Under the label “AI stocks” sits a whole supply chain: chip designers, foundries, equipment sellers, memory makers, server assemblers, all the way to the clouds and model companies. Chasing the list alone will leave you dizzy.
It is clearer to first see how the money flows. This piece starts from the cloud giants’ capex and follows one dollar through the chain: which links it passes, which layer has fat margins, and which is thin.
Where the Money Comes From: the Cloud Giants’ Capex
The source is the cloud giants’ “capital expenditure” (capex, the long-term spending on equipment and data centers).
Per each company’s official guidance from July 2026, Amazon plans about 220 billion dollars, Alphabet (Google) 195–205 billion, Microsoft about 175 billion, and Meta 130–145 billion, for a 2026 total of about 720–745 billion dollars across the four. Most of this money goes to AI chips and data centers. It is the pie the whole supply chain is splitting.
Core-Data Snapshot
The table compares company-wide profit metrics (results published through September 2026).
| Link | Representative Company | Margin / Profit Profile | Nature |
|---|---|---|---|
| Source of the money | Amazon, Alphabet, Microsoft, Meta | 2026 capex guidance about 720–745 billion dollars combined | Official guidance |
| GPU design | NVIDIA | Gross margin about 75% (Q2 FY27) | Official results |
| Wafer foundry | TSMC | Gross margin 67.7% (2026 Q2) | Official results |
| Lithography equipment | ASML | Gross margin 54.0% (2026 Q2) | Official results |
| HBM memory stocks | Micron, SK hynix | Micron gross margin 86.8% (Q4 FY26); SK hynix operating margin 76% (2026 Q2) | Official results |
| Full-system ODM | Quanta, Wiwynn | Gross margin about 5% (Quanta), 9.3% (Wiwynn), 2026 Q2 | Official results / earnings call |
| Model companies | OpenAI, Anthropic | Fast revenue growth, still investing heavily | Official / media |
The Money Map: One Dollar’s Journey
Upstream, the shovel-sellers (fattest margins). The money first flows to AI chips: GPU leader NVIDIA designs them, TSMC manufactures them, TSMC needs lithography machines from ASML, and the chips are paired with HBM high-bandwidth memory. Gross margins in this layer run from 50% to over 80%. This is where the “selling shovels” idea comes from: whichever model wins, the chip and equipment sellers get paid first.
Midstream, the system assemblers (thin margins). Chip modules go to Taiwan’s AI server ODMs to be built into full systems and racks. In Q2 2026, Quanta’s and Wiwynn’s revenue grew anywhere from over 20% to more than double year over year, yet their gross margins stayed in single digits.
The supporting links. A full rack also needs cooling and liquid cooling, power, networking, and advanced packaging. Many Taiwanese firms take part, with margins that vary widely.
Downstream, the model companies spending the money. At the end of the chain are companies like OpenAI and Anthropic that buy compute to train models. According to the FT, OpenAI’s annualized revenue was about 50 billion dollars at the end of September 2026 (its own measure). Anthropic officially reported over 47 billion dollars in May 2026, and according to Bloomberg it passed 65 billion by the end of July. The two count partner sales differently, so the numbers can’t be compared directly. According to the WSJ, OpenAI is still losing a lot of money; Anthropic’s adjusted operating profit reportedly turned positive in Q2 2026 (excluding stock-based compensation). Neoclouds that rent GPUs to these companies, like CoreWeave, posted 2.58 billion dollars in revenue in Q2 2026 and still had a net loss of 630 million dollars.
Why Margins Differ So Much
The key is “barriers” and “business model.” Only a handful of companies can make the upstream chips, equipment, and memory, so they have pricing power. Many more firms can do midstream assembly, and they compete on scale, efficiency, and delivery time, so the processing fee is thin by nature.
The business model also dilutes margins. ODM revenue often includes expensive parts that just pass through: their value counts as revenue, but the ODM only really earns the assembly and integration step. Wiwynn is an example. From April 2026, some customers’ memory moved to a “procurement on behalf of the customer” model, so that amount no longer counts in revenue or cost. Together with engineering service (NRE) income from new products, its Q2 gross margin rose to 9.3%. Change the accounting, and the margin changes with it. With ODMs, look at which step they actually earn from.
How to Read “AI Stocks”
The market puts related companies from every layer above into the “AI stocks” bucket. Knowing which layer of the money map a company sits on, and whether that layer’s margins are fat or thin, is more useful than chasing the list. But margins are past results, not a forecast, and many claims that “a certain Taiwanese firm maps to a certain global giant” are speculation by institutional investors, not company announcements. This article only describes the supply chain’s division of work and money flow; it does not list beneficiary stocks, rank stocks, or constitute investment advice.
Penchan’s summary
To see the funding web where chipmakers and clouds invest in each other, read the AI Circular Financing Map. To see whether this counts as a bubble, read Is AI a Bubble. To see how NVIDIA outsources manufacturing to Taiwanese firms, read the NVIDIA Supply Chain. For all eight gates of the hardware chain, go back to The AI Hardware Supply Chain, End to End.
Upstream shovel-sellers earn high margins behind technical barriers, midstream assemblers compete on scale for thin margins, and downstream model companies are still trading heavy investment for growth. With AI stocks, first find which layer of the chain a company sits on, then look at that layer’s margin structure. That is more practical than chasing a theme list.