“AI stocks” has been one of the hottest keywords in the Taiwan market in recent years. But under those two words sits an entire supply chain: the chip designers, the foundries, the equipment sellers, the memory makers, the server assemblers, all the way to the clouds and the model companies. The range is so wide that chasing the list alone will leave you dizzy.
There’s a clearer way to do it: rather than rushing to ask “which stocks count as AI stocks,” first understand “how the money flows.” This piece draws one money map, starting from the cloud giants’ capex and tracing which links one dollar passes through, which layer has fat margins, which is thin, and which is still investing heavily. It’s one entry in the “AI Industry Watch” series. To see the funding web where chipmakers and clouds invest in each other, pair it with the AI Circular Financing Map; to see whether this counts as a bubble, pair it with Is AI a Bubble.
Where the Money Comes From: the Cloud Giants’ Capex
The source of this chain is the “capital expenditure” (capex, the long-term investment in buying equipment and building data centers) that a handful of deep-pocketed cloud giants are pouring out.
Alphabet (Google), Amazon, Microsoft, Meta, and Oracle — these five — have official 2026 capex guidance that adds up to roughly 745 to 775 billion dollars (each has slightly different fiscal years and definitions, so this is an approximate sum). Amazon alone guides to about 200 billion dollars, Microsoft about 190 billion, and Alphabet about 180 to 190 billion. A large share of this money goes toward buying AI chips and building data centers packed with servers. In other words, the big pie everyone in the AI supply chain is splitting traces back to these few companies’ capex.
Core-Data Snapshot
The table below places the “margin thickness” of several key links on the money map side by side, where the contrast is clearest. Gross margins are each company’s most recent quarterly figures.
| Link | Representative Company | Margin / Profit Profile | Nature |
|---|---|---|---|
| Source of the money | Cloud Big Five capex | 2026 total about 745-775 billion dollars | Official guidance |
| GPU design | NVIDIA | Company-wide gross margin about 75% (Q1 FY27) | Official results |
| Wafer foundry | TSMC | Gross margin about 60% (FY2025) | Official results |
| Lithography equipment | ASML | Gross margin about 53% | Official results |
| HBM memory | Micron, SK hynix | Micron gross margin over 74%; SK hynix operating margin over 70% | Official results |
| Full-system ODM | Quanta, Wistron, Wiwynn | Gross margin only about 5%-7.6%, but revenue up around 60% YoY | Official / earnings call |
| Model companies | OpenAI, Anthropic | High annualized revenue growth (media estimate), not yet profitable | Media estimate |
The Money Map: One Dollar’s Journey
Follow one dollar of capex as it flows downward, and it passes through several layers, each with a very different “earning profile.”
Upstream, the shovel-sellers (fattest margins). The money first flows to the link that designs and manufactures AI chips. GPU leader NVIDIA runs a company-wide gross margin of about 75%, the biggest collector of this wave; it hands its chips to TSMC for foundry work, and TSMC’s gross margin is about 60%; TSMC needs lithography equipment from ASML, and ASML’s gross margin is about 53%; the chips need to be paired with HBM high-bandwidth memory, where Micron’s gross margin tops 74% and SK hynix’s operating margin tops 70%. What this layer has in common is high technical barriers and few substitutes, which is why it can earn high margins. This is where the “selling shovels” analogy comes from: no matter which model wins, those selling the chips and equipment pocket the money first.
Midstream, the system assemblers (very thin margins). Once chips and memory are packaged into GPU modules, they go to Taiwan’s AI server ODMs to be assembled into full systems and racks. This part is interesting: the revenue at Quanta, Wistron, Wiwynn and others routinely grows over 60% year over year, and the headline numbers look great, but the gross margin is only single digits (Quanta and Wistron around 5%, Wiwynn around 7.6%). The reason comes in the next section.
The supporting links. A full rack also needs cooling and liquid cooling, power, networking, and advanced packaging, and a host of Taiwanese firms take part in these links too, with margins varying widely. For instance, Vertiv, which does power and cooling, has an operating margin of about 20%, landing between the upstream and the ODMs.
Downstream, the model companies spending the money (still burning cash). At the end of the chain are companies like OpenAI and Anthropic that use compute to train models. Their annualized revenue is growing extremely fast (media estimates put each at around 25 and 30 billion dollars, and neither is an audited figure), but they have simultaneously signed astronomical compute contracts, and per media estimates they are still investing heavily and not yet profitable. Even the neocloud players that rent GPUs specifically to them, like CoreWeave, posted 2.08 billion dollars in revenue in Q1 2026 with rapid sequential growth, yet still ran a net loss.
Why Margins Differ So Much
Same chain — upstream margins of 70%, midstream margins in single digits — how can the gap be this big? The key lies in “barriers” and “model.”
The upstream chips, equipment, and memory have extremely high technical barriers, with only a handful of companies able to do them, giving them pricing power at the negotiating table, so margins are fat. Midstream system assembly is different: there are relatively many firms that can assemble servers, and the competition is over scale, efficiency, and delivery time, so the processing fee is inherently thin.
More crucially, many AI servers use a “consignment” model: the most expensive GPUs are supplied directly by the cloud customers, and the ODM only handles assembly. The revenue on the books pours in the value of those costly GPUs, so it looks huge, but what the ODM truly earns is just that small assembly-and-integration segment. So with ODMs, revenue growth does not equal earning a lot — you have to look at which segment they earn from and whether the margin moves along with it.
How to Read “AI Stocks”
“AI stocks” is an evergreen theme in the Taiwan market, and the market throws in all the related companies above — chips, equipment, memory, packaging, ODMs, cooling, power, networking.
When looking at this group, hold to one principle: understanding which layer of the money map each company sits on and whether that layer’s margin structure is fat or thin is more useful than chasing the list. But remember a few things: the gross margins cited above are financial facts from “the past up to the most recent quarter,” not a prediction of the future; whether a company can keep benefiting involves orders, yield, competition, and stock valuation; and many claims that “a certain Taiwanese firm maps to a certain link of a certain global giant” are analyst speculation, not company announcements. This article only describes the supply chain’s division of labor and money structure; it does not compile a list of beneficiary stocks, nor rank individual stocks, nor constitute investment advice.
Key Takeaways for This Piece
After looking at this money map, first remember the skeleton: AI money starts from the cloud Big Five’s nearly trillion-dollar capex, flows through GPUs, wafer foundry, equipment, memory, packaging, and server contract manufacturing, finally reaching the model companies.
The upstream shovel-sellers earn high margins on technical barriers (NVIDIA, TSMC, ASML, the memory makers); the midstream Taiwanese ODMs that assemble full systems compete on scale for thin margins (single-digit margins); the downstream model companies are still trading heavy investment for growth. When looking at AI stocks, understanding this structure is more practical than chasing the theme list.
To see the funding web behind this chain, 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 its manufacturing to Taiwanese firms, read the NVIDIA Supply Chain; to look back at all eight gates of the hardware chain, return to The AI Hardware Supply Chain, End to End.