It’s not a simple cyclical bounce, but a layering of vectors rarely seen before in networking. Taiwanese manufacturers of networking gear — switches, optical modules, access points, ground-station antennas — are getting a simultaneous boost from three directions: the AI race, the rise of low-earth-orbit (LEO) constellations, and public broadband upgrade plans. The quarter’s tailwind is triple-barreled, and that changes the nature of the game.

The AI engine

The loudest driver comes from data centers that train and serve LLMs. Large-scale GPU clusters no longer live on a few isolated racks; they are distributed fabrics that demand switches with extremely high capacity, deterministic latency, and near-zero packet loss. Taiwanese builders, from Accton to Quanta Cloud, have become the natural partners of hyperscalers because they can produce customized platforms — often based on Broadcom or Marvell silicon — in manageable volumes, with lead times that traditional vendors struggle to match. The shift to 400G and 800G links, coupled with protocols like RoCEv2 for efficient gradient transport, is inflating orders for on-switch optics and pluggable transceivers. It’s not just about volumes: the very structure of AI networking rewards those who can quickly integrate a chip vendor’s reference design with their own manufacturing muscle.

The LEO effect and the broadband push

Less discussed but just as structural is the contribution of mega-LEO constellations. Each satellite needs user terminals and ground gateways, which often resemble software-defined small radio stations. Here Taiwan’s manufacturing flexibility — able to pivot from a batch of telecom switches to phased-array antennas with minimal adjustments — represents a clear competitive advantage. In parallel, national fiber-extension plans (across Europe, North America, and Asia) are pulling demand for ONTs, OLTs, and residential CPE, a segment where Taiwanese makers have operated for decades with solid margins, but above all with volumes that help saturate production lines and lower unit costs even for the more sophisticated products bound for AI.

Who wins and who loses (and what it means for on-prem)

The direct beneficiary is the supply chain concentrated in Taiwan: not just the ODMs, but the entire ecosystem of packaging, testing, and logistics revolving around semiconductors. The losers are those whose business model relies on selling branded gear with high margins, because the commoditization of white-box switches — accelerated precisely by AI — erodes positional rents. Unsurprisingly, some large legacy vendors are re-architecting to embed more software and services, while Taiwanese ODMs gain ground directly with hyperscalers and, increasingly, with large enterprises that set up on-premise clusters.

For Italian infrastructure teams evaluating on-premise LLM deployments, this scenario has very concrete implications. A 100G/400G fabric built on open Taiwan-based switches can significantly reduce TCO compared to branded solutions, but demands integration skills, careful firmware management, and second-level support savvy. The choice is not neutral, and the speed with which Taiwanese manufacturers are responding to AI demand suggests that the window to build competitive on-premise stacks is widening, not closing. It’s not a guarantee, but a sign of hardware supply maturity that can make data sovereignty more tangible without yesterday’s prohibitive cost.