Samsung’s second-quarter 2026 profit of $58.5 billion has put the company ahead of Nvidia’s $53 billion in the same period, underscoring how the race to supply memory for artificial intelligence systems has reshaped the global semiconductor market.
The South Korean technology group drew more than 94% of those earnings from its AI memory business, led by high-bandwidth memory, or HBM. Gross margins in the division reached 52%, according to the figures cited, reflecting a market in which demand from AI data centers is still running well ahead of available supply.
The result marks one of the clearest signs yet that advanced memory has moved from a supporting role in the chip industry to one of its most profitable segments. HBM is a specialized form of memory used alongside advanced processors in AI servers. It allows large volumes of data to move quickly between memory and computing chips, an essential function for training and running large machine-learning models.
Yet the record earnings did not translate into a rally in memory stocks. Samsung shares fell 9% on earnings day, while SK Hynix dropped 15%, as traders focused less on current profits and more on the risk that today’s tight supply could eventually turn into another semiconductor downcycle.
That reaction captured the central tension now facing the industry: memory producers are making historic profits, but traders are trying to judge whether these gains represent a lasting structural shift or the peak of a pricing cycle.
Hbm becomes the center of the chip boom
The global HBM market is controlled mainly by Samsung, SK Hynix and Micron. The two South Korean producers together hold roughly 60% of total capacity, giving them unusual pricing power at a time when AI chip demand is expanding rapidly.
The market’s structure is especially tight because HBM is far more difficult and capacity-intensive to produce than standard memory. Producing one gigabyte of HBM consumes about four times the wafer capacity required for regular DRAM. That means every additional HBM order can reduce the amount of factory space available for conventional memory used in personal computers, smartphones and other consumer devices.
This shift has pushed memory costs higher across the electronics sector. HBM prices rose 90% in the first quarter, climbed another 50% to 60% in the second quarter, and are expected to rise by a further 20% in the third quarter. The price pressure has already spread beyond AI servers, raising the cost of memory components used in everyday hardware.
Samsung’s annual profit has now exceeded the combined total of its previous four decades, according to the reported figures. That milestone reflects both the scale of the current AI buildout and the unusually concentrated nature of the advanced memory market.
For South Korea, the shift has broader economic effects. Higher margins at Samsung and SK Hynix have lifted income across parts of the technology sector, while also contributing to rising prices for consumer electronics. Margins for leading Korean memory manufacturers are reported to have reached 72 cents on every $1 sold in some advanced products, a level rarely seen in the historically cyclical memory business.
Traders look past record profits
The sharp decline in Samsung and SK Hynix shares after the earnings release showed that traders are not simply rewarding current profitability. Instead, they are weighing the danger that fast-rising prices could lead to excessive production commitments, slower demand growth or a future inventory build.
Memory has long been one of the most cyclical areas of technology. Periods of shortage often lead producers to expand capacity aggressively. If supply later arrives just as demand slows, prices can fall quickly and margins can collapse. That history remains fresh for many market participants, even though the AI cycle has changed the demand profile.
Analysts watching the latest selloff described it as a cyclical adjustment after two quarters of extreme price growth. Concerns about inventory buildup remain, especially as large cloud-computing customers reassess capital spending plans after a period of heavy AI infrastructure purchases.
There are also fears that growth in AI capital expenditure could slow from its recent pace. Even a modest slowdown can weigh on sentiment because memory stocks have priced in strong demand for years ahead. When share prices reflect perfect execution, any sign of uncertainty can trigger a sharp pullback.
Still, the physical limits of production continue to support profitability. Current HBM demand is expanding three to five times faster than supply, and new fabrication plants are not expected to begin meaningful production until around 2030. That gap suggests that shortages may persist even if some customers reduce near-term spending.
Supply deficit may deepen before it eases
Kwak, a market watcher cited in the latest industry discussions, warned last Friday that the technology sector could enter its worst supply deficit in history next year. He said aggressive factory expansion plans are still failing to match the volume of customer orders arriving from AI developers, cloud providers and hardware makers.
The warning reflects a key change in semiconductor manufacturing. Chip makers are devoting more factory floor space to advanced computing components, including HBM and related packaging technologies. As a result, fewer resources are available for older standard hardware.
That shift is creating a ripple effect. Standard DRAM, mobile memory and PC components are becoming more expensive because the same factories that once supplied those markets are being redirected toward higher-margin AI products. For consumers, that can mean higher prices for laptops, smartphones, gaming devices and local storage hardware.
The issue is not only the number of factories but also the complexity of the production process. HBM requires advanced stacking, packaging and testing. Capacity cannot be added overnight, and bottlenecks can appear in equipment, materials, skilled labor and power availability.
Even if companies approve new fabrication plants now, they face long construction and qualification timelines. Equipment delivery schedules are tight, and advanced memory fabs must meet demanding performance standards before volume production can begin. That helps explain why new capacity is not expected to provide major relief for several years.
Power costs add another constraint
The memory crunch is being intensified by rising electricity demand from AI computing and digital token systems. The International Energy Agency projects that power usage from machine learning and virtual token systems will reach 1,000 terawatt-hours this year, placing fresh strain on local electricity grids.
That growth is pushing utility bills higher in regions with large data-center clusters and mining operations. Grid operators are preparing to raise power rates in some areas to manage the surge in demand, according to market watchers. Higher power costs can feed back into hardware prices because semiconductor manufacturing, AI training and server operations all require large and reliable electricity supplies.
Digital coin miners are under particular pressure. Proof-of-work networks depend on large amounts of computing power, and many mining firms operate on thin margins when token prices weaken or electricity costs rise. Without fixed power contracts, these firms can see operating costs rise sharply.
Market watchers project that 20% of plants focused on proof-of-work networks may pivot their servers toward machine-learning tasks by late 2027. The shift would reflect the changing economics of computing infrastructure. Where mining once offered the highest return for certain types of hardware, AI workloads are increasingly competing for the same power, space and cooling resources.
For crypto-asset traders, that transition adds another layer of risk. Mining firms without long-term electricity agreements or access to modern computing machines may face growing pressure as grid costs rise and AI operators bid more aggressively for capacity.
Hardware leasing gains attention
Tight server supply has also boosted demand for hardware leasing networks. Server rental rates rose 40% in March and increased another 50% last month for the newest hardware models, according to the figures cited. The rise reflects a market in which companies unable to secure their own AI hardware are turning to rented capacity instead.
Hardware leasing has become a way for some businesses to access advanced computing without waiting for direct equipment deliveries. It also allows owners of scarce servers to benefit from rising rental prices. As long as supply remains constrained, leasing platforms may continue to attract interest from companies that need computing power quickly.
However, this market also carries risks. If new server supply arrives faster than expected, rental rates could fall. If AI demand remains strong, rates may stay elevated. The outcome depends on the same forces driving the broader memory debate: demand growth, supply expansion, electricity availability and the timing of new factories.
The hardware shortage has also drawn attention from prominent short sellers. Michael Burry has reportedly taken a large short position against memory makers, betting that the current shortage has already reached its peak and that an oversupply cycle could damage profit margins by 2028 as newly funded factories begin to open.
That bearish view contrasts with the argument that HBM shortages could last well into the next decade. The difference comes down to timing. If demand slows before new capacity arrives, margins could weaken. If AI workloads continue to expand faster than production, the current profitability of memory makers may prove more durable than previous cycles.
New standards aim to ease bottlenecks
Industry groups are also trying to reduce production pressure through design changes. JEDEC, the standards organization for microelectronics, released new SPHBM4 design rules last week aimed at lowering final packaging costs.
The updated design uses a narrower 512-bit layout that can mount directly onto standard boards, which could speed up final assembly lines and reduce some integration costs. If widely adopted, the standard may help make advanced memory easier to deploy across a broader range of systems.
Still, standards alone cannot solve the shortage. They may improve efficiency and lower some production barriers, but they do not immediately create more wafer capacity, skilled labor or electricity supply. The industry still faces a multi-year challenge in matching AI-driven demand with physical manufacturing capability.
A structural shift, but not without cycle risk
The latest financial data points to a technology sector undergoing structural repricing rather than a simple rebound from a weak cycle. AI demand has changed how chip makers allocate factory space, which products receive priority and how much customers are willing to pay for advanced memory.
At the same time, the semiconductor industry has not repealed its old rules. High margins invite expansion. Expansion can eventually create oversupply. Stock prices may fall even when profits are at record highs if traders believe the best pricing environment is already reflected in valuations.
For now, the core supply picture remains tight. HBM demand is still growing far faster than supply, and meaningful new fabrication capacity is years away. That imbalance supports high profitability for Samsung, SK Hynix and Micron, even as share prices swing on concerns about capital spending and future inventory levels.
The next phase will depend on whether AI demand continues to absorb every additional unit of advanced memory that producers can make. If it does, the current repricing of memory could become one of the defining shifts in the technology sector. If demand cools before new factories are fully loaded, the same companies posting record profits today could face the familiar pressure of a cyclical downturn.
For consumers and businesses, the more immediate effect is simpler: advanced computing is making memory more valuable, more scarce and more expensive. That change is already shaping the cost of servers, personal computers, mobile devices, digital infrastructure and power consumption across the global technology economy.
Explore how institutional flows shape crypto like AI reshapes chips with Toobit’s ETF market dynamics guide.
Disclaimer: The content on this page is provided for general informational purposes only and does not represent the views or financial advice of Toobit. We make no guarantees regarding the accuracy or completeness of this information and shall not be held liable for any errors, omissions, or outcomes resulting from its use. Investing in digital assets involves risk; users should independently evaluate their financial situation and the risks involved. For further details, please consult our Terms of Service and Risk Disclosure.
