On 21 October 2026, at GTC Berlin, NVIDIA named six partners — Amkor Technology, IBIDEN, King Yuan Electronics (KYEC), Samsung Electro-Mechanics, Siliconware Precision Industries (SPIL) and Unimicron — and attached dates to capacity that does not exist yet. Jensen Huang's framing was blunt: "We are in the midst of the largest infrastructure buildout in human history. Demand is extraordinary. Capacity is the constraint." The release that followed is best read as a schedule: who is adding what, where, and when it is supposed to switch on.
Substrates: 2.5x is a sum, not a plant
NVIDIA says it plans to increase substrate capacity for its products 2.5x over the next three years. That figure is an aggregate: the agreements with IBIDEN, Samsung Electro-Mechanics and Unimicron are meant to contribute to it, and NVIDIA says more agreements are underway. No absolute baseline — panels, wafers, square metres — is given, so the multiple is the only unit on offer.
The three contributions look different. IBIDEN's deal was signed and announced earlier in 2026, with NVIDIA providing a multibillion-dollar capacity reservation advance for the Ono Plant in Japan; that expansion targets 3x the plant's current capacity, with additional capacity expected to start ramping in the first quarter of 2028. Samsung Electro-Mechanics is new to the substrate supply chain and brings immediate capacity from existing South Korean operations, while the collaboration also supports upgrades and new factories in Sejong and Busan; by 2028, production allocated to NVIDIA is projected to be a significant portion of its total annual package-substrate capacity. Unimicron's KF substrate expansion in Taiwan, backed by a capacity reservation agreement, is expected to begin adding capacity in the second quarter of 2027 and reach full capacity by mid-2028.
A "capacity reservation advance" is worth translating: the customer pays ahead to hold a place in the queue. It moves some of the expansion risk off the supplier's balance sheet and onto the buyer's, which tells you how badly the buyer wants the line built.
Packaging and testing
Amkor and NVIDIA plan to expand advanced packaging in South Korea, targeting a greater than 300% increase in interposer wafer capacity by the end of 2028, alongside Amkor's development of new manufacturing facilities there. Interposer wafers are one specific slice of packaging, not packaging in general.
On testing, NVIDIA expects its total semiconductor testing capacity to double from calendar 2026 to 2027, supported by buildouts in the U.S., Taiwan and Singapore. The $1.1 billion capacity reservation advance to KYEC is earmarked for new U.S. testing facilities and equipment, with high-volume production targeted for April 2027; expanded testing in Singapore is expected to begin high-volume production in February 2027. The KYEC advance is one of only two dollar amounts in the announcement.
Optical engines
NVIDIA and SPIL plan to expand capacity in Taiwan for optical engines used in co-packaged optics, targeting a 15x increase by mid-2027 and another doubling by the end of 2028. Fifteen times in roughly eight months is the steepest ratio in the release.
The supplier's side of the same base
On a 29 October call, optical-module maker Eoptolink said raw-material supply had been tight overall this year and was not limited to one material, with stocking lead times for core materials further lengthened as industry delivery volumes grew — meaning buyers need to plan further ahead. The company said it had prepared for fourth-quarter and next-year orders and was confident of meeting much higher delivery demand; it expected fourth-quarter capacity and output to rise substantially, was expanding rapidly at its Chengdu and Thailand plants, and was optimistic about next year. It gave no capacity, volume or revenue figures.
Read the two accounts together and the shape of the problem is visible. The plan adds optical-engine capacity downstream at a multiple, while a module maker on the same chain says inputs are hard to book and lead times are stretching. Expansion does not remove a constraint so much as relocate it.
What the calendar doesn't say
NVIDIA labels the capacity, timing and production expectations as forward-looking statements subject to risks and uncertainties. These are targets and plans, not reported output, and the release does not state capital amounts for most of the agreements beyond the KYEC $1.1 billion and IBIDEN's multibillion-dollar advance. Who funds the Samsung Electro-Mechanics and Unimicron builds, and on what terms, is not disclosed.
Dates to check
February 2027 (Singapore testing), April 2027 (KYEC U.S. high-volume production), second quarter 2027 (Unimicron's first added capacity), mid-2027 (SPIL's 15x), first quarter 2028 (IBIDEN ramp), mid-2028 (Unimicron full capacity), end-2028 (Amkor's 300%-plus and SPIL's second doubling). A calendar is not capacity; its value is that it can be checked against output, quarter by quarter.