, massive data traffic continues to drive demand for high-bandwidth, low-latency, high-capacity memory. Memory has evolved from a "supporting component" to a "core cost item."

The core vehicle of this transformation is AI Memory—a new generation of memory solutions represented by HBM (High Bandwidth Memory). It is no longer an accessory to CPUs or GPUs, but a key bottleneck determining whether AI computing power can be effectively unleashed. Understanding why AI Memory is becoming the new core of AI infrastructure not only relates to judging trends in the semiconductor industry but also directly points to a more practical question: In the memory supercycle, which assets are worth attention? How can one participate in the structural opportunities of global stock markets through platforms like Gate?

Computing Power Races Ahead, Memory Lags Behind: The Digital Truth of the AI Storage Gap

The expansion speed of AI computing power is "crushing" storage supply capacity in an unprecedented way. According to the latest forecast data from Sigmaintell, global AI server shipments in 2026 will reach approximately 3.7 million units, a year-on-year increase of 51.3%. TrendForce estimates that global AI server shipments in 2026 will grow over 28% annually. Regardless of which data source is used, double-digit annual growth in AI server shipments is a highly certain industry trend.

But the truly noteworthy data lies in the change in "per-unit storage capacity." Gartner points out that the DRAM usage per AI server is 8 to 10 times that of traditional servers, and NAND flash usage is more than 3 times. Sigmaintell further estimates that, in terms of GB capacity, DDR storage demand for AI servers in 2026 could increase 105% year-on-year, and HBM demand could increase 110%, with both types of memory products maintaining double-digit growth rates.

The explosion in demand is only half the story. The rigid constraints on the supply side are the structural support that sustains the memory "supercycle." HBM memory chips are about twice the physical size of standard DDR chips, consuming more wafer area. According to SEMI China, the HBM market size in 2026 is expected to grow 58% to $54.6 billion, accounting for nearly 40% of the DRAM market. Even though Samsung, SK hynix, and Micron have shifted 70% of new or allocable capacity toward HBM, the HBM capacity gap remains as high as 50% to 60%.

As of the first quarter of 2026, HBM capacity from the three major manufacturers was fully sold out. Micron management publicly confirmed that the company can only meet approximately 50% to 66% of actual customer demand. Nvidia CEO Jensen Huang clearly stated that the global HBM supply shortage "is not a short-term market fluctuation at all, but a structural industry dilemma that will last for years."

From "Cyclical Product" to "Strategic Asset": Reconstructing the Valuation Logic of Memory Chips

The investment logic of the traditional memory chip industry was built on "cyclical supply-demand fluctuations"—expansion, oversupply, price drops, production cuts, shortage, price increases, repeating the cycle. But the current memory upcycle driven by AI differs fundamentally from historical precedents in multiple dimensions.

Difference 1: Demand driver shifts from consumer electronics to AI infrastructure. In the past, the largest demand sources for DRAM and NAND were smartphones and PCs. But according to Sigmaintell, the proportion of DRAM demand from the smartphone category will drop sharply from 43% in 2024 to 23% in 2027. Meanwhile, by 2026, AI servers' share of total global DRAM shipment capacity will exceed 40%, expected to climb to 49% by 2027 and surpass 50% by 2028. Servers are replacing smartphones as the largest market for memory demand.

Difference 2: Supply expansion faces multiple rigid constraints. Unlike past cycles where manufacturers could quickly respond to demand by building new production lines, HBM capacity expansion faces bottlenecks including capital investment, advanced packaging processes, and fab construction lead times. SK hynix is building the Cheongju M15X wafer fab and establishing a dedicated HBM technology organization, but these capacity additions will take years. With supply response severely lagging, the medium-to-long-term existence of the supply-demand gap has high certainty.

Difference 3: Pricing power shifts from a buyer's market to a seller's market. In traditional memory cycles, downstream OEMs had strong bargaining power, and memory makers often fell into price wars during oversupply. But in this cycle, insufficient supply of high-end products, customers competing for long-term supply guarantees, and manufacturers controlling the pace of low-end expansion have collectively enhanced memory companies' pricing power. J.P. Morgan expects the average price of mixed HBM in 2027 to rise 32% year-on-year, hitting a record high.

Micron Earnings: The Strongest Validation of the Memory Supercycle

If the above logic is still at the "expectation" level, then Micron Technology's fiscal third-quarter 2026 earnings report pushes that logic into the "validation" stage.

For the fiscal third quarter ending May 28, 2026, Micron reported revenue of $41.46 billion, up 346% year-on-year and 74% sequentially, setting a record for the fifth consecutive quarter. This figure significantly exceeded market expectations of $35.84 billion. The company expects fiscal fourth-quarter revenue to reach $49 billion to $51 billion.

Even more noteworthy is the leap in profitability. Micron's third-quarter gross margin rose to 85%, with an operating margin of 81%. Among them, the DRAM business operating margin reached 81%, and the NAND business operating margin reached 78%. This profitability is far higher than the norm in traditional memory cycles—where DRAM and NAND were typically seen as highly commoditized, price-volatile markets where manufacturers quickly fell into margin compression during supply-demand reversals.

Micron also disclosed that cumulative sales of HBM4 have reached approximately $1 billion. The company's entire HBM capacity for 2026 has been fully booked. In terms of stock performance, as of 0:00 Beijing time on June 30, Micron Technology rose 1.14% to $1,145.28. Although it once fell 9.6% intraday on news of a DRAM antitrust lawsuit, buy orders returned in the final session, eventually turning positive—this price action itself reflects that the market's confidence in the long-term logic of memory has not been shaken by short-term disruptions.

From Industry Trends to Investment Mapping: Which Targets Are Worth Attention?

The structural trend of AI Memory is not an isolated stock event but a revaluation covering the entire industry chain from upstream wafer manufacturing to downstream data centers. The following main lines are worth focusing on:

HBM leader tier. SK hynix holds a leading position in the HBM market, with a shipment share of 62% as of Q2 2025 and a revenue share of 57% in Q3. Goldman Sachs expects SK hynix to maintain a dominant position in HBM3 and HBM3E at least until 2026, with an overall HBM market share above 50%. UBS predicts SK hynix could capture about 70% of the HBM4 market for Nvidia's next-generation "Rubin" platform. Micron is also catching up quickly, with HBM4 already entering mass production and supply.

AI server and data center industry chain. Sigmaintell expects global AI server shipments to approach 5 million units by 2028. The continued expansion of AI server shipments directly drives demand for DRAM, NAND, HBM, and enterprise SSDs. By 2026, eSSD is expected to surpass smartphones as the largest application area for NAND Flash.

Semiconductor equipment and advanced packaging. HBM capacity expansion relies on breakthroughs in advanced packaging technology. SK hynix is strengthening its cooperation with TSMC in advanced packaging. Wafer fab expansion directly benefits semiconductor equipment vendors.

Gate Stock Trading: How to Participate in Global Memory and AI Infrastructure Investment?

For investors looking to directly participate in the above structural opportunities, Gate's stock trading service provides a low-barrier, high-efficiency path.

In June 2026, Gate successively launched trading services for U.S. stocks (June 1), Hong Kong stocks (June 11), and Korean stocks (June 22). On June 23, Gate further upgraded stock trading to 7×24 hours all-day trading, covering pre-market, regular, after-hours, overnight, and weekend market closure periods.

Regarding AI Memory-related targets, Gate already covers core assets such as Nvidia (NVDA), Micron Technology (MU), SK hynix (000660), and Samsung Electronics (005930). Transactions are settled in USDT, eliminating the need for bank transfers or currency exchange. Users can transfer USDT from their spot account or unified account to their stock account to place orders.

It is important to note that 7×24-hour trading does not mean 7×24-hour liquidity. Liquidity may be relatively low during overnight and weekend closed periods, bid-ask spreads and price volatility may expand; price gaps may also occur due to accumulated market news across trading sessions. U.S. stocks, Hong Kong stocks, and Korean stocks are subject to different trading calendars and market rules. Investors must fully understand the associated risks and make prudent decisions.

Conclusion

2026 is becoming one of the most iconic years in the history of the memory chip industry. The global memory chip market size is expected to reach approximately $975 billion, up about 3.2 times year-on-year. Memory chips have grown 250% year-on-year, exceeding $800 billion in scale. BofA Securities describes 2026 as a semiconductor "supercycle" comparable to the boom of the 1990s.

The essence of this supercycle is not simply a supply-demand mismatch, but a systematic restructuring of the underlying hardware of infrastructure driven by the AI computing power revolution. AI Memory—high-bandwidth, high-capacity memory solutions represented by HBM—is moving from a supporting role to center stage, becoming a key variable in determining whether AI computing power can continue to expand.

For investors, understanding the significance of this structural change may be more important than chasing short-term price fluctuations. The valuation logic of memory chips is shifting from "cyclical products" to "strategic assets," and the duration of this shift may far exceed current market expectations.

FAQ

Q1: What is the difference between HBM and traditional DRAM? Why does AI need HBM?

HBM (High Bandwidth Memory) uses 3D stacking technology to vertically integrate multiple DRAM chips, greatly increasing data bandwidth and reducing power consumption. Traditional DRAM cannot meet the large-scale parallel computing requirements of AI training and inference in terms of bandwidth and energy efficiency. HBM's high-bandwidth characteristics make it the ideal companion memory for GPUs and AI accelerators, serving as a key technical path to break through the "memory wall" bottleneck.

Q2: What is the expected market size for memory chips in 2026?

Multiple institutions have given different estimates. Counterpoint Research expects the global memory chip market size in 2026 to be approximately $975 billion. TrendForce estimates the combined output value of DRAM and NAND in 2026 at $889.3 billion. WSTS forecasts memory in 2026 to grow about 250% year-on-year, with market size exceeding $800 billion. Although the statistical scopes of different institutions vary, the magnitude all points to a range of $800 billion to $1 trillion.

Q3: How large is the HBM capacity gap? How long will it last?

According to SEMI China, even though the three major manufacturers have shifted 70% of new capacity toward HBM, the HBM capacity gap remains as high as 50% to 60%. As of Q1 2026, HBM capacity from the three major manufacturers was fully sold out. J.P. Morgan expects the supply-demand gap to persist throughout 2028. Jensen Huang stated this is not a short-term fluctuation but a structural industry dilemma lasting several years.

Q4: Which AI memory-related stocks can be traded on Gate?

Gate stock trading covers U.S. stocks, Hong Kong stocks, and Korean stocks. AI memory-related targets include: U.S. stocks Nvidia (NVDA), Micron Technology (MU); Korean stocks SK hynix (000660), Samsung Electronics (005930); on the Hong Kong stock side, although there are no pure memory chip targets, tech giants like Tencent Holdings (00700) are important downstream buyers of HBM. Gate supports 7×24-hour trading, settled in USDT.

Q5: What is the specific pull effect of AI server shipment growth on memory demand?

According to Sigmaintell, DDR storage demand for AI servers in 2026 increased 105% year-on-year, and HBM demand increased 110%. In 2026, AI servers' share of total DRAM demand will exceed 40%, expected to surpass 50% by 2028. AI servers are not only growing in shipment volume, but the storage capacity per unit is also continuously increasing, creating a dual pull in terms of "quantity" and "quality."