The shape of the returns sorts the clocks
This wiki holds year-by-year returns for about 300 companies. Lay those numbers out by chain layer and read them at one moment, and even under the single word "AI" some layers have clearly passed their peak while others are just starting up. Set broker targets aside, and the returns and the chain position alone already separate what is early from what is late.
The reading rule is the one set in Many clocks: gains propagate along the chain with a lag from layer to layer. That narrows the judgment to two axes. A layer whose clock has not started still has room, and a layer that ran hardest carries the give-back. What follows sorts the chain along those two axes.
Clocks that haven't run, part one: the power execution layer
As The next bottleneck is power argued, power runs on the clock of plant permits and grid construction, slower than semiconductors. The returns confirm it. In 2026, with chips and memory cooling, the power execution layer re-accelerated instead. Vertiv, in datacenter power and cooling, rose +86% in 2026; Caterpillar in generation went from +60% to +69% year over year; GE Vernova in gas turbines was +71% and Quanta Services in grid construction +58%.
The layer still needs a distinction. The small modular reactors that ran on pre-commercial expectation (Oklo -27% in 2026, NuScale -31%) and the transformers already past their peak (HD Hyundai Electric decelerating from 300% through 120% to +7%) are in a different phase. The principle that only physical capacity turning into revenue keeps rising holds inside the layer too.
Clocks that haven't run, part two: humanoid parts and the compute follower
The earliest case in Frontiers rerate infrastructure before application is humanoids. Harmonic Drive, in joint reducers, was down -24% in 2025 and then started up +161% in 2026, with servo maker Yaskawa at +53%. The application side that builds the robots (Tesla -12%, UBTech -22%) is still held down. The infrastructure-first grammar is intact.
The compute layer has a follower clock too. While NVIDIA cooled from +171% in 2024 to +5% in 2026, AMD accelerated from -18% through +77% to +142% in 2026. The number-two in the same bottleneck layer is turning late, in the space the leader's cooling left open.
Clocks that ran hardest, part one: substrates and packaging
The 2026 explosion flagged by The bottleneck moved from silicon to the package concentrated in a single year. Samsung Electro-Mechanics rose +613% in 2026, AT&S +432%, ASE +160% and Ibiden +184%. The heat piled into the last cell of the chain's propagation.
That concentration is itself the caution signal. The principle left by The memory supercycle, that what rises fastest gives back fastest, weighs most heavily on this layer right now. Samsung Electro-Mechanics at +613% in one year is the extreme case.
Clocks that ran hardest, part two: the funding side cools first
Memory has passed its extreme as well. Kioxia printed +1,082% in 2025 and then +290% in 2026, while Nanya decelerated from +985% to +25%. The heat of the peak is leaking out.
The earlier signal sits at the demand end. As seen in Circular AI financing, this chain cools from the seller's side first. Microsoft, on the funding side, was -19% in 2026 and Oracle -28%, while Nebius, which takes that money and spends it, was still hot at +158%. If the payers cool first and the recipients later, that too is propagation with a lag, downward this time. It is the line that puts the whole AI demand chain on watch.
Clocks outside the chain
The map's boundary shows up in the same returns. Pure defensives with virtually no AI exposure drew money instead in 2026 as AI cooled: Coca-Cola +22%, Nestlé +20%, P&G +7%. Discretionary consumption fell on its own clock the other way. Nike, on weak Greater China sales and inventory correction, was -30% in 2026, moving opposite the rising chain.
Adoption does not change the clock either. Walmart and Starbucks run workflows on Azure and Gemini, but that is a tool to cut costs and is different from demand that lifts revenue. The AI money does not spread outside the chain confirms the boundary across more than twelve controls. Even full dependence on the cloud does not change the clock: Netflix, Airbnb and Disney run on AWS yet move on subscription, travel and content demand. How this map of consumption fans out from defensive through discretionary to luxury is detailed in Consumption does not run on one clock. This layer shows where the phrase "AI beneficiary" stops.
Korea and Asia consumption run on their own clocks too. Amorepacific tracks China's consumption recovery, falling three years running through 2021-2023 on the content ban and China's slowdown before rebounding on US K-beauty. Orion runs on the overseas volumes of China, Vietnam and Russia, up double digits three straight years in 2024-2026 and clearing its first KRW 500B operating profit. CJ CheilJedang, a supposed defensive, instead fell nearly every year on grain and raw-sugar costs, weak domestic demand and heavy debt. The exception is SK Telecom: a carrier that stacked an AI-infrastructure story (GPUaaS, AI data centers, an SK hynix HBM alliance) and re-rated +37.5% in 2025 despite a SIM-hack, the one name in the consumer batch to catch the AI clock.
The boundary does not end at consumption. Even the capital that funds the boom sits outside the chain. JPMorgan, the largest arranger of data-center debt, booked $95.4B of net interest income through 2026, but what lifted its stock was rates, not AI (the +44% of 2024 and +37% of 2025 were the deposit-margin spread of higher-for-longer). Bank of America moves on the same rate beta (-24% in 2022, +34% in 2024). At the opposite extreme sits Berkshire, which instead of buying chips parks $397.4B of cash in short-term Treasuries and waits, earning T-bill yields (+0.3% in 2026). The sharpest paradox is BlackRock: even as it supplies capital directly to AI data centers with Microsoft and NVIDIA through AIP, its stock returned just +6.6% in 2025, far behind the banks. To the sellers AI is revenue, to the users a cost, to the funders a rate clock. The one funding the boom does not capture it in its share price.
Verdict
Read from facts alone, the chain splits into two groups. The power execution layer and humanoid parts are early clocks still rising in 2026; substrates and memory are late clocks that already ran. The slowdown at the funding side is a leading warning laid over the top.
Where this reading diverges from broker consensus is the interesting part. Sell-side sees Korean battery materials (EcoPro BM, L&F) as high-upside buys, but the returns show they are still falling in 2026, at -46% and -16%. Buying the bottom and waiting for a clock that has not arrived collide on the same names. On the other side, Iridium after a +212% run and Samsung Electro-Mechanics at +613% face give-back, and here the consensus caution and the returns point the same way. The map neither buys nor sells. It shows which clocks have not run and which are done.