Last week we covered power semiconductors. The week before, photonics. This week the bottleneck moved beneath CoWoS.

For the last two years, the story of the AI chip shortage has had one main character and surprisingly, it isn't the chip itself.

It's been advanced packaging.

When people say NVIDIA "can't make enough GPUs," they usually mean TSMC can't finish enough of them.

The chip itself gets built fine (at least most of the time). The hard part is the step afterward, which is to stitch one or more GPUs together with its memory chips and other components into one working module.

TSMC calls that step CoWoS (Chip-On-Wafer-On-Substrate), and since 2023 it has been one of the most-watched constraints in the AI supply chain.

Our Constraint Radar flagged it again this week: CoWoS, TSMC’s advanced manufacturing capability, is sitting firmly in critical territory.

TSMC is expanding CoWoS capacity aggressively through 2026, with more planned for 2027.

But adding CoWoS lines does not automatically turn into finished AI servers. A CoWoS package still has to connect through an organic substrate, then move through module, board, and system assembly before it becomes deployable hardware.

If CoWoS capacity rises faster than the adjacent layers, the constraint shifts from TSMC’s packaging line to the next scarce input, a component almost no one outside the industry has heard of: advanced ABF substrates

Every finished AI chip sits on a substrate, which is a small and dense base board that the GPU and its memory chips are mounted directly on top of. It's the chip's foundation.

It does two unglamorous but essential jobs:

For high-end processors, that substrate is built around a material called ABF.

ABF stands for Ajinomoto Build-up Film. Yes, that Ajinomoto, the Japanese company best known for inventing MSG, the seasoning.

Decades ago its researchers discovered that a resin film derived from that same chemistry happened to be exceptional at insulating the ultra-fine wiring inside a chip substrate. It became the industry standard and never gave the position back. Today Ajinomoto supplies well over 90% of the ABF used in high-end CPU and GPU substrates.

That film is then built into finished substrates by a short list of specialist companies like Taiwan's Unimicron (the largest), Japan's Ibiden and Shinko, Austria's AT&S. A handful of firms, supplied by essentially one chemical maker.

A substrate's difficulty scales with two things:

AI is pushing both, fast. NVIDIA's current Blackwell GPU needs more substrate area than the Hopper generation before it.

The Rubin generation, arriving next, is expected to require a larger package footprint than Blackwell.

AI CPUs are also scaling the same way. As we mentioned in our latest CPU report, agentic CPUs require more cores and more CPUs stitched together. And every new generation also stacks in more wiring layers, which makes each substrate slower and harder to build.

So even if the number of AI chips held flat, the substrate required per chip is ballooning. But, as we know, the number of chips is not holding flat.

This demand for increased ABF substrates is already showing in the results these substrate makers just reported:

The price signal is just as loud. In May, Ajinomoto pushed through a roughly 30% price increase, and buyers absorbed it.

The stocks have re-rated hard as well. Unimicron is up close to tenfold over the past year, Nan Ya PCB roughly eightfold, and Morgan Stanley has upgraded both, calling ABF substrates the early stage of a demand up-cycle that lasts to the end of the decade. When these circuit-board suppliers start trading like chip designers, the price is telling you where the scarcity has moved.

The substrate makers are still constrained, but the more interesting bottleneck is moving below them, into the materials they need to turn installed capacity into shippable substrate capacity. 3 monopolies:

The tightest bottleneck inputs are unusually concentrated in a few Japanese materials suppliers.

Pays rent: the chip designers and their packaging partner. NVIDIA, AMD, the hyperscaler ASIC programs, and TSMC, all of whom need finished substrates to ship a single accelerator.

Captures rent: the substrate makers (Unimicron, Ibiden, AT&S, Nan Ya PCB) and, increasingly, the material owners beneath them like Ajinomoto, Nittobo and Mitsui Kinzoku.

Three things would loosen this: Ajinomoto announcing capacity expansion beyond the current 50%-by-2030 path, Nittobo accelerating T-glass to before mid-2027, or Mitsui lead times normalizing below 12 weeks from the current 16-20.

Until at least two of those move, expanding CoWoS capacity adds capacity on paper but not shippable units.

Here’s how the constraints moved in the other layers:

Printed circuit boards, the large boards a whole rack of chips plugs into, have been getting a lot of attention lately. Morgan Stanley's teardown of NVIDIA's next-generation Rubin rack put its total bill of materials at about $7.8 million, nearly double today's GB300 rack. Apart from the notable increase in memory and GPU, there were some overlooked components as well.

PCB content up an estimated 233% per rack, to roughly $117,000 from $35,000, and MLCCs, the tiny capacitors that condition power, up 182%.

PCB stocks have been rallying hard over the past few months, across Taiwan and China. Taiwan's PCB makers are guiding to a combined NT$150–200 billion (US$4.8–6.3 billion) of capex this year to chase the demand.

But that won't ease supply soon. Rubin's boards run past 40 layers, which makes it far harder to build than an ordinary PCB, and only a slice of the industry can make them. That capacity is already running hot, with order backlogs near three to four months. Before PCB supply loosens, it tightens.

A few important earnings this week. Read our pre-call briefs here and stay prepared.

Want the live data behind The Chokepoint? Tessara is the live supply-chain map of the AI build, for investors. We track what's binding in the supply chain and what it means for what you own. 300+ companies across compute, memory, foundry, networking, and power. Apply for early access →

See you next week,

Teng & Arvind

This article is for informational and research purposes only. It is not financial advice, investment advice, or a recommendation to buy or sell any security. Tessara Research does not publish price targets. The views expressed here reflect our analysis at the time of publication and may change as new evidence arrives. Readers should do their own research and consult a qualified financial adviser before making investment decisions.