By Sithembile Bopela
In early 2026, the global technology supply chain encountered a severe structural bottleneck in random access memory (RAM). This global shortage of memory chips, at first glance, may appear to be another cyclical swing in the semiconductor industry. However, where previous shortfalls have been driven by supply/end-user-demand mismatch, this shortage is being driven by AI infrastructure appetite, not consumer demand. Memory manufacturers are thus redirecting capacity towards high performance products used in data centres, leaving less supply for everyday electronics. Prices for standard memory have risen sharply, margins have shifted unevenly across the technology stack, and investment risks and opportunities are being re priced.
Lest we forget
There are four main categories of semiconductor chips: memory chips, microprocessors, standard chips, and complex systems-on-a-chip (or SoCs). The two main types of microchips are the logic chip, which process information to help electronic devices complete their tasks (like the central processing unit (CPU) of a computer), and the memory chip. Memory chips, which can be either NAND Flash or "working memory" chips, known as dynamic random-access memory (DRAM), are used to store information. Whichever is in question, semiconductor chips have a broad range of applications, across industries and across the globe - smartphones, laptops, servers, vehicles, and industrial machines all rely on them. However, investments into the development and manufacture of these chips has lagged the rapid growth in demand for technologies that rely heavily on these tiny engines of the modern world.
Historically, memory supply has been abundant and prices volatile, rising and falling with the global electronics cycle. That pattern has changed. The primary catalyst has been the exponential growth of artificial intelligence (AI) data centres, given the vast amounts of High-Bandwidth Memory (HBM) that advanced AI models require to function efficiently.
To meet this demand, the world's largest memory manufacturers - Samsung (005930 KS), SK Hynix (000660 KS), and Micron (MU US) - also called the "memory triopoly" have redirected their production lines away from standard DDR4 and DDR5 memory modules to focus almost exclusively on HBM.
New fabrication plants take years and billions of dollars to build. As such, the market cannot adjust quickly enough to replace this lost consumer capacity. The outcome is a structural mismatch: demand for memory is accelerating in one part of the market while supply for the rest is shrinking. Consequently, standard memory prices have surged by over 100% in the first quarter of the year, creating a deep and sustained global supply deficit.
This is not a short term disruption. Capacity decisions made today will shape supply conditions well into the latter part of this decade and beyond.
Why this is different from past memory cycles
In prior cycles, demand was largely driven by consumer electronics - PC refresh cycles, new smartphone launches, or updated gaming consoles. That demand tended to fluctuate with economic conditions - easier monetary conditions and higher disposable income promotes household spending. AI infrastructure demand behaves differently. Large cloud providers are committing capital years in advance, signing long term supply agreements to secure memory availability, often prioritising delivery over price.
In 2026 alone, the top US hyperscalers - including Microsoft, Alphabet, Meta, Amazon, and Oracle - are projected to deploy an aggregate $720 billion in capital expenditures to secure critical hardware and facility capacity, with ~75% of this allocated directly to graphic processing units (GPUs), memory clusters, and data centre real estate. These multi-year capacity lock-ins are not speculative research budgets. For these tech giants, the existential risk of falling behind in the AI transition far outweighs near-term margin compression, leading them to absorb premium component pricing and effectively buy out the semiconductor supply chain years in advance just to ensure guaranteed delivery.
On the other hand, HBM production limits how much total memory can be supplied to the market. To meet this multi-year demand, the memory triopoly has committed to unprecedented capex cycles for 2026. Micron recently increased its FY26 capex budget to $25 billion - a significant surge aimed directly at expanding its HBM footprint and advanced packaging lines. SK Hynix, currently holding the dominant share of the global HBM market, has stated its 2026 capital investments will "significantly rise" compared to the previous year (FY25: 30.2 trillion won or ~$20.3 billion), without elaborating on a specific quantum. The firm is accelerating equipment installations at its next-generation M15X fabrication plant in South Korea ahead of schedule, while also deploying nearly $4 billion for a new advanced packaging facility in Indiana. Meanwhile, Samsung is actively restructuring its existing manufacturing footprint, going so far as to discontinue legacy memory production lines to force a targeted 50% increase in its own HBM capacity this year.
However, because constructing and tooling modern semiconductor facilities is incredibly complex, even as manufacturers spend heavily on new facilities, the effective supply of standard memory grows slowly. This combination of durable demand and constrained supply has changed the industry's economics. Prices have risen sharply, inventories remain tight, and profit margins have expanded to levels rarely sustained in the past.
When the chips are down, who wins?
Memory manufacturers: structural winners The most obvious beneficiaries of this environment are the memory manufacturers themselves. Companies such as Samsung, SK Hynix, and Micron control a concentrated global memory market share and currently operate at or near full capacity. Much of their advanced output is already sold out years ahead and pricing power has improved materially, driving stronger earnings visibility - which is atypical for this otherwise volatile growth sector. According to Bloomberg Intelligence, RAM memory prices surged between 200% and 300% y/y in February as data centre demand continues to crowd out consumer product supplies.
Management teams are responding with disciplined capital spending, focused on high return projects linked to AI demand rather than indiscriminate capacity expansion. This reduces the risk of a rapid return to oversupply, which has historically hurt memory pricing. As such, this looks less like a speculative spike and more like a multi year capital cycle, anchored in physical constraints and long term contracts.
Consumer electronics: growing pressure Further downstream, the picture is less favourable. Original equipment manufacturers (OEMs) of smartphones, laptops, and personal computers rely heavily on standard memory components, and are facing rising input costs at a time when consumers are increasingly price sensitive. In many cases, companies must choose between absorbing higher costs or passing them on through price increases or reduced specifications. In 1Q26, worldwide smartphone shipments fell 4.1% y/y to 289.7 million units. The early signs are already visible: higher device prices, slower unit growth, and margin compression. Hardware refresh cycles are stretching out, and some consumers are opting to delay upgrades altogether.
This matters because many well known technology brands sit in this part of the value chain. Their business models were built for an environment of inexpensive, abundant components, however, that assumption no longer holds.
AI infrastructure: high demand, rising costs While the main drivers of the current RAM demand boom, AI infrastructure providers themselves are not immune. Demand remains strong, but the cost of building data centres is rising, and memory now represents a materially larger share of total server costs. For the largest cloud firms this is manageable. Scale, pricing power, and long investment horizons provide flexibility. Conversely, for smaller players, capital intensity is increasing and returns may be more volatile.
Physical constraints matter, again
The memory shortage highlights a broader contention that investors sometimes overlook during technology booms: digital progress is still bound by physical realities. AI is often discussed as a software driven revolution, but, in practice, it requires vast amounts of physical hardware, energy, specialised materials, and manufacturing capacity. Now, memory bandwidth - not algorithms - is emerging as one of the binding constraints. This has very important implications for capital allocation, for one, returns are flowing towards companies that own scarce and 'difficult to replicate' assets. By contrast, businesses built primarily around assembly, branding, or incremental software features face tighter margins. Ultimately, companies with little control over input costs will face ongoing pressure.
In many ways, this resembles earlier infrastructure cycles where the winners are those closest to the bottlenecks - proximity economics.
How long could this last?
As capital investments into the industry play catchup with demand, relief will come, eventually. Memory manufacturers are investing heavily in new production facilities, and some additional supply is expected from 2027 onward. However, timing matters. New fabs take years to become operational, and early output is often limited. At the same time, AI workloads are becoming more memory intensive, particularly as inference and real time applications scale.
Our base case view is that tight conditions persist through at least 2027, with incremental improvement thereafter. For reference, it typically takes 2 to 3 years for a dollar spent on a "cleanroom" to result in a chip sold to a customer - highlighting the massive catch-up phase the industry is currently entering to resolve the structural shortage. This suggests that current dynamics are not fleeting but meaningful for medium term portfolio positioning.
Conclusion
The global RAM shortage is not just a supply story but a a reminder that technological change reforms value chains unevenly throughout generations. AI is accelerating demand for physical infrastructure faster than the world can build it, and memory chips sit at the centre of this tension. Companies that manufacture essential components with limited substitutes are better positioned than those assembling end products. Which is to say, understanding where constraints sit - and who controls them - matters more in this cycle, as do strong fundamentals. While memory stocks will remain cyclical, we view the floor under earnings as being higher than in past cycles.
