Wiwynn's Texas Plant Is the AI Build-Out's Capex, Not Its Margin

Wiwynn, the cloud-AI server unit of Wistron, held the grand opening of its first U.S. assembly facility in Socorro, Texas on September 14, with Amazon as its anchor customer. On the surface this reads as another capacity win in the AI infrastructure boom: more than $1.6 billion of committed investment, roughly 1,000 new jobs moving toward 2,500 next year, machines turning out server racks for Amazon's data centers. It is a good factory. But read the opening as evidence of who is capturing the economics of the build-out, and it points the opposite way.

The plant exists because of a tariff, not a demand surprise. The force moving this work to Texas is less the pace of AI demand — real and accelerating as it is — than the geography of trade. The United States has hit Taiwan's computer and server makers with a 32 percent reciprocal tariff, and the category they make accounts for 58 percent of Taiwan's exports to the U.S. Semiconductors were carved out of that levy; server assembly was not. So the ODM industry has been forced to relocate final assembly to the customer's doorstep. Wiwynn's output has historically sat roughly 70 percent in Mexico, and the company now targets about 20 percent of total output from U.S. soil next year. Texas is not where booming demand pulled the work. It is where the tariff pushed it.

The split that matters is technical, not geographic. Wiwynn's growth is shifting away from general-purpose servers built around Nvidia GPUs toward ASIC servers built around hyperscalers' custom silicon — Amazon's Trainium line, which Amazon designs and owns. Wiwynn expects a "golden cross" in the third quarter, with ASIC and general-purpose each about half of shipments, and ASIC becoming the primary revenue driver in the fourth. That is the defining split of this cycle. A general-purpose server carries Nvidia silicon whose economics belong, in part, to Nvidia. An ASIC server is a customer-specified box whose chip, design, and pricing all sit upstream, with the hyperscaler that commissions it.

The numbers that matter are capex and margin, not the growth rate. Here Wiwynn looks less like a beneficiary than a conduit for capital. Second-quarter revenue rose 26 percent year over year and profit 23.5 percent, on earnings per share of NT$80.43 — yet gross margin was just 9.3 percent. Wiwynn's own investor materials add that the margin improvement partly reflects a shift to a procurement agency model, a change in how revenue and cost are recorded rather than a gain in pricing power. Against that reporting, the board has approved a second-half capital budget of $942 million, more than double first-half spending, on top of the more than $1.6 billion already sunk into Texas. An assembler funding that much plant at a 5.4 percent net margin is not capturing the value of what it builds. It is paying to stay inside the supply chain of a single customer — Amazon — whose AWS is growing 37 percent and whose vertically integrated, custom-silicon business is precisely what the Texas line exists to serve.

The implication is straightforward for anyone reading AI-infrastructure headlines as a rising tide. Capacity expansion across the assemblers is real, capex-heavy, and increasingly locked to wherever the tariff and the customer sit. But the value in this chain is accreting to the silicon owners and the hyperscalers who dictate the design, not to the factories bolting the boxes together. The signal to track is not the square footage Wiwynn opens or the reported growth it books. It is whether the rising ASIC share of the mix is priced at margins the company can defend, against the capital it has committed — whether the assembler is building its customers' wealth, or its own.