TL;DR · 30-second read
The Short Version
Nvidia, the company whose chips power most of today’s artificial intelligence, has signed a new $1.3 billion deal with Amkor Technology. Amkor does the final assembly of chips, sealing them into a protective casing and wiring them to their memory, in South Korea and Arizona.
With an earlier deal, the two companies’ commitments now total $3 billion. A finished chip is useless until that last step is done, and Nvidia is securing the assembly line years in advance. That tells you this little-known step can hold up the supply of artificial intelligence chips.
Even so, Amkor’s shares have fallen nearly 40% in three months.
Amkor Technology, one of the largest independent providers of semiconductor packaging and test services, has signed a new US$1.3 billion capacity expansion agreement with Nvidia covering operations in South Korea, Simply Wall St reported on October 23, 2026. The agreement lifts the two companies’ combined multi-year commitments across Amkor’s Arizona and Korea sites to US$3 billion.
The deal arrives during a steep slide in Amkor’s stock: shares were down 28.5% over the past month and 39.9% over 90 days, closing at $39.05. Over longer horizons the picture is stronger, with total shareholder returns of 23.0% over one year and 92.8% over three years.
Executive Summary
Nvidia has expanded its multi-year arrangements with Amkor Technology to a combined US$3 billion across two countries, with a new US$1.3 billion agreement focused on South Korea. Amkor is not a chipmaker in the wafer-fabrication sense. It handles the back end of the process, where finished silicon is cut, mounted, connected to memory and other dies, sealed and tested. For AI accelerators, that step has become far more complex and capital-intensive than it was for conventional chips.
The significance is in who is signing and at what scale. A capacity expansion agreement is a forward-looking arrangement: it ties future production lines to a specific customer’s needs. When the world’s dominant AI accelerator designer commits on this order to an independent packaging house, in both Asia and the United States, it signals that back-end capacity is something to be secured ahead of demand rather than bought as needed.
The market has not rewarded the news so far. Amkor’s shares have fallen sharply over the past three months, and the debate around the company has shifted to execution: whether new facilities fill quickly enough, and whether weakness in its smartphone-oriented communications business offsets AI gains.
Why a Packaging Deal Is an AI Supply Decision
Modern AI accelerators are not single chips. A typical high-end design places one or more processor dies beside stacks of high-bandwidth memory on a shared base layer, an approach known as 2.5D packaging, or spreads connections across a dense redistribution layer in what is called high-density fan-out (HDFO). Amkor has described growing engagement in both techniques for AI data center and high-performance computing customers. Until that assembly is done, a wafer of finished processors cannot ship as a usable product, so packaging capacity sets a ceiling on how many accelerators reach the market regardless of how many wafers the foundries produce.
That is the mechanism behind the $3 billion figure. Customers sign multi-year capacity expansion agreements, rather than placing ordinary purchase orders, when they expect an input to be scarce or slow to build. Advanced packaging lines need specialised equipment, cleanroom space and qualification time, so capacity cannot be added on short notice. A combined US$3 billion commitment spanning two countries indicates Nvidia is planning its back-end supply on a multi-year horizon, the same way the industry has long planned wafer capacity.
The arrangement also extends Nvidia’s packaging base beyond the foundry. Its leading accelerators have been widely reported to rely on TSMC’s in-house advanced packaging. Deepening ties with an independent assembler gives Nvidia a second route for at least part of that work. The terms have not been disclosed, so the share of volume involved is unknown, but the direction is clear. The parties most affected are other accelerator designers and cloud providers competing for the same advanced lines, who may find more of that capacity already spoken for.
Two Sites, Two Different Jobs
The geography matters. South Korea is home to a dense memory supply base, and the new US$1.3 billion agreement adds capacity there. The Arizona commitment places advanced packaging in the same state where TSMC is building US wafer fabs, which could allow more of the accelerator supply chain, from wafer to packaged chip, to be completed on US soil.
For a buyer of AI infrastructure, that dual footprint is a form of risk management. Concentrating back-end assembly in a single country or a single supplier exposes accelerator output to local disruptions, export rules or capacity crunches. Spreading committed capacity across Asia and North America does not remove those risks, but it reduces the odds that one constraint halts supply.
The Utilization Question Behind the Share Price
Amkor’s stock fell 39.9% over 90 days even as these commitments grew, which suggests investors are weighing execution risk as heavily as the headline value. Two pressure points stand out in the current debate: prolonged softness in Amkor’s communications segment, which is tied largely to smartphone chips, and the possibility that new facilities run below capacity if AI-related packaging demand disappoints.
The second risk is the important one for infrastructure watchers. Packaging plants are expensive to build and costly to leave idle, so a capacity agreement’s value to Amkor depends on its structure. It matters whether it includes prepayments, minimum volumes or cost-sharing, or simply signals intent. Investor sentiment is split. One widely followed community valuation pegs fair value at $74.50 against the $39.05 close, built on AI packaging growth and a richer product mix. That estimate depends on the same demand assumptions the sell-off is questioning.
Background
Semiconductor manufacturing splits into a front end, where circuits are etched onto silicon wafers at foundries, and a back end, where those wafers are cut into individual dies, packaged and tested. Amkor Technology is one of the largest independent back-end providers, known in the industry as an outsourced semiconductor assembly and test (OSAT) company, with operations across the United States, Japan, Europe and Asia Pacific. Historically much of its business came from communications chips, especially for smartphones.
The AI boom changed the economics of the back end. High-end accelerators combine processors and high-bandwidth memory in complex packages that require specialised lines. TSMC’s in-house advanced packaging has been widely reported as a key supply constraint for AI chips in recent years. That has pushed packaging from a commodity step toward a strategic input, and Amkor has been expanding AI-focused capacity, including in Arizona, to meet demand from data center chip designers. Source: Amkor Technology (AMKR) Lands $1.3 Billion Nvidia Deal As Fair Value Stays In Focus (Simply Wall St). Report on Amkor’s new US$1.3 billion Nvidia capacity agreement in South Korea and its share-price performance.Sources