Celestica says awarded programs already put 2028 demand above its 2027 outlook

Celestica's investor-day signal is not a guidance bump—it's a structural break in how AI infrastructure gets procured. When locked-in programs alone push 2028 demand well past 2027 revenue targets, hyperscaler capex has shifted from quarterly rolling to annual lock-in. The hardware supply chain is entering a pre-sale era, a cadence last seen during the 2019 5G buildout. Technically, Celestica sits at the intersection of AI switching (800G/1.6T) and custom ASIC board-level assembly. The 2028 step-up almost certainly maps to third-generation custom accelerator ramp (TPU v6 / MTIA v3 class), while cluster scaling from 10K to 100K+ units makes the networking BOM share overtake compute for the first time. CoWoS packaging, HBM4, and liquid-cooling bottlenecks will surface in H2 2026, and whoever controls allocation then controls the next order cycle. Competitively, Foxconn is squeezing share with a full-stack server-plus-networking play, while Jabil locks into specific hyperscaler programs. Celestica's moat is networking depth—hard to replicate for pure assemblers—but customer concentration remains a live, un-hedged risk. The real 12-24-month tail effect isn't in Celestica's P&L. It's in the supply chain's capex cycle: when EMS players plan capacity on a 2-3-year horizon, upstream equipment and materials visibility stretches beyond 18 months, rewriting the valuation anchor for back-end semiconductor manufacturing entirely.