NVIDIA Advances KYEC $1.1bn in Supply Chain Push

“Demand is extraordinary. Capacity is the constraint.” Jensen Huang, who called the moment “the largest infrastructure buildout in human history”, used GTC Berlin on 21 October to set out which factories NVIDIA is paying to have built.

In an announcement at 11:33 BST, NVIDIA said six suppliers would expand capacity in the less visible stages of making an AI chip: the substrates the silicon sits on, the packaging that joins chips together, the testing that checks them, and the optical engines that will carry data between them. The figure that stands out is a $1.1bn capacity reservation advance to King Yuan Electronics (KYEC) for new testing plants in the US. When KYEC disclosed the deal in a Taiwan filing last week, it did not say how much NVIDIA was paying. Every capacity number in the release is a target, and NVIDIA’s own.

The money buys factory space in advance

The common mechanism is the capacity reservation. NVIDIA pays a supplier up front, the supplier builds or upgrades a plant, and the output is set aside for NVIDIA. KYEC’s filing described this structure, with the prepayment offset against future service fees and the agreement running until it is used up. NVIDIA says the arrangements are meant to strengthen “the resiliency and redundancy” of its supply chain, by enabling both new and existing suppliers to add capability.

Substrates are the largest part of the plan. NVIDIA says agreements with IBIDEN, Samsung Electro-Mechanics and Unimicron will together increase substrate capacity 2.5 times over the next three years, “with more underway”. Not all of it is new. The IBIDEN agreement, a “multibillion-dollar” advance for the company’s Ono plant in Japan, was signed and announced earlier this year; the release restates it, with a target of three times the plant’s current capacity from the first quarter of 2028.

Samsung Electro-Mechanics is the newcomer. It will supply from existing South Korean operations at once, with upgrades and new factories in Sejong and Busan to follow. NVIDIA projects that by 2028 its allocation will be “a significant portion” of the company’s total package substrate capacity, without putting a number on it.

Partner (work, location)	Capacity target	Timing
IBIDEN (substrates, Japan)	3x at Ono plant	Ramp from Q1 2028
Unimicron (substrates, Taiwan)	Not given	From Q2 2027; full by mid-2028
Amkor (interposer wafers, South Korea)	More than 300% increase	By end of 2028
KYEC (testing, US)	$1.1bn advance; no capacity figure	High volume from April 2027
SPIL (optical engines, Taiwan)	15x, then a further doubling	Mid-2027; end of 2028

Source: NVIDIA, 21 October 2026. All figures are NVIDIA targets.

Testing doubles, and optics grows fastest

Amkor’s part is packaging. NVIDIA and Amkor plan to more than triple interposer wafer capacity in South Korea by the end of 2028. NVIDIA says the expansion, alongside new Amkor facilities in the country, will support higher production volumes for its current and next-generation platforms.

On testing, NVIDIA expects total capacity to double between 2026 and 2027, with buildouts in the US, Taiwan and Singapore. KYEC’s US plants are due to reach high-volume production in April 2027, and expanded operations in Singapore in February 2027.

The steepest ramp is in optics. NVIDIA and Siliconware Precision Industries (SPIL) plan to increase capacity in Taiwan for optical engines used in co-packaged optics fifteen-fold by mid-2027, and to aim for another doubling by the end of 2028. NVIDIA links the expansion to the growing connectivity needs of large-scale computing infrastructure, where co-packaged optics is meant to carry data between chips and racks as light rather than over copper.

The plan touches Japan, South Korea, Taiwan, Singapore and the US, but not Europe, despite being unveiled in Berlin. The first test of whether the targets hold comes early next year, when the Singapore testing lines are due to reach volume in February, two months before KYEC’s US plants.