Key points
- An AI data center is a machine with seven distinct layers: GPUs, memory, storage, CPUs, networking, packaging/manufacturing, and power. Each layer has its own public companies, its own bottlenecks, and its own investment story.
- The GPU gets the headlines, but the money flows through the whole machine. Memory and packaging have been the tightest constraints of the buildout.
- This page maps every layer to the companies that make it, with links to our deeper explainers on each.
- Run any name here through our stock score tool to see what its actual filings say.
Every AI model you've ever used runs on a physical machine, and that machine is more than a pile of Nvidia chips. It's a stack of components that each do one job, made by different companies, with different economics. Understanding the layers is the fastest way to understand where the money in the AI buildout actually goes, because a dollar spent on a data center gets split across all of them.
Here's the whole machine, layer by layer, with the public companies behind each part.
1. The GPU: the engine
The graphics processing unit does the actual AI math. Training a model means running trillions of parallel calculations, and GPUs are built for exactly that. Nvidia (NVDA) dominates this layer so completely that its data center revenue has become the single most-watched number in markets. AMD (AMD) is the credible second source, and the big cloud companies design some chips of their own, Google's TPU being the best-known example. A custom chip built for one specific job like that is called an ASIC, a term we cover in our AI investing glossary.
2. Memory: feeding the engine
A GPU is only as fast as the data reaching it. That's the job of HBM, high-bandwidth memory, stacks of memory chips sitting right next to the GPU die. Only three companies on earth make it at scale: Micron (MU), SK Hynix (Nasdaq: SKHY, after its July 2026 US listing), and Samsung. Memory has been the tightest bottleneck of the entire buildout, demand outran supply, and the memory makers went from boom-bust commodity businesses to the scarcest capacity in tech. The full story is in our HBM explainer.
3. Storage: where the data lives
Training data, model weights, and checkpoints have to live somewhere. Hard drives and solid-state drives are the unglamorous layer that quietly rerated as AI data centers multiplied. Seagate (STX) and Western Digital (WDC) make the drives; SanDisk (SNDK) makes the flash. Storage rarely leads a rally. It tends to follow the buildout with a lag, because drives get bought when data centers get filled, not when they get announced.
4. The CPU: the general manager
GPUs do the math, but a conventional processor still runs the show, orchestrating workloads, managing the operating system, moving data. Intel (INTC) and AMD supply the x86 server CPUs; Arm (ARM) licenses the architecture behind a growing share of custom data center chips, including the ones Nvidia pairs with its own GPUs. This layer is also where the US government planted its flag: it owns roughly 10% of Intel following an August 2025 deal that converted CHIPS Act funding into equity, part of the industrial policy push to bring chipmaking home.
5. Networking and optics: the nervous system
Thousands of GPUs training one model have to talk to each other constantly, and at these speeds that conversation happens over light. Arista Networks (ANET) builds the switches, Marvell (MRVL) and Credo (CRDO) the connectivity silicon, and Coherent (COHR) and Lumentum (LITE) the optical transceivers that move data between racks. As clusters grow, a larger slice of each data center dollar goes to simply wiring the machine together, which is why optics names now trade like AI stocks.
6. Packaging and manufacturing: putting it together
Someone has to actually build all these chips, and then fuse the GPU and its memory into a single working part. TSMC (TSM) manufactures most of the world's leading-edge AI silicon, the foundry model we explain in our TSMC piece. The step after fabrication, advanced packaging, bonds the memory stacks to the processor, and it remains a genuine chokepoint: packaging capacity, not chip production, has repeatedly set the limit on how many AI processors ship, and demand is still outrunning it. Amkor (AMKR) is the big outsourced packager, and Intel is spending heavily to become a packaging force, a push it underlined by hiring SK Hynix's former CEO, Seok-Hee Lee, to run that business.
7. Power and cooling: the constraint nobody priced in
All of this hardware eats electricity at a scale the grid wasn't built for, and the machines run hot enough that cooling is now an engineering discipline of its own. Vertiv (VRT) and Eaton (ETN) supply the power infrastructure, Trane (TT) the cooling, and generation names from Constellation (CEG) to the nuclear startups feed the load. We've covered why electricity became the binding constraint on the whole trade.
Why the layers matter more than the layer
Money rotates through this stack. When GPU names get expensive, investors hunt the next bottleneck, memory had its turn, then optics, then power. A supply problem in any single layer caps the whole machine, which is why a packaging shortage can move Nvidia's stock and a memory shortage can move everyone's. Reading the stack as one system, instead of chasing whichever layer is loudest that week, is the closest thing this trade has to a map.
Want to pressure-test any company on this page? Every name links to its live SEC filings above, and our stock score tool will grade its actual filings, insider activity, dilution, profits, debt, in about ten seconds.
Sources
- Related coverage: What is HBM, and why AI memory stocks matter
- Related coverage: What is a chip foundry? TSMC explained
- Related coverage: AI's power problem
- Related coverage: The AI investing glossary
- Company filings via SEC EDGAR, linked per ticker above
This is general market commentary and opinion, not investment advice. Markets can go down as well as up, and you can lose money. Always do your own research and consider speaking with a licensed financial professional before making any investment decision.
