NVIDIA Prepaid $1.1B for Test Capacity: Buyer Readout

NVIDIA closed 22 October at USD 212.21, down 4.4 percent, on turnover of USD 26.5 billion and a market capitalisation of USD 5.11 trillion, after a session that took the Philadelphia Semiconductor Index down more than 5 percent and dragged AMD, Intel, Micron, TSMC and ASML with it. By midday on 23 October the stock had clawed part of that back and was trading neck and neck with Apple for the title of most valuable listed company, roughly USD 5.10 trillion against Apple at USD 5.12 trillion. The selloff was about rates, oil above USD 100 and doubts over AI capital spending, not about anything NVIDIA shipped.

The capacity story that matters to buyers came from Goldman Sachs on 23 October. Its team summarised Jensen Huang's GTC Berlin keynote into three points: NVIDIA is expanding several links of its supply chain at once, which implies materially larger GPU shipments ahead; the DSX POD reference design compresses data centre deployment from months to about two weeks; and physical AI revenue already reached USD 11 billion in the first half of fiscal 2027, more than 30 percent above the whole of fiscal 2026. Goldman kept a Buy rating and a 12 month target of USD 300, with fiscal 2027 revenue of USD 412.7 billion and fiscal 2028 revenue of USD 718.3 billion.

The most concrete procurement signal of the week was reported on 21 October: King Yuan Electronics, the Taiwanese test house known as KYEC, will build a new production facility in the United States, and NVIDIA is supporting it with about USD 1.1 billion of capacity reservation prepayments. That is the same instrument NVIDIA and Tesla used with Samsung Electro-Mechanics for FC-BGA substrate lines. Prepayment for capacity rather than for parts is becoming standard practice at the top of the AI stack.

For component buyers the readout is threefold. First, test and packaging capacity is being reserved years out, so merchant capacity at outsourced assembly and test houses will stay tight for parts that share the same lines. Second, standardised rack designs such as DSX POD and the 800 VDC power architecture pull power shelves, busbars, cold plates and high speed connectors into repeatable bills of material, which shortens qualification but concentrates demand on a small approved vendor list. Third, the layer counts and panel sizes behind AI accelerators keep high layer PCB and substrate lead times long, so any board that shares that capacity should be quoted with optionality, not a single source.

Practically, buyers should ask three questions of every AI adjacent BOM this quarter. Which of my parts share test, substrate or HBM capacity with accelerator production. Which of my designs can move to a standardised rack power architecture without requalification. And what is my fallback if a distributor allocation is cut in the first weeks of a quarter. Our component catalog and quote desk covers power, interface and support ICs across those designs, and our PCBA assembly service team can quote alternate board builds where a substrate dependent part is the bottleneck. Capacity letters signed now are worth more than price concessions later.