The Allied Stack Closed Ranks — The Week the Non-China Semi Bloc Re-Wired Capex To Each Other

The week the wall finished being built. On Sunday, Mainichi and FT reports converged on the same picture: Japan, the US and the EU formally coordinated a tighter joint blockade on semiconductor manufacturing equipment exports to China — and four days later, Nvidia's China data-center GPU share was reported at effectively zero. The point isn't that China is suddenly cut off; the point is that the allied semiconductor bloc — Korea, Taiwan, Japan, the US — spent this week behaving as if China simply isn't in the addressable market anymore, and started re-routing capex, partnerships, and material flows toward each other instead of toward the customer they used to share.

That re-routing showed up in the numbers first. Korea's May semiconductor exports printed $29.4B, up 154.29% YoY — a fourth straight month above $24B, on a trajectory now visibly accelerating into HBM4E. Taiwan's Cathay-NTU economists raised their 2026 GDP forecast to 10.1% from 5.8%, attributing the entire upgrade to AI capex revisions. Nomura's Taiwan team kept CoWoS at the core of its 2026 EPS-up-50% call. Japan's Nikkei briefly touched 70,000 on Tuesday — BOJ hike to 1.0% notwithstanding — led by Advantest and Kioxia. And the US side of the bloc kept signing the cash checks: Sharon AI inked 72MW with Nvidia for six years, Rackspace committed 30MW of AMD Instinct, DataBank closed $1.45B for Red Oak in Texas. The interesting datapoint is what's missing from that list: not a single one of these commitments is China-bound.

Where it gets specific is the equipment stack. Korea's back-end vendors all closed orders this week that ratify the bloc's HBM4/HBM4E build-out as a coordinated, not opportunistic, ramp. Techwing (089030) won its first cube prober supply contract from SK Hynix after passing March's qual test, joining its existing Samsung book. ISTE landed an initial HBM-dedicated FOUP cleaner order from Samsung. Yes-T's tools became essential for HBM4/HBM5 lots. LG Innotek pushed its FC-BGA business from PC CPUs deeper into AI servers — directly into TSMC's substrate-anxious customers. Notice the pattern: Korean equipment, Korean memory, but the demand pull is Taiwan-side AI accelerators. The bloc is internally cross-wired now.

The Taiwan side of the cross-wiring is even tighter, and more strained. WinWay Tech (6223) chairman Ko Chang-lin told shareholders capacity is "never enough" and the company is now negotiating prepayments from AI customers to fund expansion. Gallant Precision (3131) said current wet-process equipment capacity meets only one-third to one-quarter of orders through 2030, with lead times stretched past twelve months. Advantech's (2395) Edge AI book shows 43% of early-2026 orders carrying lead times over six months. TSMC's (2330) CoPoS roadmap moved into its 2026 validation phase with 310×310mm substrates, positioning Taiwan panel makers as the next layer of the allied substrate stack. And on the US side, TSMC and Amkor formalized a deeper packaging-and-test partnership the day before — quietly making Amkor the bloc's allied-side backup to TSMC's own packaging crunch, with the China-touching alternative (CXMT, SMIC) explicitly outside the line.