How memory is the AI bottleneck

Last updated: 5th October 2026 | Author: Cameron MacDonald

Summary

AI’s memory demands are fundamentally different to the memory requirements of everything else.
The AI memory crunch is likely to squeeze a multitude of other industries, including vehicles and telecoms.
For investors, the constraint could also be an opportunity.

What is AI memory and why is it different to traditional memory?

More traditional computer memory chips can be thought of as buildings connected with highways. However, the roads can only be widened – or a computer made faster – to a certain point, limiting the speed of the overall computer.

AI workloads are incredibly dense; when answering a question or performing a task, AI must execute billions of mathematical calculations simultaneously, bringing in massive amounts of data to produce an answer.

« If the memory cannot feed the chip with sufficient speed, a bottleneck is created, ensuring an unacceptable response time. »

High bandwidth memory, or HBM, fixes this issue by stacking the ‘buildings’ with these layers connected via silicon, essentially adding more lanes on the highways.[1]

However, while HBM fixes the bandwidth issue, the increasing complexity and size of AI models, along with the massive amount of data needed to train new models are making the physical memory chips an issue for AI buildout.

Source: McKinsey & Company. Data as of 26.02.2026. Chart shows expected data. For illustrative purposes only.

In 2024, the scale of the challenge was already clear to those in positions of authority – U.S. Commerce Secretary Gina Raimondo said that “the volume of chips [AI companies] project they need is mind-boggling”.[2]

Since then, demand for memory has increased dramatically as AI deployment becomes ever more present. SK Hynix (a South Korean company manufacturing memory chips essential to GPUs) reported a revenue increase of 198% year-on-year in April.[3] All three of Taiwan Semiconductor Manufacturing Company’s (TSMC) fabricators are fully booked with 52 to 78-week lead times.[4]

What is the manufacturing issue affecting AI build out?

To meet the requirements of AI companies, memory is expected to reach 30% of these companies’ spending in 2026.[5] AI companies are being forced to fundamentally alter their expenditure to meet demand.

AI hyperscalers are willing to sign three- to five-year non-refundable contracts just to secure supply of chips they need, and according to UBS Group, between 20% and 30% of dynamic random-access memory (DRAM) shipments may now be covered under these agreements.[6]

While this purchase behaviour indicates strong desire to purchase greater memory and chip capacity, there is a broader issue. For example, if Nvidia wants to sell more chips, they can only do so if TSMC produces more of them. The chips and computing power are not necessarily the issue; it is the assembly of the entire package that is producing the bottleneck and forcing prices for memory to rise.

What are the likely consequences of higher memory prices?

AI-driven memory demand is structural. Samsung, SK Hynix, and Micron – the three companies that together accounted for around 88% of global DRAM revenue in the second quarter of 2026 – have gradually reallocated resources to try and offset the demand, moving 93% of combined production to memory for AI data centres. [7] [8]

Consequently, consumer DRAM and NAND memory are in critically short supply, and DRAM prices have almost doubled since early 2025. As a result of this focus on facilitating AI workloads, other industries could see increased price pressure, and those companies at the lower end of each market could face a significant squeeze as costs continue to remain high.

Medical prices, car prices, telecoms, and the gaming industry are all facing price increases. For instance, a typical smartphone could see production costs increase by $30,[9] something that is likely to be passed onto the consumer.

What could this mean for investors?

The scale of the memory supply issue is significant. The lead times for production of current orders is long and added manufacturing capacity will take months if not years to match current demand.

At the same time, the demand for AI as a service is only expanding, Meta is investing a projected $135 billion to $145 billion in 2026, having raised the lower end of its guidance in July – nearly double that of 2025.^[10] Alongside this massive expenditure are the suppliers of the physical memory powering these AI models; Samsung, SK Hynix, and Micron have pricing power currently. Micron, for example, reported fiscal fourth-quarter revenue of $54.2 billion on 30 September 2026, up around 379% year-on-year.^[11] The fact is that these companies are producing a necessary part of the AI build out, and higher prices may not deter customers.

The manufacturers of AI memory could stand to benefit as AI continues to evolve and become more embedded within societies globally.

This material does not constitute a marketing document. It is not an invitation to invest but to be read for educational purposes only. Past performance and forecasts are not reliable indicators of future results.

[1] https://www.scientificamerican.com/article/high-bandwidth-memory-is-a-bottleneck-for-ai-chips/

[2] https://www.nbcnews.com/business/business-news/what-is-nvidia-what-do-they-make-ai-artificial-intelligence-rcna140171

[3] https://stockanalysis.com/quote/krx/000660/transcripts/547968-q1-2026/

[4] https://siliconanalysts.com/analysis/foundry-allocation-status-q1-2026

[5] https://x.com/SemiAnalysis_/status/2039870546582630470

[6] https://www.bloomberg.com/opinion/articles/2026-06-01/ai-rally-a-trillion-dollar-question-for-memory-chipmakers?srnd=phx-opinion-markets&embedded-checkout=true

[7] https://tech-insider.org/memory-chip-shortage-2026-ai-consumer-electronics/

[8] https://evertiq.com/news/2026-09-07-dram-industry-revenue-surges-595-in-q2-as-supply-lags-demand

[9] https://www.bbc.co.uk/news/articles/c1dzdndzlxqo

[10] https://www.kucoin.com/news/flash/meta-q2-2026-earnings-miss-estimates-capex-guidance-rises-to-145-billion

[11] https://www.sec.gov/Archives/edgar/data/0000723125/000072312526000018/a2026q4ex991-pressrelease.htm

Frequently Asked Questions

What is AI memory?

AI memory refers to the memory technology used to help AI systems access and move large amounts of data quickly. AI models rely on huge volumes of information, so memory plays a key role in helping chips process tasks at speed

Why does AI need different memory to traditional computing?

AI workloads are much heavier than many traditional computing tasks. When an AI model answers a question or completes a task, it may need to carry out billions of calculations while pulling in large amounts of data. If the memory cannot keep pace, the whole system slows down.

What is high bandwidth memory?

High bandwidth memory, or HBM, is a type of memory designed to move large amounts of data to the processor much faster. AI workloads need this because they rely on huge volumes of data being accessed at the same time. The article compares this to adding more lanes to a highway: HBM stacks memory layers and connects them through silicon, allowing more data to move at once and helping reduce bottlenecks.

Why are memory chips becoming a bottleneck for AI?

AI workloads require enormous amounts of data to be processed simultaneously, creating far greater memory demands than traditional computing. While AI chips have become increasingly powerful, they can only perform efficiently if memory can supply data quickly enough. When memory cannot keep pace with the processing power of AI chips, it creates a bottleneck that limits performance.

High-bandwidth memory (HBM) helps address this challenge by allowing data to move faster between memory and AI processors. However, the rapid growth of AI infrastructure has driven demand for advanced memory technologies, putting pressure on supply and making memory a critical constraint for further AI scaling.

Adapting precision warfare: How US defence spending priorities are changing

October 2026

How memory is the AI bottleneck

Last updated: 5th October 2026 | Author: Cameron MacDonald

Summary

﻿ « If the memory cannot feed the chip with sufficient speed, a bottleneck is created, ensuring an unacceptable response time. »

What is the manufacturing issue affecting AI build out?

What are the likely consequences of higher memory prices?

Frequently Asked Questions

What is AI memory?

Why does AI need different memory to traditional computing?

What is high bandwidth memory?

Why are memory chips becoming a bottleneck for AI?

Related Funds

More Articles

Investment Strategies

About Us

Resources

Leaving HANetf's Website

« If the memory cannot feed the chip with sufficient speed, a bottleneck is created, ensuring an unacceptable response time. »