Statementdog (財報狗) EP541 Notes: Once Chips Got Big, Round Wafers Stopped Paying

Magical-realist oil painting cover: an enormous square glass panel suspended in a fog-filled cleanroom, its surface reflecting countless small square chips, beside it a round wafer left with only its trimmed corners, light passing through the gap between the square and the circle

Without compass and square,
you cannot make a circle or a square.
—— Mencius, “Li Lou I” (Warring States period); translation mine

These are my personal notes after listening to Statementdog (財報狗) EP541 (released 2026-07-26), an industry breakdown of panel-level packaging. Not a transcript, not official show content. Please support the original if you want the full version. What follows is what the episode sparked, plus my own reading.

What this episode is about, in one line

Geometry has started sending invoices.

Wafers are round. Chips are square. Cutting squares out of a circle wastes the corners, and that has been true since the beginning of the industry. Nobody cared, because the chips were small and so were the wasted corners. AI changed the dimensions. Once a single packaged die swells to several times the reticle limit, the waste stops being a rounding error and becomes half your capacity.

The arithmetic in the episode is vivid. At the current generation of roughly 5.5 reticles per chip, a 300mm wafer yields somewhere around nine units, for an area utilisation of about 58 percent. More than four-tenths of the wafer holds nothing. The obvious fix follows: if you’re placing square things, use a square carrier, and make it big. That is panel-level packaging.

The main points

The driver is chip size, not a technology upgrade. Panel level is not a “better grade” than wafer level. The episode makes a point of correcting this, because the Chinese term for “level” also reads as “rank.” It’s a classification, not a ranking. Panel level is having its moment purely because chips got too big for round carriers to make economic sense.
The real barrier is called warpage. Packaging stacks many materials, each with a different coefficient of thermal expansion. Make the area large, run it through high-temperature steps repeatedly, and the whole board deforms, so the lithography and the vias no longer line up. The benefit of scaling up the panel is linear. The difficulty is not.
Tooling built for circles doesn’t transfer cleanly to squares. Spin coating relies on centrifugal force to even out thickness, which cannot produce even corners on a square. Plating concentrates charge at the four corners. A lot of what looks like “just a different shape” means rebuilding a process line.
Glass solves two problems at once. Its coefficient of thermal expansion is very low, which holds warpage down. Replacing today’s organic resin materials with glass also reduces high-frequency signal loss in transmission.
Three camps, three tempos. Panel makers (the episode focuses on Innolux) bring years of experience moving large glass sheets and coating films, and are entering from the low- and mid-end where linewidth requirements are looser; they’re already in volume production. OSATs have been at this longest — the episode notes Powertech starting around 2018 and committing serious capex this year to a 500 by 500mm line, with ASE beginning at 300 by 300mm. The foundries are going after the hardest tier: TSMC is working on moving its interposer process onto a square worktable and evaluating glass carriers, while Intel picked the more aggressive route, skipping the current interposer approach to bet directly on glass.
The money at the high end is a long way out. The episode’s estimate is that meaningful revenue contribution from high-end panel-level packaging at TSMC and Intel arrives somewhere around 2029.

Extensions

One: when an industry’s bottleneck shifts from performance to geometry, it usually means a new phase.

For twenty years the story in semiconductors was linewidth: make the transistor smaller. Nothing in this episode sits on that axis. It’s about shape, area utilisation, and how materials expand when heated. When a mature industry starts rebuilding its infrastructure to satisfy geometry, the room on the original axis has narrowed and value is migrating sideways.

That’s a useful signal for anyone hunting bottlenecks. Bottlenecks don’t stay in one layer. While everyone watches leading-edge process nodes, the thing actually capping output may have become “how do you keep a square flat.” And problems of that second kind usually get solved by an entirely different set of companies.

Two: glass solving two problems at once matters more than it sounds.

A new material that solves one problem gets adopted at a speed set by how much that problem hurts. A material that solves two unrelated problems at once — warpage and signal loss — draws pressure from two different departments, and faces far less internal resistance. Historically, substitutions like that arrive faster than the market expects.

I’d apply one brake to that. Willingness to adopt is not the same as speed of ramp. Yield sets the tempo, and yield is both the hardest thing to observe from outside and the easiest thing for a company to describe optimistically. So the sensible way to track this isn’t listening for announcements. It’s watching whether capex actually lands and whether line dimensions actually step up.

Three: “meaningful contribution around 2029” is narrative fuel, not a holding period.

Semiconductor roadmaps have a habit of being revised, in both directions, essentially every cycle. That isn’t dishonesty. Before yield crosses a threshold, the timeline genuinely is an estimate.

The problem is that the market trades it years early. A theme that contributes revenue in 2029 can move share prices in 2026, and for those three years what you own is not earnings. What you own is the assumption that the schedule doesn’t slip. That’s the part of this episode I’d hold onto: not that you shouldn’t buy a trend, but that you should know whether you’re buying low-end volume that already collects money or a high-end promise still sitting in a lab. Those look nearly identical in a press release and completely different in the accounts.

A note on where this leaves things

I’m increasingly convinced that the valuable part of industry research isn’t knowing a trend exists. It’s knowing which step of the trend you’re standing on.

Panel-level packaging is the perfect example. It is simultaneously a shipping business (low and mid-end), an engineering problem in pilot production (high end), and a materials bet where the route is still being chosen (glass). All three share one name, and their risk profiles have nothing in common. In conversation they get blended together, so “this company does panel-level packaging” sounds like a single statement while potentially meaning three very different things.

Unpacking a hot term back into its actual stage is, at the moment, the most worthwhile piece of dull work I know of in investment research.

What this episode is about, in one line

The main points

Extensions

A note on where this leaves things

Further reading