Everyone assumes CoWoS spilling over to ASE and Amkor just means the OSATs add… · @Leoskie_L · TwKit

@Leoskie_L

Everyone assumes CoWoS spilling over to ASE and Amkor just means the OSATs add a few more lines. But to actually catch that overflow, they first need machines big enough to bond an entire AI chip package in one go. That's what today's Korea Economic TV exclusive is about. Per the report, Hanmi Semiconductor will launch its "TC Bonder 120" for 2.5D packaging at year-end. Amkor and ASE are running quality tests, samples come at year-end, and volume shipments start in 1Q27. This is a Korean media exclusive. Hanmi, ASE and Amkor haven't announced anything, and a quality test is not an order. A TC (thermo-compression) bonder uses heat and pressure to attach chips. In 2.5D, it places several dies, GPU, CPU and HBM, onto one large interposer. Hanmi's existing 2.5D tool is 40mm class. The new one handles 120mm x 120mm. If the old one is square too, that's roughly 9x the area. That's my math. Why so big? Because interposers keep growing. On the Oct 15 call, C.C. Wei said TSMC is already in volume production on 5x-reticle advanced packaging, with 9x in development. One reticle is about 26x33mm, or 858mm². 5x is about 4,290mm², 9x about 7,700mm². 120mm square is 14,400mm². A tool's bonding range and an interposer's area aren't the same kind of size, but the order of magnitude tells you Hanmi is preparing for packages beyond 9x. That's my inference. This ties back to my SemiAnalysis post. Their March piece said that when CoWoS really runs short, the overflow options are ASE/SPIL, Amkor and Intel's EMIB. But spillover isn't just handing over the orders. The other side needs equally large tools and equal yields. Today's story is one piece of equipment on that spillover path. How big the spillover gets is disputed. SemiAnalysis's March view was that CoWoS is "tight but easing" and the real constraint is front-end wafers. Commercial Times (Oct 5) cited market estimates that even with TSMC's expansion, CoWoS will still be 15-20% short, and a local advisory firm estimated ASE's FoCoS monthly capacity at about 20k wafers by end-2026 and 40-60k by end-2027. I can't verify the advisory numbers with the company, so treat them as estimates. Think of a print shop. Customers used to print A4. Now everyone wants full-size posters. Adding more A4 copiers doesn't help. You need a large-format printer, and those aren't available on demand. Hanmi isn't first to this business, though. On Jul 29, ASMPT announced wins for more than 50 chip-to-substrate TCB tools from OSAT customers for AI chips, calling itself the main supplier and process tool of record for the application. Besi's CEO said on the July call that "CoWoS-like" orders started coming in from 4Q25 and are "already a significant part of the orders." So in 2.5D, Hanmi is the challenger. Its strength is HBM: the report says Hanmi's HBM TC bonders supply Micron and SK hynix with about 70% share. That's the report's number. Hanmi has been busy. Per the report, on Sep 22 it shipped its "2.5D FC Bonder 75" for large interposers to a global customer, on Oct 16 it set up "Hanmi USA" in San Jose, and at year-end it will show a sample of its second-gen hybrid bonder, timed for HBM4 16-high and above. On the Taiwan side, there are other tools on the same line. GPTC (Grand Process Technology) makes CoWoS wet-clean equipment. A Goldman report puts its share at about 50% at TSMC CoWoS and close to 100% at ASE and SPIL (as relayed by a local advisory report). So if ASE adds CoWoS-like lines, GPTC's cleaning tools probably follow (inference). All Ring Tech does back-end underfill dispensing and lid attach, and analysts say its order visibility runs to 2Q27. But the equipment makers are stuck too. GPTC's chairman said at SEMICON West on Oct 15 that the bottleneck now is "people and materials." EFEM lead times were 8 months two months ago and have stretched to 10, and some parts take up to two years. So my read: if Hanmi's timeline holds, 1Q27 shipments line up with the ASE 2027 expansion the advisory firm expects. But there are a lot of ifs. The quality tests have to pass, ASMPT has to leave room, and if SemiAnalysis is right that front-end wafers are the ceiling, extra back-end tools just wait for chips. That's also inference. I started out thinking packaging was short on TSMC capacity. By the end, I found even the packaging tools are queuing for parts. $AMKR $ASX #Hanmi 042700.KS #ASE http://3711.TW #GPTC 3131.TWO #AllRingTech 6187.TWO #ASMPT http://0522.HK #CoWoS #AdvancedPackaging Not investment advice.

曝光 258 · 评论 0 · 点赞 0 · 书签 0