AI’s next bottleneck is memory

The artificial intelligence boom is exposing a new constraint on computing power as demand for the memory used alongside advanced processors begins to outstrip supply.

Graphics processing units, or GPUs, such as those produced by Nvidia have become the most recognizable semiconductors behind AI. But their ability to perform enormous numbers of calculations depends on memory chips that can deliver data at comparable speeds. As AI models grow more computationally demanding, that relationship is driving an extraordinary shift within the semiconductor market.

Gartner, a technology research and advisory firm, forecasts that manufacturers will generate $837 billion from sales of memory chips in 2026, almost four times as much as in 2025, with revenue exceeding $1 trillion next year. Memory is expected to account for 54% of all semiconductor revenue this year, up from 27% in 2025, according to its August 2026 semiconductor forecast.

The pressure is beginning to spill into the rest of the technology industry. Memory manufacturers have limited production capacity, leaving AI systems competing with computers, smartphones and other electronics for it. As more resources flow toward data centers, AI investment is beginning to affect the cost and supply of devices far removed from them.

Why AI needs so much memory

The performance of an AI processor depends not only on how quickly it can calculate, but on how quickly it can access the data required for those calculations. A powerful GPU can process enormous quantities of information, but some of that computing capacity goes unused if the memory supplying the data cannot keep pace.

The need to move ever-larger quantities of data has driven demand for high-bandwidth memory (HBM), an advanced form of dynamic random-access memory (DRAM). HBM places stacks of memory chips close to the processor, allowing large quantities of data to move between the two simultaneously and reducing the time processors spend waiting for information.

The quantities required are substantial. A June 2026 study from the Semiconductor Industry Association and Deloitte found that a single advanced AI computing system can contain more than 4,500 semiconductor chips, including processors, memory and networking chips.

As companies build more AI systems, demand for memory is rising alongside demand for the processors themselves. Gartner expects revenue from DRAM sales to increase 247% in 2026, driven in part by the greater amount of memory being installed in AI servers.

Supply cannot catch up overnight

The most advanced memory cannot simply be produced in unlimited quantities. HBM is more difficult to manufacture than conventional DRAM, while building and equipping new semiconductor facilities requires enormous investment and can take years.

In its 2026 Key Questions on Energy and AI report, the International Energy Agency (IEA) found that a shortage of HBM had emerged and was expected to persist through at least the end of 2027. The agency placed memory alongside electricity, grid connections, and chip-manufacturing capacity among the physical constraints that could limit the expansion of AI infrastructure.

Relieving that pressure will require substantial new manufacturing capacity. SK Hynix, the South Korean company with an estimated 58% of the global HBM market, broke ground in August on a $4 billion memory facility in Indiana. Yet volume production of its next-generation HBM4E chips there is not expected until the third quarter of 2029. The company expects the global memory shortage to persist through 2030 and has approved roughly $38 billion in investment through 2031.

Even Nvidia is confronting the constraint. The company said in August that shortages of memory components were limiting supply as demand for its AI systems continued to rise, while higher memory and other component costs were putting pressure on profit margins.

When AI competes with consumer electronics

Finite manufacturing capacity means supplying more memory for AI can leave less available elsewhere. International Data Corporation (IDC), a technology market research firm, found in its June 2026 analysis that server demand was growing faster than supply could respond as manufacturers prioritized HBM, while PCs and smartphones faced higher component costs.

The memory industry is also highly concentrated, with advanced chips produced largely by South Korea's SK Hynix and Samsung Electronics and U.S.-based Micron. Their decisions about where to invest manufacturing capacity can therefore reverberate across the global electronics industry.

The pressure is already reaching consumers. Apple raised prices for MacBooks and iPads in June as memory and storage-chip costs surged, providing one example of how a shortage originating partly in data-center demand can travel through the technology supply chain.

The bottleneck moves

The memory shortage illustrates how solving one constraint on AI infrastructure can simply move the bottleneck elsewhere.

GPUs initially commanded attention because companies could not obtain enough of the processors required to build advanced AI systems. As investment has accelerated, pressure has spread into high-bandwidth memory, advanced semiconductor packaging, networking equipment and electricity infrastructure. Producing more processors is of limited value if companies cannot secure the memory required to use them effectively.

Gartner expects additional memory capacity to enter the market in 2027, but forecasts supply and demand will remain tight as AI infrastructure consumes more memory. The enormous increase in memory revenue suggests that the semiconductor industry's center of gravity has already begun to move.

The consequences extend well beyond companies building AI models. Every expansion of high-end memory production requires capital, factory space and manufacturing resources that could otherwise serve another part of the electronics industry. AI’s appetite for computing power is therefore becoming a force across the technology economy, influencing what gets manufactured, where investment flows and ultimately what consumers pay.