The hidden engine of AI: Why memory speed matters

Tom Bailey:

Hello everyone and welcome to today’s webinar. My name is Tom Bailey, Head of Research at HANetf.

Please note that this webinar is for professional investors only.

Today we will be looking at one of the most talked-about themes of the year: memory and memory chips.

Memory is crucial to the workings of AI. If you look at a computer, it has processors and memory, and these two components work together. Memory has emerged as one of the key bottlenecks in the AI buildout, leading to significant increases in memory prices and earnings growth for some memory-focused companies.

We recently launched an ETF tracking this theme in partnership with Defiance, a US-based ETF firm. This is our fourth ETF launched with Defiance.

The memory ETF has been listed on Borsa Italiana and Xetra in Germany and will soon be coming to the London Stock Exchange.

To discuss this theme, I am joined by Brendan Cavanaugh, Chief Strategy Officer at Defiance ETFs.

Brendan will provide an overview of the theme and the product before we move into a Q&A session.

Brendan Cavanaugh:

Thanks, Tom.

My name is Brendan Cavanaugh, Chief Strategy Officer at Defiance ETFs.

As Tom mentioned, this is the fourth fund we have launched in partnership with HANetf, and we are very excited about both the partnership and this specific ETF.

To give some background on Defiance, we were founded in 2018 with the goal of being first to market with high-conviction, next-generation investment themes through the ETF structure.

Fast forward to 2026, and we now manage more than $13 billion across approximately 80 ETFs.

Our product range includes leverage and inverse single-stock funds, option-based income ETFs, and our largest and most prominent area: thematic ETFs.

That is where DRAM fits in.

To understand the opportunity, it makes sense to look at the AI value chain and where different components fit.

We are all using AI more than ever. It is easy to open ChatGPT, Claude, Copilot or another AI platform, enter a prompt and receive information.

However, people sometimes overlook all of the infrastructure required to make these systems work.

The AI value chain begins with raw data collection and extends through to the interfaces users interact with. Along the way, there are multiple important stages, including data processing, computing, chips and memory.

Every part of this ecosystem needs to work together for AI systems to function effectively.

We are currently in a major expansion phase for AI infrastructure.

When looking at the AI value chain, memory is where we see a significant bottleneck emerging.

In recent years, compute performance has increased much faster than memory bandwidth.

Essentially, GPUs are able to process data significantly faster than memory can provide that data.

This creates a bottleneck because the processor is waiting for memory to deliver information.

The challenge becomes even greater as AI models continue to increase in size and complexity.

When looking at where costs are captured across AI infrastructure, memory represents a meaningful proportion of the overall system cost.

Brendan Cavanaugh:

One of the interesting things about memory is that it is a critical component across almost every area of technology.

When people think about semiconductors, they often focus on processors and the companies designing the chips that perform calculations.

However, memory plays a different but equally important role.

Every computing system requires both processing capability and the ability to store and access data efficiently.

As technology advances, the relationship between compute and memory becomes increasingly important.

The growth of AI has highlighted this because AI workloads require much greater volumes of data to be processed at much higher speeds.

Tom Bailey:

A lot of the discussion around AI has historically focused on training large models. How important is memory for inference as AI moves into wider use?

Brendan Cavanaugh:

Inference is becoming increasingly important.

Training is the process of building an AI model, while inference is when that model is used to generate outputs, answer questions or complete tasks.

As AI becomes integrated into more applications, the amount of inference activity is expected to increase.

This requires significant computing resources, including memory.

The shift from AI being mainly a research and development tool towards being integrated into everyday applications creates additional infrastructure demand.

Data centres are the physical foundation of the AI ecosystem.

The expansion of AI requires companies to build larger and more powerful data centres capable of supporting increasing workloads.

Within these facilities, memory is a crucial component alongside processors, networking equipment and power infrastructure.

The growth of hyperscale data centres has therefore become an important driver of demand for advanced memory technologies.

Increasing memory supply is a complex process.

Semiconductor manufacturing requires specialised facilities, highly skilled engineers and advanced equipment.

Even when companies decide to increase investment, new capacity can take years to become operational.

There are also challenges around maintaining manufacturing quality as technology becomes more advanced.

This means supply responses often lag behind changes in demand.

Brendan Cavanaugh:

One of the interesting things about memory is that it provides exposure to a fundamental part of the technology ecosystem.

Many technology themes focus on specific applications or consumer trends.

Memory is different because it is an underlying infrastructure component.

Whether AI develops through large language models, robotics, autonomous systems or other applications, these technologies require computing infrastructure.

Memory is a key part of that foundation.

Tom Bailey:

Could you explain the rationale behind creating a dedicated memory ETF?

Brendan Cavanaugh:

The goal was to provide targeted exposure to companies involved in the memory ecosystem.

Many investors are familiar with AI and semiconductor themes, but memory is a more specific area within that broader landscape.

The ETF is designed to capture companies involved across different parts of the memory value chain.

This includes companies focused on memory production, manufacturing equipment, technology development and related infrastructure.

The memory ecosystem extends beyond companies that manufacture memory chips.

It includes businesses involved in:

• Memory chip production
• Semiconductor manufacturing equipment
• Testing and packaging technologies
• Supporting technologies used in memory development

The broader approach reflects the fact that the growth of memory requires an entire ecosystem of companies.

An ETF provides investors with diversified exposure across a group of companies rather than requiring them to select individual businesses.

Within emerging technology themes, different companies can benefit at different stages of development.

A diversified approach can provide exposure across the broader ecosystem.

Brendan Cavanaugh:

The index focuses on companies that have meaningful exposure to the memory theme.

The methodology identifies businesses involved in areas such as memory production, memory-related technologies and the broader semiconductor ecosystem.

The aim is to provide exposure to the companies positioned within the memory value chain.

Memory requirements are likely to continue evolving alongside AI development.

As models become more complex and applications expand, the need for efficient data storage and processing is expected to remain important.

The relationship between computing power and memory performance will continue to be a key consideration for AI infrastructure.

Like all technology markets, the memory industry faces potential challenges.

These include:

• Changes in AI investment trends
• New technological developments
• Increased competition
• Changes in supply and demand conditions
• Broader semiconductor market cycles

The memory industry has historically experienced periods of expansion and contraction, and these dynamics remain relevant.

Tom Bailey:

Thank you, Brendan.

It has been very useful to understand the role memory plays within the broader AI infrastructure buildout.

The discussion highlights that while much attention is focused on AI applications and processors, the supporting infrastructure is equally important.

Memory is a key component that enables AI systems to process and manage the increasing volumes of data they require.

Thank you everyone for joining us today.