How and When the Memory Chip Shortage Will End

United States - Ekhbary News Agency

How and When the Memory Chip Shortage Will End

If it feels like everything in technology is about AI these days, that’s because it is. Nowhere is this more evident than in the computer memory market. The demand and profitability for the type of Dynamic Random-Access Memory (DRAM) used to power Graphics Processing Units (GPUs) and other accelerators in AI data centers are so immense that they are diverting supply away from other applications, causing prices to surge dramatically. According to Counterpoint Research, DRAM prices have already climbed by 80-90% so far this quarter.

The leading AI hardware companies report that they have secured their chip supplies through 2028. However, this leaves everyone else—manufacturers of PCs, consumer gadgets, and all other devices requiring temporary storage for billions of bits—scrambling to cope with scarce supply and inflated costs. How did the electronics industry find itself in this predicament, and more importantly, how will it extricate itself?

Economists and memory experts point to a confluence of factors: the industry's historic boom-and-bust cycle in DRAM and an AI hardware infrastructure build-out of unprecedented scale. Barring a significant collapse in the AI sector, it is projected that years will pass before new capacity and technological advancements can bring supply into alignment with demand. It's even possible that prices will remain elevated even after equilibrium is achieved.

To grasp the dynamics of this situation, understanding the primary driver of supply and demand fluctuations is crucial: High-Bandwidth Memory (HBM). HBM represents the DRAM industry's strategy to circumvent the slowing pace of Moore's Law by employing 3D chip packaging technology. Each HBM chip comprises up to 12 thinly sliced DRAM chips, known as dies. Each die contains numerous vertical connections called through-silicon vias (TSVs). These dies are stacked atop one another, interconnected by arrays of microscopic solder balls aligned to the TSVs. This DRAM stack—approximately 750 micrometers thick, resembling a brutalist office block more than a slender tower—is then mounted on a base die, which manages data transfer between the memory dies and the processor.

This sophisticated technological assembly is positioned within a millimeter of a GPU or other AI accelerator, linked by as many as 2,048 micro-scale connections. HBMs are typically attached to two sides of the processor, integrating the GPU and memory into a single packaged unit. The objective behind this close proximity and high-speed connection to the GPU is to overcome the "memory wall"—the energy and time barrier encountered when trying to feed the terabytes per second of data required by large language models (LLMs) into the GPU. Memory bandwidth is a critical bottleneck limiting the operational speed of LLMs.

While HBM technology has been available for over a decade, and DRAM manufacturers have continuously enhanced its capabilities, its importance to GPUs has surged in tandem with the increasing size of AI models. This rise in significance has come at a considerable cost. SemiAnalysis estimates that HBM typically costs three times more than other memory types and accounts for 50% or more of the total cost of a packaged GPU.

Industry observers concur that the DRAM sector is inherently cyclical, characterized by periods of substantial booms followed by severe busts. With new fabrication plants (fabs) costing $15 billion or more, companies are understandably hesitant to expand capacity. They often only possess the financial resources to do so during peak market conditions, as explained by Thomas Coughlin, a storage and memory expert and president of Coughlin Associates. However, the construction and commissioning of such a fab can take 18 months or longer, effectively ensuring that new capacity comes online well after the initial demand surge has passed. This often leads to market saturation and depressed prices.

The roots of the current cycle, according to Coughlin, trace back to the chip supply panic during the COVID-19 pandemic. To mitigate supply chain disruptions and support the rapid shift to remote work, hyperscale data center operators like Amazon, Google, and Microsoft procured vast inventories of memory and storage, artificially inflating prices. Subsequently, as supply chains normalized and data center expansion slowed in 2022, memory and storage prices plummeted. This downturn persisted through 2023, compelling major memory and storage firms, including Samsung, to cut production by as much as 50% in an effort to prevent prices from falling below manufacturing costs. This was an unusual and desperate measure, as companies typically rely on running their facilities at full capacity to recoup their investments.

Following a recovery that began in late 2023, "all memory and storage companies became very wary of increasing their production capacity again," Coughlin notes. "Consequently, there was little to no investment in new production capacity throughout 2024 and most of 2025."

This lack of new investment is now colliding head-on with a massive surge in demand fueled by new data centers. Globally, nearly 2,000 new data centers are either planned or under construction, according to Data Center Map. If all these facilities are completed, it would represent a 20% increase in the global supply, which currently stands at approximately 9,000 facilities.

Should the current build-out continue at its projected pace, McKinsey predicts that companies will invest $7 trillion by 2030. The bulk of this investment—$5.2 trillion—is earmarked for AI-focused data centers. Within that segment, $3.3 trillion is expected to be allocated to servers, data storage, and network equipment.

Nvidia, the GPU manufacturer, has been the most significant beneficiary of the AI data center boom. Its data center business revenue soared from just under $1 billion in the final quarter of 2019 to $51 billion in the quarter ending October 2025. During this period, its server GPUs have not only demanded exponentially more gigabytes of DRAM but also an increasing number of DRAM chips. The recently launched B300 model utilizes eight HBM chips, each a stack of 12 DRAM dies. Competitors are largely mirroring Nvidia's adoption of HBM; for instance, AMD's MI350 GPU also incorporates eight 12-die HBM chips.

With such overwhelming demand, HBM is increasingly contributing to the revenue streams of DRAM manufacturers. Micron, the third-largest producer behind SK Hynix and Samsung, reported that HBM and other cloud-related memory constituted 17% of its DRAM revenue in 2023, a figure that jumped to nearly 50% in 2025. Micron forecasts the total HBM market to expand from $35 billion in 2025 to $100 billion by 2028, surpassing the entire DRAM market size of 2024.

Ekhbary News Agency