The question this wiki exists to answer
This wiki begins with one question: what rose first, and what hasn't yet? The AI cycle looks as if it pushed all 259 nodes this wiki tracks in the same direction, but lay the yearly returns out node by node and the order splits apart. NVIDIA rose first, +239% in 2023 and +171.3% in 2024, then cooled to +38.9% in 2025 and +5.1% in 2026. SK hynix only turned up later, +361.1% in 2025 and +194.7% in 2026. Micron came late too, +240.5% then +325.5%. Same chain, different clocks.
To read that order, you first have to fix where the money enters. Without knowing the chain's mouth, you cannot say what ought to move first, nor whether a layer that hasn't risen yet will rise late or never. So the chronicle's opening chapter starts not with the order of returns but with the entrance of the money.
The entrance looks wide
The demand side looks diverse. OpenAI (the peak of generative-AI demand), Anthropic (run-rate revenue around $30B), Oracle (Stargate, roughly $300B and 4.5GW), CoreWeave (a backlog near $100B): many names, each betting on different chips and different clouds. On the surface, demand runs in several streams.
But trace these names back through their balance sheets and they converge on a handful of payers. Oracle's Stargate is underwritten by OpenAI's contract, and OpenAI's ability to pay leans on its Azure commitment to Microsoft. CoreWeave's backlog is filled by NVIDIA's investment and hyperscaler volume. They appear to enter through different doors, but follow the source of the money and the doors are few.
The money enters through one door: hyperscaler capex
The ultimate payers are four. Microsoft cash capex of $64.6B (FY2025, +45%), Alphabet $91.4B (+74%), Amazon $131.8B (+59%), Meta $69.7B (+87%). Those four alone poured about $357B into infrastructure in 2025, and 2026 guidance is larger still: Alphabet $175-185B, Amazon around $200B. Even Oracle's Stargate and CoreWeave's backlog are ultimately collateralized by this capex budget and by the GW commitments of OpenAI and Anthropic. AI demand looks diverse, but the money enters through a single door: hyperscaler capex.
A narrow entrance is a risk signal in itself. If demand that looked like many streams is really a function of four or five budget lines, then when those budgets shake, the whole chain turns back in the same direction.
Look anywhere on the chain and that entrance is reflected
Read the concentration node by node and the same narrow entrance keeps reappearing. NVIDIA's revenue of $215.9B (FY2026, +65%) comes from direct customers at 22% and 14%, with hyperscalers as the end customers. Move upstream and it holds. SK hynix's single largest customer is NVIDIA at about 27% (roughly ₩11T), Micron's largest single customer is 17% (its cloud segment), TSMC's largest is NVIDIA at 19%, and Broadcom's top five are 40% with one distributor at 32%. Each node looks like an independent business, yet the largest block of its revenue points at the same few names. If NVIDIA's end customer is a hyperscaler, and the largest customer of SK hynix, Micron and TSMC is in turn NVIDIA, then the revenue of all three layers traces back to a single line: the capex of the same four firms. The chain is wide, 259 nodes, but run its revenue back and it meets at the same entrance.
The entrance is even self-referential
This entrance is not even purely external demand. NVIDIA invested $2B in CoreWeave, and CoreWeave is OpenAI's second-largest customer. Amazon supplies Anthropic with roughly 500,000 Trainium2 chips to train Claude (Project Rainier) while being an equity investor in it. That loop, in which the seller supplies equity and volume to the buyer, is treated separately in AI invests in itself. When the entrance is narrow and, on top of that, self-referential, the entire supply lined up above it rests on the same assumption.
Verdict
Demand scatters across many names, but the money enters through one point: hyperscaler capex. That narrow entrance is reflected in every concentration figure along the chain, and the self-referential loop tightens the concentration further. A narrow entrance means the demand risk shared by all 259 nodes is gathered at a single point.
The question that remains is this: did money entering through so narrow a door really spread all the way down the chain? Whether that propagation was real, and whether it had an order, is tested with returns in Did the AI cycle propagate along the value chain?.