The lag itself draws the chain
Lay the yearly returns of the roughly 210 companies this wiki tracks (its nodes) side by side, and a rule emerges: the best year slides back one layer at a time. After NVIDIA led with +239% in 2023, Lam Research only turned +139.5% in 2025, and Samsung Electro-Mechanics reached +614.3% in 2026. Chips rose first; a year later equipment and memory turned; a year after that, substrates and components followed. The propagation was real, and it had an order. The best year sliding back one layer at a time is itself the evidence that the rise was not luck but movement along the chain.
The money of the AI chain enters through a single door: hyperscaler capex. Meta's cash capex hit $69.7B in FY2025, up 87% year over year; Microsoft's reached $64.6B, up 45%. The demand of the entire chain hangs on a handful of balance sheets (just how narrow that entrance is was taken up in Where AI capex begins). That narrow entrance leaves a question. Did the money really propagate all the way down the chain, or did it heat a few nodes near the door and fade? The annual returns of some 210 nodes are the material to test it.
The chips rose first
The rise began at the bottom of the chain, the chips closest to final demand. NVIDIA led with +239% in 2023, with AMD (+127.6%) and Broadcom (+104.9%) following the same year. Half a beat downstream, the foundry hub TSMC peaked a year later, at +92.9% in 2024, as the money crossed from design into manufacturing.
Watch NVIDIA's trajectory: +239% (2023) → +171.3% (2024) → +38.9% (2025) → +5.1% (2026). The best year sits at the very front of the chain, and the gains ease the further back you go. Money arrives first where the bottleneck is sharpest and substitution hardest. The chip touching final demand was that place, so it rose first and rose most.
Equipment and memory turned a year later
While the chips ran, the layer above them fell. In 2024 ASML was -7.7%, Lam Research -6.9%, Tokyo Electron -4.4%. Demand was already running, yet upstream trailed downstream, and that inversion is the core evidence for propagation. Then, in 2025, that upstream turned as one: Lam Research +139.5%, Tokyo Electron +61.6%, ASML +55.8%, Applied Materials +59.6%. Lam had been down the prior year, then jumped +139.5% the next, showing the lag as plainly as anything on the chain.
In that same 2025, memory climbed hardest of any layer on the chain: SK hynix +361.1%, Micron +240.5%, Kioxia +1082.1%. That Advantest, the near-monopoly of HBM test, printed +196% (highest among the tool makers) is of one body with that memory demand. Over the same window the demand-side big tech stayed quiet: Meta +13.1%, Microsoft +15.6%, a world away from their triple digits of 2023. The rise was plainly moving from downstream to upstream.
Substrates and components rose last
The last layer to move was the physical infrastructure that holds, links and cools the chip. Samsung Electro-Mechanics is the archetype: flat at -2.5% in 2024, then +110% in 2025 and a further +614.3% in 2026. Ibiden, a global leader in FC-BGA substrates, was pressed down -38.6% in 2024 before rebounding to +261% in 2025; AT&S followed at +398.7% in 2026, Simmtech at +392.9% in 2025. Ajinomoto, the near-monopoly of the ABF dielectric those substrates depend on, had its best year of six at +63.4% in 2026.
The optics that link chips trace the same path: Lumentum +339.1% (2025), Coherent +120.6% (2026). Down to the power and cooling that run the datacenter (Vertiv +101% in 2026, Delta +183.8% in 2025), the layer differs but the rhythm is identical: the chip rises first, then the infrastructure that holds, links and cools it. Why the substrate became this cycle's bottleneck is taken up in The package matters as much as the chip.
The lag is the lag of the financial statements
Why does each layer slip a year? Because demand is already running, but the capacity investment to meet it takes time to land in revenue and earnings. When a hyperscaler orders GPUs, NVIDIA's results jump at once; but the orders for the tools that make those GPUs, the memory built on those tools, and the substrates those chips sit on arrive quarters and years apart. The re-rating lag is simply the physical time it takes for cash to flow down the chain. Structure sets the order, and time unfolds that order one layer at a time.
Verdict
The rise moved along the chain from chips (2023-24) to equipment and memory (2025) to substrates and components (2026), over one to three years. The prediction that the best year slides back layer by layer is confirmed in the data. The hypothesis that value-chain structure explains the order of the rotation holds. Money from a narrow entrance did, in fact, reach the far end of the chain.
What remains is amplitude. The propagation passed through every layer, but one layer is where the re-rating showed up most extreme of all. As Kioxia's lone +1082.1% suggests, how a single point of HBM demand re-rated the entire memory asset class is taken up next in Memory: the steepest beneficiary of AI.