The year the board outran the chip
In 2026, the board that holds and links the chip outran the chip itself. While NVIDIA cooled to +38.9% and +5.1% across 2025-26, Samsung Electro-Mechanics, which makes the high-layer substrate (FC-BGA) those GPUs sit on, rose +614.3% in 2026, AT&S +398.7%, and Ibiden +192.2%. The returns the companies this wiki tracks (its nodes) posted in 2026 point to one thing: the bottleneck has moved from silicon to the package.
An AI accelerator's performance does not come from the chip alone. Two process steps bind the GPU die and the HBM stacks into one body: the high-layer substrate (FC-BGA) they sit on, and the advanced packaging (CoWoS and kin) that stitches them densely onto a silicon interposer. That the next constraint would move to the seam between chip and board, once memory filled, was already flagged (the start of that migration was traced in The memory supercycle). The 2026 returns confirm the move.
The substrate ran ahead of the chip
Start with the board. Ibiden rose +261% in 2025 and +192.2% in 2026; Austria's AT&S climbed +188.9% then +398.7%. In Korea, Samsung Electro-Mechanics rose +614.3% in 2026, LG Innotek +282.2%, Simmtech +392.9% in 2025, and Daeduck +250.2%; Taiwan's Unimicron posted +175.7% and +157.6%, and Nan Ya PCB +188.2% and +190.2%. The private Shinko (No.2 in FC-BGA) sits in the same layer.
What stands out is that these gains ran ahead of the chip itself. While NVIDIA managed only +38.9% and +5.1% across 2025-26, the substrate names jumped triple digits. That listed and private names across Japan, Austria, Korea and Taiwan all moved the same way is itself the clue. It points to a common force on the substrate step itself rather than any one company's results.
The real chokepoint is advanced packaging
Above the substrate sits another bottleneck: the advanced packaging that mounts GPU and HBM onto a silicon interposer and links them. TSMC's CoWoS capacity is that neck. TSMC rose +56% and +44.3% in 2025-26 and effectively rations the volume of its largest customer, NVIDIA (19% of revenue). The overflow spills to the OSATs: world No.1 ASE rose +65.6% and +159.9% (advanced-packaging revenue of $1.1B, up 144%), and Amkor +58.4% and +118%.
Dig one layer deeper and it is equipment. Besi, which makes hybrid-bonding tools, rose +35% and +73.1%; ASMPT, in thermocompression bonding, +47.6% and +89%. The narrower the process that joins the chips, the more bargaining power pools in the tools that make it possible. Substrates are spread across several countries, but CoWoS capacity is concentrated in essentially one place, TSMC, so this bottleneck is narrower geographically too.
And beneath it all, ABF
One layer below the substrate sits a quieter monopoly. The interlayer dielectric in every FC-BGA substrate is 80-90% Ajinomoto ABF film, and effectively 100% for high-end CPUs and GPUs. The film of a company known for seasoning is inside the package of every AI server in the world. Ajinomoto rose +63.4% in 2026, less spectacular than the substrate names, but the chain always hides a narrower neck one layer deeper. If the substrate is the chip's bottleneck, ABF is the substrate's bottleneck.
Why 2026 in particular
That the numbers cluster in 2026 is no accident. The order traced in Did the AI cycle propagate along the value chain?, chips (2023-24) → memory and equipment (2025) → substrates and components (2026), has its final square right here. The reason is physical. AI accelerators must fit a larger die and more HBM into one package, and high-layer, large-area substrates and CoWoS capacity carry low yields and long equipment and cleanroom lead times, so they cannot be added quickly. Demand steps up with training and inference scaling; supply grows only on a yearly cadence. That lag showed up as the triple-digit premium of 2026.
Verdict
The bottleneck moved from silicon to the package. The substrate (Ibiden, AT&S, Samsung Electro-Mechanics), advanced packaging (TSMC's CoWoS, ASE, Amkor), and the ABF beneath them (Ajinomoto): the entire layer that holds and links the chip re-rated ahead of the chip itself in 2026. The proposition holds again: a bottleneck's return scales with how badly it is needed, not with how famous it is.
But substrates and CoWoS both eventually expand, since capacity is a problem inside the chain. Once the packaging is filled too, the last constraint sits outside the chain entirely, in the physical world that runs all of it: power. In Power is AI's next bottleneck, the chain leaves the semiconductor behind.