What are AI infrastructure stocks? The 12 layers, explained

AI infrastructure stocks are the companies that get paid to build and run the physical side of AI: the chips, memory and networking inside a data center, the buildings, cooling and power around it, and the clouds that rent the finished capacity out. They sit underneath the AI products people use, and they collect the capital spending before those products earn anything.

The catch is how wide the label has become. Supercycle tracks 67 semiconductor stocks alone, and about 320 companies across 12 infrastructure layers, with AI exposure scores from single digits to 100. "AI infrastructure" covers NVIDIA and it covers a copper miner. This post is about telling them apart: what each layer sells, who pays it, and why two stocks in the same layer can carry completely different exposure to the buildout.

Where does the money come from?

From capital expenditure. Seven companies at the center of the buildout, Amazon, Microsoft, Alphabet, Meta, Oracle, CoreWeave and Nebius, spent $657 billion on capex over their last four reported quarters, on the figures in the September capex roundup, and the five that guide on a calendar year add up to between $600 billion and $634 billion for 2026. That spending is the revenue of every layer below it. The glossary entry on hyperscaler capex covers why it is a harder signal than any amount of AI commentary: guidance is public, expensive to walk back, and turns into orders down the whole chain.

The energy side is measurable too. The IEA estimates data centers used about 415 terawatt hours of electricity in 2024, around 1.5 percent of global consumption, and projects that to roughly double to 945 terawatt hours by 2030, with AI-accelerated servers accounting for almost half of the increase. That is the demand the power, grid and cooling layers are being paid to meet.

What are the 12 layers of AI infrastructure?

Supercycle groups the universe into 22 buildout layers. Twelve are infrastructure. The table runs from the ground up. Counts and scores are rendered from the live data at each nightly build; the examples are the highest-scoring names in each layer at the time of writing.

Layer	What it sells	Who pays it	Tracked stocks	Examples (score)
Raw materials	Copper, rare earths, uranium, lithium, substrates	Everyone above, indirectly	37	AXT (68), Ibiden (62)
Power and energy	Generation: gas turbines, nuclear, fuel cells, on-site plants	Data center operators, via power purchase agreements	57	GE Vernova (55), Oklo (85)
Utilities and grid	Transmission, substations, switchgear, electrical contracting	Data center operators and ratepayers	42	Eaton (50), IES Holdings (62)
Chip equipment and foundries	Fabrication tools and the fabs that make the chips	Chip designers	37	TSMC (78), Applied Materials (78)
Semiconductors	GPUs, accelerators, custom AI chips	Hyperscalers, neoclouds, AI labs	67	NVIDIA (97), Broadcom (78)
Memory and storage	HBM, DRAM, NAND, enterprise storage	Chip makers and server builders	9	SK hynix (80), Micron (78)
Networking and optics	Switches, interconnects, optical modules	Cluster builders	24	Astera Labs (92), Arista (80)
EMS and manufacturing	Assembling the servers, racks and boards	Chip and server brands	9	Celestica (85), Jabil (72)
Data centers and hardware	Server systems, racks, the buildings and the REITs that own them	Cloud providers and enterprises	64	Super Micro (90), Dell (55)
Cooling and thermal	Liquid cooling, chillers, thermal management	Data center builders	9	Vertiv (85), Trane (52)
Neoclouds	Rented GPU capacity, and nothing else	AI labs and hyperscalers	22	CoreWeave (97), Nebius (97)
Cloud and hyperscalers	General-purpose cloud with AI capacity inside it	Millions of customers	41	Microsoft (75), Alphabet (70)

Two things the table shows. The money enters at the bottom right, with the hyperscalers and neoclouds, and flows up the left column: cloud capex buys chips, chips need memory and networking, all of it needs a building, power and cooling, and the building needs copper. And the layer with the most stocks, semiconductors, is not the layer with the highest exposure. Neoclouds, networking and contract manufacturing average higher, because more of what they sell goes only into AI clusters.

How do AI infrastructure companies make money?

Three models, and they behave differently when spending slows.

Selling hardware per unit. NVIDIA, Micron, Vertiv, Arista. Revenue arrives when a chip, a module or a cooling unit ships, so it tracks capex almost quarter for quarter, in both directions. This is where the highest scores cluster and where a cut in guidance shows up first.

Selling capacity by contract. Neoclouds, data center REITs, and the power producers that sign power purchase agreements. Revenue arrives over years under a signed contract, often take-or-pay, so it lags capex on the way up and holds on the way down, until a contract is renegotiated. The neocloud explainer covers who owns the chips in that model, which is where its risk lives.

Earning a regulated or percentage return. Utilities, electrical contractors, engineering firms. A utility earns an allowed return on the substation it builds for a data center; a contractor earns a margin on the job. Steady, and diluted by everything else the company does, which is why most of this layer scores in the Growing or Peripheral tiers.

Why do two stocks in the same layer score so differently?

Because the layer says what a company sells and the score says how much of the company depends on it. Supercycle's AI Exposure Score is a 0 to 100 measure of how much of the business would go away if AI spending stopped. In the power layer, Fermi scores 92 because it exists to build an AI campus, and Chevron scores 8 because gas for data centers is a rounding error on an oil major. Both are "AI power stocks" in a headline. They are not the same exposure.

The distribution across the whole universe makes the point. Of 508 tracked companies, 29 score 90 or above, the Pure Play tier. 160 score below 30, Peripheral. The middle three tiers hold the rest. An "AI infrastructure" basket built from a headline list is mostly middle and bottom.

What is the catch?

• It is a capex cycle. The glossary has the pattern: suppliers are rewarded while the spending runs and punished when it peaks, and the durability of AI capex is the open question of this market. The per-unit hardware model feels a cut first.
• A handful of buyers. Seven companies account for most of the tracked spending. When one changes guidance, every layer below it moves, and the capex roundup shows none has cut yet in 2026, which is either reassuring or the thing to worry about.
• Concentration inside the layers. CoreWeave is the anchor tenant at a third of the neocloud layer. Microsoft anchors three of the four most-searched neoclouds. The neocloud analysis walks through who depends on whom.
• AI-washing. The glossary entry has the test: does money from the buildout actually land in this company's revenue? Supercycle's inclusion bar drops companies that only use AI internally, and the score handles the ones that pass the bar with a small share.

How to check whether a stock is really AI infrastructure

1. Find it on the leaderboard or its stock page and read the score and the rationale beside it. The rationale says what the company sells into the buildout and to whom.
2. Open its layer on the categories page and compare it with the top-scoring names in the same layer. A wide gap means AI is a small share of the business.
3. Check the buildout tracker for the company in a campus, capex or power-deal row. A stock that never appears there is being paid, if at all, from further down the chain.
4. Use the screener to filter the layer by score and size, which is the fastest way to separate the pure plays from the diversified names.

What to watch

The next capex round runs from late October to mid-December, and every layer above reprices on it. The IEA's 2030 number is the demand case for the power and cooling layers; the tracker's status column, where a campus moves from announced to under construction, is the same case in megawatts. And the tier counts on the leaderboard, 29 Pure Plays today, are the measure of how much of "AI infrastructure" is actually the real thing.