AI Stocks: What Is Inside the AI Infrastructure Theme

The theme applies one test: does revenue depend meaningfully on AI training or inference capital spending? In practice that means accelerators and high-bandwidth memory, semiconductor foundries and the lithography and tooling behind them, advanced packaging and the specialty materials it consumes, high-speed interconnect and optics, server integration, and the cloud platforms that rent the finished compute.

The word doing the work is capital spending. Enormous numbers of companies now use AI, and almost all of them say so. Using AI makes a company a customer of this theme, not a member of it. Drop that distinction and the roster expands until it is simply a list of large technology companies, which is the failure mode of most funds and lists carrying the AI label.

The second structural choice is that the theme follows the chain down rather than picking the most visible layer. Owning only accelerator designers is a bet on one architecture in one part of one cycle. Owning only the cloud platforms is barely an AI position, because AI is a minority of what moves them. Holding the chain means the theme captures spending at the point it is committed, the point it is fabricated, the point it is wired together, and the point it is resold. For the general idea, see thematic investing.

This is where an AI budget lands first. Accelerators do the arithmetic that training and inference are made of, and high-bandwidth memory sits next to them because a processor starved of data is an expensive idle object. The layer is the most direct expression of the theme, because its revenue is recognised while clusters are being built rather than after anything built on them earns money. That timing is the single most important thing to understand about the whole roster, and the page returns to it below.

One more constituent sits in this layer without carrying a full entry, because its contribution is narrow but genuinely load-bearing.

• Rambus (RMBS). Supplies memory interface chips and related silicon IP for high-density server memory modules. It is a small, specific link in the same chain: denser AI servers need more memory modules, and those modules need interface silicon. The exposure is real but indirect, and licensing-driven businesses recognise revenue on their own rhythm rather than in step with shipments.

How this layer relates to the rest. Compute sets the pace for everything else. Every accelerator ordered pulls through foundry capacity, packaging slots, memory, switches, optics, and eventually rented cloud capacity, which is why an order slowdown here is felt in all four layers rather than one.

The companies in the compute layer design chips. Almost none of them make them. Leading-edge fabrication, the lithography that patterns the transistors, the deposition and etch tooling around it, the advanced packaging that bonds logic to memory, and the ultra-pure materials the whole process consumes are separate businesses with separate economics. This layer has the longest lead times and the narrowest chokepoints in the theme, which is exactly why it is included rather than assumed away.

Five more constituents fill out the manufacturing chain. None is a household name and none is optional, which is the point of including them: they are the tooling, packaging, and chemistry that make the layer work, and each has a narrower business and thinner diversification than the two names above.

• Applied Materials (AMAT). The largest semiconductor equipment company, supplying the deposition, etch, and inspection tooling that every new or upgraded fab has to install. It is the broadest read on fab construction in the theme, and like all equipment names its revenue follows capacity decisions rather than chip demand directly.
• Amkor Technology (AMKR). The largest United States headquartered outsourced assembly and test provider, doing the advanced packaging that bonds logic and memory into a finished accelerator. Packaging capacity has been one of the visible constraints on AI accelerator supply, which is why an otherwise low-profile business belongs in the roster.
• Entegris (ENTG). Supplies the high-purity materials, filtration, and handling products that fabs consume continuously. Leading-edge nodes use more of these per wafer than older ones, so the exposure grows with the mix shift toward advanced processes rather than only with fab count.
• MKS Inc. (MKSI). Builds the vacuum systems, lasers, and process-control subsystems that sit inside the equipment other companies sell to fabs. It is a supplier to the suppliers, which makes the AI link real but two steps removed from any accelerator order.
• Element Solutions (ESI). Provides specialty chemistries used in semiconductor assembly and advanced packaging through its MacDermid Alpha business. Advanced packaging is chemistry-intensive, so more complex accelerator assembly consumes more of this, though the wider company also serves markets with nothing to do with AI.

How this layer relates to the rest. Manufacturing sets the ceiling. Demand for accelerators only becomes revenue at the pace this layer can fabricate, package, and test them, so a packaging or tooling constraint caps the compute layer regardless of how strong order books look, and the cloud layer cannot rent capacity that was never built.

A pallet of accelerators is not a training system. Thousands of chips only behave like one machine if they can exchange data fast enough, which turns networking from a supporting cost into a design constraint. This layer covers the back-end switching that wires clusters together, the optical components and fiber that carry the traffic, and the server integration that ships the whole assembly to a data-center floor. It is the least glamorous part of the theme and the part that scales most directly with cluster size.

Two more constituents complete the layer, both of them ways to own the buildout without owning a chip designer.

• Corning (GLW). The largest manufacturer of optical fiber and cable, the physical medium AI data centers consume in far greater quantity than conventional cloud facilities because of how densely clusters are interconnected. Optical communications is one segment among several, so display and specialty materials businesses dilute the AI read considerably.
• Dell Technologies (DELL). One of the largest integrators of accelerator-based AI servers, taking chips from the compute layer and shipping configured, supported systems to cloud and enterprise buyers. It is the most demand-visible name in the layer and the thinnest-margin one, because integration captures volume rather than pricing power, and a large legacy PC and infrastructure business sits alongside it.

How this layer relates to the rest. This layer is what converts silicon into usable capacity. It sells to the same small group of buyers as the compute layer and on the same schedule, so it offers very little diversification against a capital-spending pause, but it does capture spending that goes up faster than chip count as clusters get larger.

The three layers above sell equipment. This one buys it, assembles it into a service, and rents it by the hour. That makes the cloud platforms simultaneously the largest customers in the theme and the companies whose spending decisions everything else keys off, which is a genuinely unusual structure and the reason the theme holds both sides. They are also the only constituents whose AI exposure sits inside businesses large and profitable enough to absorb a bad year in the buildout.

Two more cloud platforms round out the layer, included for how differently they express the same exposure.

• Amazon (AMZN). Operates AWS, the largest cloud platform, designs its own Trainium and Inferentia accelerators, and has invested in a leading model developer. It is the broadest read on renting AI capacity at scale, wrapped inside a retail business that dominates the reported numbers and dilutes the AI signal more than any other cloud constituent.
• Oracle (ORCL). A smaller cloud platform by overall share that has taken on outsized AI training workloads through Oracle Cloud Infrastructure, which makes it the most AI-levered name in the layer and the least diluted. That is the whole case for including it and the whole risk: it is building capacity against contracted future demand, so it carries the buildout's financing risk more visibly than the larger platforms do.

How this layer relates to the rest. This layer is both the demand signal and the counterweight. Capital-spending guidance from these companies is the number the rest of the roster trades on, and because their revenue comes overwhelmingly from software, advertising, retail, and conventional cloud, they do not move with AI capex the way the equipment layers do.

There is a fifth layer, and its absence from the roster is deliberate rather than an oversight. Above the cloud platforms sit the model developers and the applications built on them, and that is the layer that has to eventually produce the revenue that justifies everything below it. A data center is a cost until something running inside it is worth paying for.

Public markets can barely reach that layer. The largest independent model developers are privately held, so there is no share to buy. The listed exposure that does exist runs through the cloud platforms that host them and have invested in them, which is one of the reasons Microsoft, Alphabet, Amazon, and Oracle sit in the theme rather than being treated as ordinary software companies. Applied-AI software firms are a different exposure again, tracking enterprise software budgets rather than capital spending, which is why they are excluded here and covered by the AI agents theme.

Naming the gap is more useful than papering over it. This theme is a position on the construction of AI capacity. It is not a position on AI applications succeeding, and the two are related but not the same bet.

Read top to bottom, the theme is a dependency chain with a timing problem inside it. Cloud platforms commit capital. That commitment becomes orders for accelerators and memory, which becomes foundry wafer starts, packaging slots, lithography systems, tooling, and materials, which becomes switches, optics, fiber, and integrated servers, which becomes rentable capacity. Every link is paid before the link above it earns anything.

That sequence is the most important analytical fact about the roster, and it cuts both ways. It is why the picks-and-shovels layers have shown the clearest revenue: they are paid out of capital budgets, in advance of any return on those budgets. It is also why their results are not independent confirmation that the spending will continue. An accelerator sold is evidence that someone chose to build, not evidence that building was profitable.

The honest name for the risk is circularity. The buildout is funded by a small group of companies that are simultaneously each other's customers, suppliers, and investors. Suppliers sell into cloud platforms, cloud platforms back model developers, and model developers commit to buying capacity from those same platforms. None of that is improper, and demand paid for in cash is still demand. But it does mean the apparent breadth of the theme is narrower than a twenty-name list looks, because the same underlying decision by a handful of buyers is being counted at four different layers.

The practical consequence is that these constituents do not diversify each other as much as their variety suggests. A capital-spending pause would be felt by NVDA, MU, AMKR, ANET, and DELL at roughly the same time and for the same reason. What genuinely behaves differently is the cloud layer, where advertising, retail, and conventional software revenue dominates the results, and the far end of the manufacturing layer, where ASML sells against multi-year fab plans rather than this quarter's orders. Understanding that is more useful than any ranking of the 20.

A membership test is only credible if it excludes things. These are the names and categories people most often expect to find here, and the specific reason each one does not qualify.

• OpenAI and Anthropic. Privately held, so there is no listed security to include. This is the theme's largest structural gap rather than a technicality: the companies whose model revenue has to eventually justify the buildout are not directly ownable, and public exposure to them runs indirectly through the cloud platforms that host and back them.
• Data-center power, cooling, and electrical equipment. Electricity and thermal capacity are real constraints on the buildout, but the businesses that solve them are driven by grid, utility, and construction cycles that long predate AI and will outlast this phase of it. They belong to the data center power theme, where power supply is the thesis rather than a dependency.
• Applied-AI software and agent platforms. Companies selling AI-powered software to enterprises consume compute rather than supply it, and their revenue tracks software budgets, seat counts, and sales cycles instead of capital spending. That is a different economic exposure with a different timetable, and it sits in the AI agents theme.
• Broad cloud and enterprise SaaS. Most software delivered from a data center has nothing to do with AI training or inference capex, and adding AI features to an existing product does not change what drives the business. Those names sit in the cloud computing theme, which is defined by subscription software economics rather than by hardware spending.
• Analog, embedded, and IP-licensing chipmakers. They are semiconductor companies, which is not the same qualification. A microcontroller supplier to industrial customers is exposed to factory automation and autos, not to accelerator demand. Those names sit in the semiconductors theme, which is deliberately broader and less AI-specific than this one.
• Meta, Apple, and Tesla. All three spend heavily on AI, but they spend it on themselves. They are buyers of the infrastructure this theme sells, and owning them is a position on their own products, advertising, devices, or vehicles rather than on the buildout. Being a large AI consumer is not the inclusion test.

Power is the exclusion worth dwelling on, because it is the one people argue with. Electricity and cooling are genuine constraints on the buildout, arguably tighter ones than silicon. They are excluded here because the companies that supply them answer to utility capital cycles, permitting, and grid economics, so they would dilute a theme defined by AI capital spending with a very different driver. They have their own roster in the data center power theme. The same logic sends broad chipmakers to the semiconductors theme and subscription software to the cloud computing theme. A company can be an excellent business, and genuinely connected to AI, and still be the wrong expression of this particular theme.

All 20 names, grouped by the layer they occupy rather than ranked, so the shape of the theme is visible in one view.

Full entries above cover 10 of the 20, the names that define what their layer is. The rest are narrower businesses the layer does not work without, which is a different kind of importance and worth seeing rather than hiding.

The passive route answers a different question. An index defines what counts as AI, then assigns weights you do not control. The funds this theme names as proxies each capture part of the chain and miss the rest: SMH and SOXX are semiconductor funds, so they express the compute and manufacturing layers well, hold some of the cluster layer, and hold essentially none of the cloud layer. QQQ is the reverse, dominated by the cloud platforms with a large accelerator position attached. VGT and XLK are broad technology funds, so a meaningful share of what you own has nothing to do with the buildout at all.

A theme inverts the trade. You know exactly which names you own, which layer each represents, and what weight each carries, and you accept a narrower roster and the work of maintaining it. Neither is automatically better. The fund is the simpler instrument and the theme is the more deliberate one, and plenty of people hold a broad fund as a core with a small thematic tilt beside it. The one thing worth checking either way is overlap, because the largest names in this theme are already substantial positions inside most broad index funds.

A list of 20 names is an input, not a portfolio. What turns one into the other is structure: which layers you want exposure to, what weight each name carries, and whether the concentration you end up with was chosen or inherited.

• Decide the layer mix first, then the names. Moving weight from compute to manufacturing or cloud changes the character of the position far more than swapping one accelerator designer for another.
• Count the correlation, not the count. Twelve equipment names sharing four end customers is closer to one position than to twelve. Breadth inside a layer buys less than breadth across layers.
• Set target weights that sum to 100. Equal weighting is a choice, and so is tilting toward the cloud platforms for their dilution. Both are defensible. Not deciding is what leaves you concentrated by accident after one name runs.
• Check what you already own. The largest constituents here are usually already sizeable positions inside broad index funds, so adding them can concentrate a portfolio while feeling like diversification.
• Frame it against the S&P 500. A narrow thematic position should be judged against a broad benchmark, particularly this one, where several constituents are among the benchmark's largest members already.

This is what Walnut is built for. You describe the thesis, the AI assistant proposes constituents and weights you can edit, the portfolio tracks as one performance line against the S&P 500, and you place trades you approve yourself at your own broker. Walnut is informational and does not tell you which stocks to buy.

For the companion view of which AI names are most widely held and discussed, see best AI stocks. For the passive route in detail, see best AI ETFs.

The AI infrastructure theme is 20 companies across four layers, and the layering is the whole idea. The compute layer sells the accelerators and memory that AI budgets are spent on first. The manufacturing layer fabricates, packages, and supplies the materials for silicon that nobody in the compute layer makes themselves. The cluster layer wires chips into machines large enough to train on. The cloud layer buys all of it, rents it out, and is the only part of the roster whose results do not live or die on capital spending.

Understood as a flat list of AI stocks, the theme looks like twenty ways to own the same trend. Understood as a chain that is paid from the bottom up, it is a structure with one open question sitting on top of it: whether the model and application layer, the part public investors can barely own, eventually earns enough to keep the spending going. That question is the theme, and it is what you are deciding whether to own. Nothing here is a recommendation, and Walnut is not an investment adviser.