
Custom Chips: Don't Value Them Like Platform Companies
Chiplet has been talked up as a platform story for the past two years. Break a big chip into small dies, and customers piece together as much compute as they want — sounds like building blocks. North Pole Xiongxin's customized high-performance computing plus AI acceleration solution follows the same narrative.
I went through their technical roadmap and public materials. My biggest takeaway is one sentence: what this business really sells is interconnect; the chip itself comes second.
There's a detail in the material I read that's telling. When Intel's Meteor Lake first went Chiplet, what the industry talked about most was how the dies connect to each other — nobody much cared how many dies were cut. AMD has done Chiplet for years, and apart from 3D V-Cache, its die-to-die interconnect has always been described as "primitive." Cutting dies isn't hard; the hard part is making them talk to each other like a single chip after they're cut.
That's where pricing power lives. Whoever can hold down interconnect IP, advanced packaging, and yield all at once has the bargaining power. Just doing die design and assembly essentially saves the customer tape-out money — you're earning engineering service fees, and the ceiling is visible.
So my judgment is direct. This kind of company shouldn't be valued as a platform semiconductor company; it should be valued as an engineering services provider with reusable IP modules in hand. Looking at order count means little; what matters is whether it has turned interconnect and packaging solutions into modules that can be reused, so the next customer requires less physical design work from scratch. If it can't do that, every project starts over, and both margin and valuation get compressed.
One-line advice: if you're also watching this track, make reusable interconnect IP your first screening criterion — it's more useful than looking at the customer list.
At this stage my stance is tracking, not building a position. The track's ceiling is genuinely high, but the competitive landscape hasn't converged, and going in too early means paying tuition on someone else's behalf.
Physix Frontier