Everyone knows who is winning the AI boom: the chipmakers, and the GPU above all. But honestly, a GPU is a strangely helpless thing on its own; it cannot store the data it produces, generate the electricity it burns, dissipate the heat it produces, or talk to the thousands of chips it works beside. A different industry fills each of those gaps, and each industry also indirectly benefits from this AI wave. In this article, we will discuss the different ideas or industries that also benefit from this wave of AI spending.
Memory and storage
Memory might be where the ripple is the sharpest. AI accelerators (the specialised chips that run AI, the most common being GPUs) depend on a faster, denser type of memory called high-bandwidth memory (HBM). The difficulty is that HBM is expensive to produce, as manufacturing it requires 3-4 times the production capacity of the standard memory used in laptops and phones. Since capacity is fixed, HBM has higher profitability, and every production line a manufacturer shifts towards HBM takes capacity away from conventional memory.
The consequences have been significant. Memory prices have risen sharply, and memory manufacturers like Micron and Samsung are extending their agreements from quarterly or annual terms to 3- to 5-year terms. Micron even announced it would exit the Crucial consumer business to prioritise data-centre customers. Most major vendors and analysts do not expect meaningful relief until around 2028. Storage also faces the same pressure, since AI training data has to be stored somewhere. The effect ultimately reaches consumers in the form of laptops, phones and storage. If you want to build a PC now, it might not be a good idea.
Now, there is a dedicated memory fund, the Roundhill Memory ETF (DRAM). However, it is new (launched in April) and highly concentrated, with roughly three companies accounting for three-quarters of its weight. For steadier exposure, a broad semiconductor ETF such as the iShares Semiconductor ETF (SOXX) or the VanEck Semiconductor ETF (SMH) holds the major memory makers alongside the wider industry.
Power and electrical infrastructure
If memory is the squeeze, then power is the possible constraint. Global data-centre electricity demand is projected to double, reaching about 945 TWh by 2030, roughly equal to Japan’s current annual consumption. In the U.S., data centres accounted for about half of electricity demand growth in 2025, and IT load capacity could roughly double from ~80 GW in 2025 to ~150 GW by 2028, according to Bloom Energy.
The demand spreads through several layers of the economy. Electricity must be generated, which is renewing interest in natural gas and nuclear capacity. Then it must also be delivered and managed, which includes switchgear, transformers, and backup power systems. There is already a shortage of transformers in the US, and lead times for switchgear have also ballooned. 30-50% of large data centres scheduled for 2026 are expected to be delayed, partly due to power constraints and equipment shortages.
For exposure, the First Trust Nasdaq Clean Edge Smart Grid Infrastructure ETF (GRID) concentrates on electrical equipment and grid companies. At the same time, the broader Global X U.S. Infrastructure Development ETF (PAVE) captures the construction and heavy-equipment side of the buildout. However, its exposure extends to broader infrastructure such as freight rail.
Cooling
The tremendous power required is converted into heat, which leads directly to cooling. According to S&P Global, traditional enterprise and cloud data centres are designed for 8-15 kW/rack, while modern AI data centres are engineered for 50-100 kW/rack. At that density, circulating cold air over the components is no longer sufficient. The industry is therefore shifting to liquid cooling, which channels fluid directly past the hottest parts, removing heat far more effectively. With liquid cooling, it enables higher power densities,150 kW/rack from Blackwell/Hopper superpods. According to Trendforce, the deployment of liquid cooling in AI data centres rose from 14% in 2024 to 33% in 2025, with 40% of AI data centres expected to use it in 2026.
This has created an entirely new category of equipment: coolant distribution units, cold plates and specialised plumbing. Cooling has gone from a routine consideration to a distinct investment theme. There is no dedicated cooling ETF; most cooling specialists sit inside larger industrial or electrical companies, with the major beneficiaries best accessed individually.
Optical communication
So AI training does not run on a single chip but across thousands working in concert, and that coordination depends on moving enormous volumes of data between them at high speed. The key point is that each additional chip can increase and scale interconnect complexity, which means a larger number of connections are needed, so the demand for optical components such as transceivers, switches and fibre grows faster than chip count itself. The traditional copper cabling runs into physical limits at high speeds and over longer distances, including signal loss, interference, and higher power requirements. Light carried through optical fibre largely avoids these constraints, which is why the industry is shifting toward it.
The technology is also advancing further. Optical components have historically been housed in pluggable modules at the board edge, but the industry is increasingly moving toward silicon photonics and co-packaged optics, placing optics much closer to the ASIC to reduce electrical loss and support higher data rates. As clusters scale toward tens of thousands of accelerators, the time chips spend waiting for data to move between them grows, and the network itself could become as much of a performance bottleneck as the chips. There is no clean ETF for optical communication specifically, hence exposure typically requires picking individual names.
The fifth perspective
Beyond these four beneficiaries mentioned above, others could also benefit. As AI circuit boards become denser, they will place greater demand on passive components (e.g., capacitors, inductors, resistors), while conventional CPUs still need to coordinate each server. These are genuine tailwinds and what’s happening in the market, but not a guarantee of returns. Memory, especially, has a long history of boom and bust; much of the optimism may already be priced in, and the whole chain rests on the thesis that AI capital spending will continue to grow. The point of this article is to map where the ripples flow rather than to recommend specific names. Each section can be dug deeper and expanded into an individual article or case study by itself; this is just scratching the surface. The story of AI actually extends well beyond just a single chip.
