GlobalFoundries ties its 1.1 billion euro Dresden expansion to a 7nm-class FD-SOI process for edge AI

Spread across a million wafers of yearly capacity, the roughly 1.1 billion euros set aside for the SPRINT project ^[9] comes to about 1,100 euros a wafer ^[13]. The spend pays for the wafers the expansion adds, or rather for whatever part of the million is new. SPRINT enlarges a fab that already runs ^[9], so the added wafers are probably fewer than a million and each one costs more than 1,100 euros ^[13]. The report does not say what Dresden makes today, how much of the added capacity goes to the 7nm-class process, or whether any of it is committed to European customers.

Design kits for the 40UX and 22UX variants are due in September 2026 ^[8], and production is set for 2028 ^[2]. Depending on the month, designers get 16 to 27 months between the two ^[14]. Crypto Briefing calls it roughly two years ^[11]. The process this money is attached to is planar FD-SOI. It sits on an insulating layer to cut power leakage, and it differs from the FinFET processes the report associates with TSMC and Samsung ^[4].

There are two plausible ways the capacity gets used. In the first, most of the new capacity goes to 40UX and 22UX work, and the 7nm-class node takes a small first slice in 2028 ^[8][2]. In the second, GlobalFoundries gives FDX Fusion a large share of the million wafers, and the expansion becomes a bet on that node. Crypto Briefing's description of the two announcements as strategically linked points that way ^[12].

I think the first is likelier. The capital buys capacity for a process family that has been in volume production since 2019 and has shipped more than 3 billion chips ^[6], and the report puts the 40UX and 22UX variants on the same roadmap ^[8]. The case against that view is the company's own framing: it timed the node's 2028 start to the expansion ^[2] and put a serviceable market estimate of more than $18 billion by 2030 on physical AI ^[3]. A published allocation giving FDX Fusion most of the added capacity would prove the first reading wrong.