AI Boom Fuels DRAM Shortage and Price Surge

If it seems today that everything in technology revolves around AI, that’s because it is. And nowhere is this more true than in the computer memory market. The demand and profitability for the type of DRAM used to power GPUs and other accelerators in AI data centers is so enormous that it is diverting memory supply to other uses and sending prices skyrocketing. According to Counterpoint Research, DRAM prices are up 80-90% so far this quarter.

The biggest AI hardware companies say they have secured their chips until 2028, but that leaves everyone else — PC makers, consumer gadgets and anything else that needs to temporarily store a billion bits — scrambling to keep up with scarce supply and inflated prices.

How did the electronics industry get into this mess and, more importantly, how will it get out of it? IEEE Spectrum » asked economists and memory experts for explanations. They argue that the current situation is the result of a collision between the DRAM industry’s historic boom-and-bust cycle and an AI hardware infrastructure buildout unprecedented in its scale. And, barring a major collapse of the AI ​​sector, it will take years for new capabilities and new technologies to align supply with demand. Even then, prices could remain high.

To understand both sides of the story, you need to know the main culprit behind the supply and demand oscillation, High Bandwidth Memory or HBM.

What is HBM?

HBM is the DRAM industry’s attempt to short-circuit Moore’s Law slowdown by using 3D chip packaging technology. Each HBM chip is made up of 12 lightweight DRAM chips called dies. Each chip contains a number of vertical connections called silicon vias (TSV). The dies are stacked on top of each other and connected by arrays of microscopic solder balls aligned on the TSVs. This DRAM tower (well, at about 750 micrometers thick, it’s more of a brutalist office building than a tower) is then stacked on top of what’s called the base chip, which shuttles the memory chips and processor back and forth.

This complex piece of technology is then placed within a millimeter of a GPU or other AI accelerator, to which it is connected by up to 2,048 micrometer-scale connections. The HBMs are attached to both sides of the CPU, and the GPU and memory are grouped into a single unit.

The idea behind such tight, highly connected compression with the GPU is to break down the so-called memory wall. This is the energy and time barrier to importing the terabytes per second of data needed to run large language models in the GPU. Memory bandwidth is a key limiter on LLM execution speed.

As a technology, HBM has been around for over 10 years and DRAM manufacturers have worked to improve its capabilities.

As the size of AI models has increased, so has the importance of HBM for the GPU. But this comes at a cost. Semi-analysis estimates that HBM typically costs three times as much as other memory types and constitutes 50% or more of the cost of the bundled GPU.

Origins of the memory chip shortage

Memory and storage industry observers agree that DRAM is a highly cyclical industry, with huge booms and devastating busts. With new factories costing at least US$15 billion, companies are extremely reluctant to expand and may only have the cash to do so during boom times, says Thomas Coughlin, a storage and memory expert and president of Coughlin Associates. But building such a plant and making it operational can take 18 months or more, virtually guaranteeing that new capacity arrives well beyond the initial surge in demand, flooding the market and driving down prices.

According to Coughlin, the origins of the current cycle can be traced back to the chip supply panic caused by the COVID-19 pandemic. To avoid supply chain stumbles and support the rapid shift to remote work, hyperscalers — data center giants like Amazon, Google and Microsoft — have been buying up huge inventories of memory and storage, thereby raising prices, he notes.

But then supply became more steady and data center expansion slowed in 2022, causing memory and storage prices to fall. This recession continued into 2023 and even led large memory and storage companies such as Samsung to cut production by 50% to try to keep prices from falling below manufacturing costs, Coughlin says. This was a rare and quite desperate move, as companies typically have to run their factories at full capacity just to recoup their value.

After a recovery began in late 2023, “all the memory and storage companies were very cautious about increasing their production capacity again,” Coughlin says. “There has therefore been little or no investment in new production capacity in 2024 and most of 2025.”

The AI ​​data center boom

This lack of new investment is colliding with a surge in demand for new data centers. Worldwide, nearly 2,000 new data centers are currently planned or under construction, according to Data Center Map. If they were all built, it would represent a 20 percent increase in global supply, which currently stands at around 9,000 installations.

If current development continues at pace, McKinsey predicts that companies will spend $7 trillion by 2030, the bulk of which ($5.2 trillion) will be on AI-driven data centers. Of that, $3.3 billion will be spent on servers, data storage and networking equipment, the company predicts.

So far, the biggest beneficiary of the AI ​​data center boom is undoubtedly GPU maker Nvidia. Revenue from its data center business grew from just $1 billion in the final quarter of 2019 to $51 billion in the quarter ended October 2025. During that period, the GPUs in its servers demanded not only more and more gigabytes of DRAM, but also an increasing number of DRAM chips. The recently released B300 uses eight HBM chips, each of which is a stack of 12 DRAM dies. Competitors’ use of HBM largely mirrors that of Nvidia. AMD’s MI350 GPU, for example, also uses eight 12-chip chips.

With such demand, a growing share of DRAM makers’ revenue comes from HBM. Micron, the third-largest producer behind SK Hynix and Samsung, reported that HBM and other cloud-related memory grew from 17% of its DRAM revenue in 2023 to nearly 50% in 2025.

Micron expects the total HBM market to grow from $35 billion in 2025 to $100 billion by 2028, larger than the entire DRAM market in 2024, CEO Sanjay Mehrotra told analysts in December. It reached this figure two years earlier than Micron had anticipated. Across the sector, demand will outstrip supply “significantly… for the foreseeable future,” he said.

Future DRAM supply and technology

“There are two ways to solve DRAM supply problems: by innovating or by building more factories,” says Mina Kim, economist at Mkecon Insights. “As scaling DRAM has become more difficult, the industry has turned to advanced packaging…that simply uses more DRAM.”

Micron, Samsung and SK Hynix together represent the vast majority of the memory and storage markets, and all three have new factories and facilities in the works. However, these measures are unlikely to contribute significantly to lowering prices.

Micron is building an HBM factory in Singapore that is expected to be in production in 2027. And it is retooling a factory it purchased from PSMC in Taiwan that will begin production in the second half of 2027. Last month, Micron broke ground on what will be a DRAM factory complex in Onondaga County, New York. It will not be in full production until 2030.

Samsung plans to begin production at a new factory in Pyeongtaek, South Korea, in 2028.

SK Hynix is building an HBM and packaging facility in West Lafayette, Indiana, which is expected to begin production by the end of 2028, and an HBM manufacturing plant it is building in Cheongju is expected to be completed in 2027.

Speaking about his perception of the DRAM market, Intel CEO Lip-Bu Tan told attendees at the Cisco AI Summit last week, “There won’t be any relief until 2028.”

Since these expansions will not be able to contribute for several years, other factors will be necessary to increase supply. “Relief will come from a combination of incremental capacity expansions by current DRAM leaders, yield improvements in advanced packaging, and broader diversification of supply chains,” said Shawn DuBravac, chief economist at the Global Electronics Association (formerly IPC). “The new factories will only help marginally, but the quickest gains will come from learning the processes, better [DRAM] stacking efficiency and closer coordination between memory suppliers and AI chip designers.

So, will prices drop once some of these new plants come online? Don’t bet on it. “In general, economists find that prices fall much more slowly and reluctantly than they rise. DRAM today is unlikely to be an exception to this general observation, especially given the insatiable demand for computing,” says Kim.

Meanwhile, technologies are in the pipeline that could make HBM an even bigger consumer of silicon. The standard for HBM4 can accommodate 16 stacked DRAM dies, even though current chips only use 12 dies. Getting to 16 has a lot to do with chip stacking technology. Conducting heat through the HBM “layer cake” of silicon, solder and support material is a key limiter for going higher and repositioning the HBM inside the package to get even more bandwidth.

SK Hynix claims an advantage in thermal conduction thanks to a manufacturing process called advanced MR-MUF (mass reflow molded underfill). Further down the road, an alternative chip stacking technology called hybrid bonding could facilitate thermal conduction by essentially reducing the vertical distance between chips to zero. In 2024, researchers at Samsung proved they could produce a 16-high stack with hybrid bonding, and they suggested 20 chips weren’t out of reach.