NVIDIA books supplier capacity years in advance

At the GTC conference in Berlin, NVIDIA announced plans to expand production together with six contractors — Amkor, IBIDEN, KYEC, Samsung Electro-Mechanics, SPIL and Unimicron. Part of that capacity the company is booking up front, that is, paying for it before the new lines reach volume.

“Demand is extraordinary. Capacity is the constraint,” said NVIDIA founder and chief executive Jensen Huang. In his words, the industry is “in the midst of the largest infrastructure buildout in human history.” It is telling exactly where the money is going: not one of the six agreements concerns the production of the dies themselves. All of them cover the later stages — substrates, advanced packaging, testing and optical engines.

Exactly where production is hitting the ceiling

NVIDIA broke the expansion down into four segments, each with its own contractors:

• Substrates. The agreements with IBIDEN, Samsung Electro-Mechanics and Unimicron should together deliver a 2.5-fold capacity increase over three years. For IBIDEN’s plant in Ono, Japan, NVIDIA is providing a multibillion-dollar prepayment to book capacity, with the goal of reaching three times the current level; the additional volumes will begin ramping from the first quarter of 2028. Unimicron is expanding output of KF substrates in Taiwan: the first additional volumes are expected in the second quarter of 2027, with full capacity by mid-2028. Samsung Electro-Mechanics is joining the supply chain immediately, from its existing Korean sites, while modernizing production in parallel and building new plants in Sejong and Busan; by 2028 a notable share of its annual package substrate capacity will be allocated to NVIDIA.
• Advanced packaging. Amkor is building up capacity in South Korea and constructing new sites there. The goal is more than 300% growth in interposer wafer output by the end of 2028. The expansion is tied to current and next generations of NVIDIA platforms.
• Testing. KYEC is building new sites in the United States against a $1.1 billion prepayment from NVIDIA, with volume production planned for April 2027. Overall, NVIDIA expects to double its semiconductor testing capacity from calendar year 2026 to 2027 through construction in the US, Taiwan and Singapore; the Singapore site is due to reach volume production in February 2027.
• Optics. SPIL is expanding production in Taiwan of optical engines for co-packaged optics — optics integrated into the chip package. A 15-fold increase by mid-2027 and a further doubling by the end of 2028.

Why NVIDIA is paying before delivery

Packaging and testing contractors as a rule expand cautiously: the equipment is expensive, and utilization depends on someone else’s demand cycles. NVIDIA, judging by the terms of the deals, is taking part of that risk on itself through prepayments and capacity booking agreements — $1.1 billion for KYEC, a multibillion-dollar sum for IBIDEN, a separate agreement with Unimicron. In return the company is probably locking in volumes it would otherwise have to compete for.

The second layer of meaning is geography. The six partners operate in the US, Japan, South Korea and Taiwan, with testing additionally spread across the US, Taiwan and Singapore. NVIDIA itself describes this as strengthening the resilience and redundancy of the global supply chain.

What this means for those who work with graphics

The key thing in the announcement is the dates: almost all the new capacity reaches volume in 2027–2028. That means the capacity constraint will persist over the coming quarters, and with it, most likely, the current priority given to data center customers in shipments compared with workstations and render farms. The cost of renting accelerators in the cloud and the lead times for professional hardware, judging by these plans, will not start to ease until the new lines in Ono, Sejong and Taiwan are up and running.

The fifteenfold expansion of optical engine output points to a bet on co-packaged optics: the next round of performance gains is probably tied not only to the individual accelerator, but also to connectivity inside the rack and between racks. For studios building distributed rendering and model training on their own fleets, this is a reference point for the next few years.

The very fact that NVIDIA is financing its contractors’ expansion up front shows how narrow the stretch between a finished wafer and an assembled module has become. If the trend holds, direct investment by chip designers in someone else’s packaging and testing capacity will go from an exception to the norm, and the state of the supply chain will become a permanent factor for hardware buyers.

Community reaction

In the feeds discussing today’s NVIDIA announcement from the Berlin conference, the tone is mostly confident and interested, without outrage. Commentators from the industry and among investors agree that the company considers demand for compute extraordinary, and that the constraint now lies not in the GPU die but in substrates, packaging, testing and optics. The multibillion-dollar prepayments to suppliers, including Ibiden and KYEC, are read as booking capacity years ahead and as a sign that the internal volume forecast for 2027–2028 is above market expectations. The doubts concern the timing, not the decision itself. People note that the new lines will not appear before early 2027 and that the multiples in the announcement still have to be seen in real output, not just in a press release. Under the company’s official post there is also a more down-to-earth hope: growth in packaging and testing capacity could ease the shortage of graphics cards.

“Substrates are the new oil. You don’t prepay like this unless your internal volume forecast for 2027–2028 is well above what the market expects.”

— @bobziyangding (X)

“The names with stated timelines are packaging, testing and substrates, the unremarkable middle of the chain. The earliest date is February 2027: orders and prepayments run roughly a year ahead of the capacity appearing.”

— @demian_ai (X)

“For gamers this is the supply story behind graphics card prices and availability: the bottleneck often turns out to be packaging and testing, not the chips themselves. More capacity of that kind should mean less shortage and less panic buying when the next card launches.”

— @birajdutta_ (X)

What explains prepayment like this for lines coming in 2027–2028: fear of a substrate shortage, the race for optics, inflated AI demand?