There is no end in sight for the GPU supply problem

The shortage may last because AI data centers are absorbing memory faster than manufacturers can produce it. Suppliers are also shifting production toward high-bandwidth memory, or HBM, instead of consumer-oriented memory. That leaves less capacity for GPUs, PCs, and other electronics.

Expansion cannot happen instantly. SK Hynix chairman Chey Tae-won said building adequate wafer supply would take at least four to five years. Manufacturers must also move from older HBM generations to HBM4 and HBM4E, which further disrupts production during the transition.

The pressure is already extending into future sales. Micron reportedly sold 75 percent of its 2027 HBM output, while Samsung sold almost 80 percent of its 2027 memory output. Leaders expect supply to be much tighter in 2027 and 2028, and production may remain 20 percent below demand until at least 2030. Some executives think prices could plateau sooner if cheaper Chinese alternatives arrive.

A graphics processing unit, or GPU, is a specialized processor that performs many calculations at once. It powers visual graphics, but it also handles the parallel workloads used to train and run AI models. GPUs therefore need substantial memory capacity and speed to keep their computing cores supplied with data.

The article focuses on memory because GPU performance is constrained when data cannot move quickly enough. High-bandwidth memory gives AI systems faster connections and more bandwidth. Without enough of it, powerful GPU hardware cannot operate at its full potential. Memory shortages can therefore limit both AI infrastructure and consumer graphics cards.

That shared demand is creating a squeeze. AI data centers need huge stores of HBM, while consumer GPUs and PCs need other DRAM products. Both compete for limited wafer capacity. As suppliers redirect that capacity toward higher-margin HBM, consumer GPU supplies decline and prices rise. The result is a shortage driven by memory, not just processor production.

AI data centers are projected to consume roughly 70 percent of global DRAM supplies this year. That enormous share leaves less memory for consumer GPUs, computers, consoles, and other devices. The shortage is especially severe because HBM, the memory AI systems prefer, uses wafer capacity that could otherwise produce traditional DRAM.

The price impact is already dramatic. A JP Morgan report projects that DRAM prices will rise 400 percent between 2024 and the end of 2026. HBM also sells for roughly 30 percent more than traditional DRAM, giving manufacturers a strong financial reason to prioritize it. As demand grows, suppliers are committing more capacity to AI memory.

The numbers point to continued pressure rather than a quick correction. HBM could represent nearly 30 percent of DRAM production in 2027, compared with 20 percent the year before. Some expect it to reach almost 40 percent by 2029. Meanwhile, major suppliers still describe future memory supplies as tight despite expanding production.

When memory suppliers favor AI data centers, consumer electronics receive less production capacity. GPUs, laptops, consoles, and accessories all depend on memory, so reduced supply raises component costs and limits the number of devices available. Consumers then face higher prices or delayed purchases.

The article gives striking examples. Nvidia's RTX 5090 was selling for $4,920 above its $1,999 launch price, according to the Tom's Hardware tracker. GameStop was marketing Sony's PlayStation 5 Pro for about $500 above MSRP. The Consumer Price Index for software and accessories also rose nearly a quarter over two years.

The damage is spreading beyond graphics cards. The Economist reported similar increases for video game consoles and laptops, alongside an unusual decline in PC shipments. Resellers can exploit scarce inventory, while manufacturers prioritize more profitable AI-related products. Unless memory production catches up, consumer prices and supply problems may persist through the late 2020s.

High-bandwidth memory, or HBM, is a specialized form of dynamic random access memory. It provides faster connections and more bandwidth than traditional DRAM. That makes it suited to workloads that must move very large amounts of data quickly between memory and processors.

AI developers use HBM to train and run their models. Training processes large datasets, while running models also requires rapid access to their stored information. The article says demand for HBM already far outpaces current market supplies. AI data centers therefore need vast stores of this specialized memory as they expand.

HBM's importance creates a wider supply problem because it competes for the same wafer capacity as ordinary DRAM and other memory products. Suppliers are transitioning to newer HBM generations, including HBM4 and HBM4E, which further limits output during the changeover. HBM could account for nearly 30 percent of DRAM production in 2027, up from 20 percent the previous year.

Memory manufacturers prioritize HBM because it combines exceptional demand with better margins. AI developers need large quantities of HBM for training and running models, while data centers are expanding rapidly. Suppliers follow that demand because AI customers are willing to secure substantial future capacity.

The financial incentive is clear. The article says HBM sells for roughly 30 percent more than traditional DRAM. Producing it uses wafer capacity that could support consumer memory, but manufacturers can earn more by serving AI infrastructure. The market is also concentrated: Micron, Samsung Electronics, and SK Hynix control roughly 90 percent of global DRAM production.

This strategy hurts consumer supply. Micron stopped direct-to-consumer sales last year to dedicate resources to AI production. HBM could make up nearly 30 percent of DRAM production in 2027, compared with 20 percent the year before. The three major suppliers still expect long-term shortages even while expanding capacity.

Semiconductor capacity cannot expand like ordinary factory output. Manufacturers need new wafer capacity, specialized equipment, and production lines before they can make more memory. Moving from one memory generation to another also interrupts supplies while factories are reconfigured. These steps make a shortage difficult to fix quickly.

The article highlights the scale of the delay. SK Hynix chairman Chey Tae-won said it would take at least four to five years to build adequate wafer supply. Suppliers are also transitioning toward HBM4 and HBM4E, creating another production challenge as they move to newer technology. Expansion plans therefore add capacity gradually rather than immediately.

Demand continues rising during that wait. AI data centers are projected to consume roughly 70 percent of global DRAM supplies this year. Micron and Samsung have already sold much of their 2027 output, while SK Hynix expects 2027 to be the industry's worst supply year. Production may remain 20 percent below demand until at least 2030.