LG Innotek showed a package substrate bigger than eleven reticle-limited dies
Its first flip-chip substrate for AI accelerators runs past 100mm on a side, with mass production targeted for 2027 and the glass cores that keep it flat for 2028.
The thing that limits how big an AI package can get is not the chip. It is the board underneath it.
LG Innotek showed its first flip-chip ball grid array built for AI accelerators at the KPCA Show in Incheon on 9 September 2026. The substrate runs more than 100mm on a side, which the company says is about six times the area of the PC substrate sitting next to it on the stand. Mass production is targeted for 2027.
Set that against the hard ceiling on a single piece of silicon. A conventional EUV exposure field caps any single die at roughly 858 square millimetres, and no amount of ambition moves that number. The new substrate is close to twelve times that area.
Which is the point. This is a carrier for a package of many dies and stacks of memory, not for a chip.
Run the arithmetic in the release and two products turn out to be one
LG Innotek also plans to show an 85mm square board, and separately an "ultra-large-area" sample about 40 percent bigger than that one. Do the multiplication and 40 percent more area than an 85mm square is a square of about 100mm.
They are the same board, described twice.
If you are counting product lines from press releases, count carefully. (Trade-show copy does this more often than it admits.)
The technology that matters is not the size
Two other things on the stand do more work than the dimensions.
Chip embedding puts silicon inside the substrate instead of on top of it, which shortens the connection between chip and board and cuts resistive loss. And glass replaces the core layer, to hold the board flat.
Warpage is the reason large-area organic substrates are hard. A big board bends as it heats, the circuits distort, and yield falls. So the glass date, which LG Innotek puts at 2028, is arguably the more important one here, because flatness is what makes area usable.
There is a quieter improvement called Cu-Post, which stands copper pillars on the substrate and puts the solder ball on top of them. LG Innotek says it cuts the spacing between solder balls to roughly a fifth of what it was, and that copper conducts heat more than seven times better than the lead it replaces. More connections, cooler package, same footprint. All three of those are what a customer asks for when the package is already at the size limit of the tool that makes it.
Is any of this a business yet?
Is any of this a business yet? The package solutions division made 99.6 billion won of operating profit in the first half of 2026, with margin up from 6.5 percent to 10.6 percent and profit nearly doubling year on year. That is real momentum, on real products.
But the company's stated goal is a trillion won of operating profit from this business by 2031. Annualise the half just reported and the target asks for roughly five times the current run rate in five years, on a substrate that no named customer has qualified.
Which is the fair objection. A sample at a trade show is not a design win, and the high end of this market belongs to incumbents. On the Taiwanese side, Unimicron's advanced substrates already sit inside TSMC's 3DFabric alliance, which is the seat everyone wants. Nobody at LG Innotek claimed an accelerator customer this week.
The same reticle ceiling is pushing the lithography end of the industry too, which we wrote about in TSMC will start using High-NA EUV in 2030. Packaging is how the industry buys time until then.
We would expect a named customer before the end of 2027, or the 2027 production date to move. Watch which of those happens first. And if you follow the packaging supply chain, watch the glass date harder than the area.
