Key points
- The AI trade is a stack: toolmakers, foundries, memory, networking, data centers, power, hyperscalers.
- The bottleneck keeps moving, from GPUs to memory to power, and whoever owns it gets the pricing power.
- Hyperscaler capex is the number that matters most; the chain's concentration makes it fragile.
Everyone wants to "buy AI," but almost nobody can tell you what that means past a single ticker. The AI trade is not one stock or even one sector. It is a supply chain that starts with literal sand and ends with a chatbot answering your email, and every layer has its own margins, its own bottleneck, and its own way of blowing up when the layer above or below it stumbles.
This is the map. We go top to bottom, and at each stop we flag the one thing that matters most: how that layer pushes and pulls on the others. One rule to anchor everything: money flows down the stack (the hyperscalers write the checks) and product flows up (a wafer becomes a chip becomes a server becomes a service).
1. The toolmakers
Before anyone makes a chip, someone makes the machines that make the chip. The most defensible real estate in all of AI, because the moats here are measured in decades.
ASML (ASML) is the whole ballgame: the only company on earth that makes the EUV lithography machines needed to print the smallest transistors. Applied Materials (AMAT), Lam Research (LRCX), and KLA (KLAC) handle the other steps (deposition, etch, inspection). Synopsys (SNPS) and Cadence (CDNS) make the software chips are designed in; Arm (ARM) licenses the underlying CPU blueprints.
How it ripples: This layer is the leading indicator for everyone below it. When the foundries order more equipment, it shows up here a year before the chips exist. Watch it to see what the foundries believe about demand twelve to eighteen months out.
2. The foundries and the chips
TSMC (TSM) is the center of gravity of the entire trade: it physically manufactures the leading-edge chips for almost everyone who matters. Samsung is number two; Intel (INTC) is trying to build a real foundry rival, and whether it succeeds is one of the sector's biggest open questions.
On design, Nvidia (NVDA) is the franchise, and its real moat is CUDA, the software lock-in. AMD (AMD) is the credible challenger. But the quiet threat to Nvidia is custom silicon: the hyperscalers hate paying Nvidia's margins, so they design their own chips (Google's TPU, Amazon's Trainium) with help from Broadcom (AVGO) and Marvell (MRVL). Every time a customer builds its own accelerator to escape Nvidia, Broadcom often gets paid.
How it ripples: The accelerator market is not a one-horse race. It is Nvidia versus a coalition of its own biggest customers, and that tension sets pricing for the whole stack.
3. Memory and packaging (the real bottleneck)
A GPU is useless without memory fast enough to feed it. That memory is HBM, and only three firms make it: SK Hynix (the leader), Samsung, and Micron (MU). When people say AI is "supply constrained," memory is usually what is constrained.
Even with the GPU and the HBM, you have to bond them together. TSMC does this with a process called CoWoS, and its packaging capacity has been one of the single biggest limits on how many accelerators actually ship (Amkor (AMKR) and ASE (ASX) help here too).
How it ripples: This layer is the throttle on the entire trade. The number of GPUs the world gets is set by HBM supply and packaging capacity, not by Nvidia's design office. When GPU lead times stretch or shrink, the cause almost always sits here.
4. Systems, networking, and optics
A loose GPU does nothing. Super Micro (SMCI), Dell (DELL), and HPE (HPE) build the servers. Then comes the part most people miss: training a model means thousands of GPUs working as one, so the network is the computer. Arista (ANET) and Cisco (CSCO) sell the switches, Nvidia sells its own networking, and the telecom names (Nokia (NOK), Ciena (CIEN)) are increasingly in the mix as connecting data centers becomes its own demand story. Broadcom shows up again, making the switch chips inside the gear.
All that data moves between racks as light, which means optical transceivers from Coherent (COHR), Lumentum (LITE), and Fabrinet (FN). And it runs hot enough to need liquid cooling, where Vertiv (VRT) is the marquee name.
How it ripples: This layer scales with cluster size, not just chip count, and clusters keep getting bigger. That makes optics and networking a high-beta way to play the same trend as Nvidia, with more upside in a build-out and more pain in a slowdown.
5. The buildings (data centers and neoclouds)
The chips have to live somewhere with power and cooling. Equinix (EQIX) and Digital Realty (DLR) are the landlords of the internet. The neoclouds, CoreWeave (CRWV) and Nebius (NBIS), do one thing: buy huge GPU fleets and rent them by the hour. They borrow heavily to buy depreciating chips, which is fantastic in a boom and brutal if utilization ever cracks. A pure-play bet on compute staying scarce.
How it ripples: This is where abstract AI demand turns into a signed lease and a power contract, which makes it the cleanest read on whether the build-out is real.
6. Power and energy (the constraint nobody priced in)
Here is the plot twist. We spent two years worried about chips. The real ceiling might be electricity. An AI data center draws power on the scale of heavy industry, which turned a pile of sleepy utility stocks into AI plays.
Vertiv (VRT), Eaton (ETN), and Schneider make the power and cooling gear between the grid and the rack: arguably the best risk-adjusted corner of the map. On generation, Constellation (CEG), Vistra (VST), and Talen (TLN) became darlings when hyperscalers started signing deals to buy their output directly, and gas-turbine maker GE Vernova (GEV) caught the same wave.
How it ripples: Power is now a hard ceiling on everything above it. A hyperscaler can own all the GPUs it wants, but with no power connection the cluster does not run. Watch the interconnection queues and power deals: they gate the whole map.
7. The buyers (hyperscalers) and apps
The bottom of the stack and the top of the money. Microsoft (MSFT), Alphabet (GOOGL), Amazon (AMZN), Meta (META), and Oracle (ORCL) set the tempo for everything. Their combined capex is the single most important number in the AI trade. They are also the strangest players on the map: simultaneously Nvidia's biggest customers and, through their custom chips, its biggest long-term threat.
How it ripples: Every dollar that moves up the stack starts as a hyperscaler capex decision. That is also the bear case in one sentence: if the apps at the bottom cannot eventually earn enough to justify the spending at the top, the capex that feeds the whole stack gets cut, and the pain flows all the way back up.
The rules that move the whole thing
- Capex is the heartbeat. Hyperscaler spending sets the tempo. It is the one number to watch above all others.
- The bottleneck moves. First GPUs, then HBM and packaging, now increasingly power. The constraint never disappears, it relocates, and whoever owns it has the pricing power.
- Leverage cuts both ways. The further you sit from the end customer and the more your content scales with cluster size (optics, networking, neoclouds), the more violently you move in both directions.
- Concentration is the systemic risk. One dominant lithography maker, one dominant foundry, three memory makers, one dominant packaging process. That concentration is why the trade has been so profitable and exactly why it is fragile.
That is the map. Bookmark it. Every story on the Wire sits somewhere on it, and once you can place a headline on the right layer, you stop reacting to news and start reading it.
AIStockWire publishes briefings and analysis on the AI economy. Nothing here is investment advice. Do your own research before making any investment decision.