The Pre-Print Signal — The Week Delta Electronics and Yageo Both Hit Limit-Down Before Their Own Earnings Calls
The layer adjacent to AI compute — power, passives, auto-memory — was repriced -10% before it reported, while back-end packaging printed record quarterly margins.
The Pre-Print Limit-Down
The single most unusual ticker in Taiwan on July 28 was not TSMC. It was Delta Electronics (2308) and Yageo (2327) — both tripping the 10% limit-down circuit breaker on the same day, both immediately ahead of their own scheduled investor calls. This was not a sell-off in response to a miss. This was a sell-off in advance of the print. It shows how thinly the AI supply-chain volume narrative was actually held in price.
Delta is the rack-scale liquid cooling and 800V power leader — the single biggest Taiwanese beneficiary of the Vera Rubin platform expanding cooling content from a GPU-scale story to a rack-scale story. Yageo is the world's #1 MLCC supplier and a critical component vendor into AI server power stages. That both names hit -10% one day before their earnings calls tells you that the tape was not willing to wait for the datapoint labelled "AI capex is still strong." It reached a pre-emptive conclusion that "the volume around AI has already peaked" — and it acted before the companies got to argue otherwise.
The 3rd-Largest Point Drop — With Half the Top 50 Hitting Limit-Down
The TAIEX closed the same session down 2,031 points (-4.65%) at 41,603 — the third-largest single-day point drop on record. TSMC (2330) broke below NT$2,300 in the afternoon session, and more than half of Taiwan's top-50 stocks hit the 10% limit-down. By the following morning (July 29), the TAIEX opened down another -4.93%, with three concept baskets — automotive memory, silicon IP, and III-V compound semis — collapsing near -9%.
The identity of the wrecked sectors matters. This sell-off was not led by TSMC or MediaTek — the actual AI compute names. It was led by (1) automotive memory, (2) silicon IP, and (3) compound semis. In other words, the market surgically targeted volume that AI data centers will not absorb. Winbond and Nanya also hit limit-down again, but that continues last week's CXMT-IPO storyline; the new information this week is that the crosshairs have expanded down one more layer, into the components adjacent to AI compute rather than the compute itself.
The Back End Printed Records
During the same three days, on the other side of the tape from the sell-off, here is what actually happened.
- Macronix (2337): Q2 net profit NT$7.7B (+706% YoY, +335% QoQ), gross margin 64.4% (+23.6ppt QoQ), board authorised NT$15.8B extra capex. Specialty memory is on a completely different trajectory from CXMT's commodity DRAM.
- Powertech (6239): Q2 net profit NT$2.22B (+131% YoY, +20% QoQ), EPS NT$3.00 — a 2.5-year high. FOPLP (fan-out panel-level packaging) yields rising, mass production imminent.
- Chipbond (3264): Q2 EPS NT$2.44, beating consensus by 22.3%, with AI ASIC test capacity entering mass production.
- Win Semiconductors (3105): Q2 revenue NT$5.26B (+39% YoY), Q3 guide +10% revenue and 30%+ gross margin. LEO satellite and AI data-center demand layered on top of the smartphone seasonal.
- Adlink (6166): Q2 revenue NT$4.46B — all-time quarterly high, +52% YoY / +29% QoQ, driven by edge-AI and semiconductor equipment.
- Sigurd (6257): Approved an additional NT$2B capex through its testing subsidiary specifically to serve a major US AI chip client.
The common thread across those six companies is exactly one thing: they are either back-end firms serving actual GPU/ASIC volume (packaging, test), or niche players exposed to a specific data-center content item (edge-AI servers, LEO satellite RF, specialty NOR). The split is legible. The layer the market sold this week — "AI-adjacent" volume — is not the layer the market is paying for.
The Delta Paradox
The story that Vera Rubin expands liquid cooling from GPU-level to full-rack-level got reinforced this week. That reads as a phase transition where AI-server content-per-server growth outruns unit growth, and Delta should be the most direct Taiwanese beneficiary. And yet Delta hit -10% the day before its own earnings call.
There are two ways to read the mismatch. (A) The market had already fully priced content growth, and was ready to sell on any confirmation — this -10% is overreaction. (B) The market already knows Delta has not yet printed the content transition in this quarter, and got out before management could confirm — this -10% is precise pre-emption. Which interpretation is right will be determined by Delta's guidance next week. That is the single most-load-bearing print in Taiwan for the next fortnight.
PM Positioning
The signal from the Taiwan tape this week is not "AI is over." It is "the AI supply chain is being reclassified." The market is drawing a wall between the volume-driven adjacent layer (power, passives, auto memory, silicon IP) and the content-driven core layer (leading-edge foundry, AI back-end packaging and test, niche specialty). The Delta limit-down is the first stress test of that wall. The Macronix / Powertech / Chipbond prints show the far side of the wall is still standing.
Positioning: Taiwan back-end packaging (Powertech, Chipbond, Sigurd), specialty memory (Macronix), and compound-semi niches (Win Semi) have kept their earnings defensibility through this week's rout. The pre-emptive sell-off in power and passives, meanwhile, needs earnings-call confirmation before it can be leaned into either direction. Watch item: the Delta and Yageo guidance calls this week decide whether the split is real or gets unwound in three sessions.
Key Sources: - Yageo and Delta Electronics Hit Limit-Down Ahead of Earnings Calls; AI Demand in Focus (cnYES, 2026-07-28) - Taiwan Stocks Post 3rd-Largest Point Drop Ever on Rate-Hike Fears and AI Doubts (cnYES, 2026-07-28) - Macronix Q2 Net Profit Soars 7x YoY to NT$7.7B; Board Approves NT$15.8B Extra Capex (cnYES, 2026-07-28) - Chipbond Q2 EPS Beats by 22% as AI ASIC Test Capacity Enters Mass Production (cnYES, 2026-07-27) - Vera Rubin Expands Liquid Cooling from GPU to Full-Rack AI Systems, Changing the Growth Calculus (cnYES, 2026-07-28) - plus 55 more
If this analysis was helpful · ☕ Support Us · ✈️ Telegram