AI infrastructure

How does AI infrastructure work?

AI infrastructure is the physical and software backbone that trains and serves large AI models. Think of it as layers stacked on top of each other: accelerator chips do the math, high-bandwidth memory feeds them data, networking ties thousands of chips into one cluster, foundries fabricate the silicon, and hyperscale clouds rent the finished compute. The AI infrastructure theme groups the companies that own each of those layers rather than the consumer apps sitting on top.

The layer that gets the most attention is compute. NVIDIA (NVDA) supplies the dominant training accelerator, and its CUDA software stack is what makes those chips sticky. AMD (AMD) is the most credible alternate accelerator path, while Broadcom (AVGO) builds the custom AI silicon that hyperscalers design in-house. Micron (MU) supplies the HBM memory stacked next to those accelerators. When AI training budgets grow, demand flows straight through these AI infrastructure names because they sit at the center of every cluster.

What is the AI infrastructure supply chain, from foundry to networking?

No accelerator exists without a supply chain behind it, and the AI infrastructure theme deliberately spans that whole chain. TSMC (TSM) fabricates virtually every leading-edge AI chip, patterning transistors on EUV lithography systems that only ASML (ASML) makes. Applied Materials (AMAT) supplies the deposition and etch tooling each new fab needs, and specialty suppliers like Entegris (ENTG), MKS Instruments (MKSI), and Element Solutions (ESI) provide the chemistries and sub-systems those fabs consume. Amkor (AMKR) handles the advanced packaging (CoWoS, chiplets) that turns bare die into finished accelerators.

Once chips are made, they have to talk to each other at enormous speed, which is the networking layer of AI infrastructure. Arista (ANET) builds the back-end switches that wire hyperscaler AI clusters together, Marvell (MRVL) supplies the optical DSPs and custom interconnect silicon, and Corning (GLW) manufactures the optical fiber that AI data centers consume in far greater volumes than traditional cloud. Each link compounds: more accelerators means more memory, more packaging, more switches, and more fiber, which is why the AI infrastructure theme treats the supply chain as one connected thesis.

How do AI infrastructure companies make money?

AI infrastructure companies monetize the AI buildout in two broad ways: selling the hardware that goes into clusters, and renting the finished compute. On the hardware side, chip designers (NVDA, AMD, AVGO) and the foundry, equipment, memory, and networking suppliers behind them (TSM, ASML, AMAT, MU, MRVL, ANET) book revenue as hyperscalers and AI labs place orders. Their economics track AI capex directly, so the AI infrastructure theme rises and falls with how much the largest buyers are willing to spend on training and inference.

The second engine is the cloud platform layer. Microsoft (MSFT) through Azure, Alphabet (GOOGL) through Google Cloud and its TPU silicon, Amazon (AMZN) through AWS and Trainium, and Oracle (ORCL) through OCI rent AI compute by the hour to companies that do not want to build their own data centers. These hyperscalers are simultaneously the biggest customers of the chip makers and the biggest sellers of finished AI capacity, which is why they anchor the AI infrastructure theme alongside the silicon names. As of 2026 the two engines feed each other: hyperscaler spending funds the chip vendors, and the chip vendors enable the cloud capacity hyperscalers then resell.

What gets a stock into the AI infrastructure theme?

Revenue meaningfully driven by AI training or inference capex: GPUs and accelerators, optical and high-speed interconnect, semiconductor foundries and equipment, specialty materials, hyperscaler cloud spend, or the platforms hyperscalers buy from.

AI infrastructure stocks

Every public name that fits the AI infrastructure thesis, with the rationale for inclusion. Click any ticker for the full stock guide. The basket above starts equal-weighted; you set your own target weights inside Walnut.

How to invest in AI infrastructure

There are a few ways to get exposure to the AI infrastructure theme, and Walnut is not an investment adviser, so what follows is descriptive rather than a recommendation. The most concentrated approach is buying individual stocks that fit the thesis directly, for example a pair like NVDA and TSM that pairs the dominant accelerator with the foundry that makes it, or a slightly wider set adding AVGO, ANET, and MRVL across the networking layer. This is the tightest expression of the AI infrastructure thesis, but it concentrates risk in a handful of correlated names. A second route is ETF proxies: SMH and SOXX are the most semiconductor-heavy (dominated by NVDA, TSM, AVGO, AMD), QQQ blends the hyperscalers with NVDA, and VGT and XLK give broader tech-sector exposure. The limitation is that none of these is a pure-play AI infrastructure vehicle. They dilute the theme with chips and software that have little to do with AI capex, so your AI infrastructure exposure is real but blended down.

The third route, and the one Walnut is built for, is constructing a dedicated AI infrastructure basket. You describe the thesis to Walnut's AI assistant (for instance, AI infrastructure spanning compute, memory, networking, foundry, and packaging), and the assistant proposes a set of constituents drawn from names like NVDA, AVGO, TSM, ANET, MRVL, and MU, along with suggested target weights and the rationale for each. You review every constituent and weight, adjust anything you want, and then fund the basket through your own existing broker. You approve every order before it is placed; Walnut never trades on your behalf. The result is a single AI infrastructure basket that tracks as one performance line you can compare against SMH or QQQ, giving you a more targeted expression of the AI infrastructure theme than any off-the-shelf ETF can offer.

ETFs used as passive proxies for AI infrastructure

If you want the theme as a single ticker rather than as a basket, these are the ETFs people most commonly use. Each has trade-offs (concentration, expense ratio, sector overlap) covered in the individual ETF guides.

FAQ