AI chips are not only limited by wafers any more. On October 21, 2026, Amkor announced a new $1.3 billion agreement with Nvidia to expand capacity in South Korea and said its Nvidia agreements in Arizona and South Korea now total $3 billion, per Amkor's investor release. The same day, Nvidia lined up a wider set of packaging, substrate and test expansions with five other suppliers, summarized by Mark LaPedus in Semiecosystem.
This post lists the deals, the stated targets, and a number check: the "$3 billion" total does not equal the two announced figures. It follows our July coverage of Nvidia's first Amkor Arizona deal and GB300 packaging.
TL;DR
| Supplier | What Nvidia is doing | Stated numbers |
|---|---|---|
| Amkor | Expands advanced packaging in South Korea | New $1.3B agreement; more than 300% interposer wafer capacity by end of 2028 (Semiecosystem); Amkor says Arizona plus Korea total $3B |
| SPIL (ASE unit) | Expands optical engines for co-packaged optics in Taiwan | 15x increase by mid-2027, another doubling by end of 2028 |
| IBIDEN | Capacity reservation advance for the Ono plant, Japan | Multibillion-dollar advance; capacity to 3x current levels, more ramping from Q1 2028 |
| Samsung Electro-Mechanics | Joins Nvidia's substrate supply chain | ABF substrate lines in Sejong and Busan; supply beginning 2026 |
| Unimicron | Capacity reservation for KF substrate expansion in Taiwan | Adds capacity from Q2 2027, full by mid-2028 |
| KYEC | Capacity reservation advance for U.S. testing | $1.1B; high-volume production targeted April 2027 |
| Overall | Substrate capacity and test capacity | Substrate capacity 2.5x over three years; testing capacity to double from 2026 to 2027 |
The Amkor numbers: claimed versus verified
Amkor's release says: "With this latest commitment, Amkor's multi-year agreements with NVIDIA in Arizona and South Korea total $3 billion." The new Korea deal is $1.3 billion. The earlier Arizona agreement, announced July 23, 2026, was described by coverage including TechRepublic as a $1.5 billion multi-year advanced packaging and development agreement.
| Item | Amount | Source |
|---|---|---|
| Arizona agreement (July 2026) | $1.5 billion | Amkor and press coverage of the July announcement |
| Korea agreement (Oct 21, 2026) | $1.3 billion | Amkor release, Semiecosystem |
| Sum of the two | $2.8 billion | Our arithmetic |
| Stated total | $3 billion | Amkor release |
The difference of about $0.2 billion is not explained in the release or in the coverage we read. Possible explanations include an unannounced amendment to the Arizona agreement or rounding, but those are guesses. Treat "$3 billion" as Amkor's stated total, and $2.8 billion as the sum of the figures that have been itemized publicly. Both numbers are prepayments and commitments, not revenue.
What each deal is for
Amkor: interposers in Korea
The Korea expansion targets a greater than 300% increase in interposer wafer capacity by the end of 2028, alongside new manufacturing facilities, according to Semiecosystem. The release says the prepayments "align Amkor's manufacturing capacity with NVIDIA's long-term growth plans." An interposer is the silicon layer that links GPU dies to memory stacks in advanced packages, so interposer capacity is a direct limit on how many top-end accelerators can be built. Nvidia's Debora Shoquist said the work helps scale production of current and next-generation platforms while strengthening the resilience of the global supply chain.
SPIL: optical engines
In a separate deal, SPIL will expand Taiwan capacity for optical engines used in co-packaged optics, with a target of a 15x increase by mid-2027 and another doubling by the end of 2028. Co-packaged optics put optical connections next to the processor to ease networking bottlenecks in large clusters.
Substrates: IBIDEN, Samsung Electro-Mechanics, Unimicron
Nvidia plans to raise substrate capacity for its products by 2.5x over three years. IBIDEN gets a multibillion-dollar capacity reservation advance for its Ono plant in Japan, aiming for 3x current capacity with additional ramp from the first quarter of 2028. Samsung Electro-Mechanics will build dedicated ABF substrate lines at Sejong and Busan and plans to supply Nvidia starting in 2026, with Nvidia-allocated output projected to become a significant portion of its annual package substrate production by 2028. Unimicron's KF substrate expansion in Taiwan begins adding capacity in the second quarter of 2027 and reaches full capacity by mid-2028.
Testing: KYEC
Nvidia's $1.1 billion capacity reservation advance to KYEC supports new U.S. testing facilities and equipment, with high-volume production targeted for April 2027. Nvidia expects total semiconductor testing capacity to double from 2026 to 2027, with buildouts in the U.S., Taiwan and Singapore; Singapore operations are expected to start high-volume production in February 2027.
TSMC was not part of the announcement, though it provides foundry and packaging services for Nvidia and many others.
Timeline of the stated milestones
| Date | Milestone | Supplier |
|---|---|---|
| 2026 | Samsung Electro-Mechanics begins supplying ABF substrates | Samsung Electro-Mechanics |
| February 2027 | Singapore testing operations start high-volume production | Nvidia test buildout |
| April 2027 | KYEC U.S. testing facilities reach high-volume production | KYEC |
| Q2 2027 | Unimicron begins adding KF substrate capacity | Unimicron |
| Mid-2027 | Optical engine capacity up 15x | SPIL |
| Mid-2028 | Unimicron reaches full capacity | Unimicron |
| Q1 2028 | Additional IBIDEN Ono capacity begins ramping | IBIDEN |
| End of 2028 | Interposer wafer capacity up more than 300%; optical engines double again | Amkor, SPIL |
Advanced packaging in plain terms
A modern AI accelerator is not one chip. It is several compute dies and stacks of high-bandwidth memory mounted side by side on an interposer, then on a package substrate, and wired out to the board. Each layer has its own supply chain.
- Interposer. A silicon layer with very fine wiring that lets dies and memory talk at very high bandwidth. Capacity is limited by specialized equipment.
- ABF substrate. The multilayer organic board under the interposer, historically a tight market with long lead times.
- Test. Every package must be tested, and complex multi-die parts take longer to test than single chips.
- Optical engines. Components that convert electrical signals to light at the package edge, used in co-packaged optics to cut power in networking.
When one of these layers is short, the whole accelerator is short, no matter how many wafers the foundry makes. That is why Nvidia is reserving capacity at each layer rather than one.
Why it matters
Our reading: when the buyer with the best balance sheet starts financing its suppliers' capex through advances and capacity reservations, it suggests packaging, substrates and test, not wafers, are where the constraint has moved. That is interpretation, not something Nvidia states, but it fits the facts in the announcements: advances and reservations, multi-year ramps into 2027 and 2028, and targets expressed as multiples of existing capacity.
For people who build with AI, the effect is indirect but real:
- Supply for Rubin-era systems. Packaging, substrates and testing gate how many accelerators ship. These deals aim to raise that ceiling over 2027 and 2028, not immediately.
- Prices and lead times. Constrained packaging supports high accelerator prices and long waits. Our coverage of the AI server component shortage and of Samsung's HBM4 yield progress shows the same pattern in other parts of the stack. See also Michael Dell's view of the memory shortage.
- Geography. The deals spread capacity across the U.S., Korea, Japan, Taiwan and Singapore, which reduces concentration risk.
- Networking. The SPIL targets show Nvidia is betting on co-packaged optics for large systems, linking to the new factory-scale designs we covered in the DSX pod claims.
How to read a "capacity multiple"
Announcements like "15x" and "more than 300%" sound huge, but they are only as meaningful as the base. A 300% increase means roughly quadrupling if read as growth, or tripling if "increase to 3x" is meant; the sources use both phrasings for different deals (IBIDEN is "3x current levels", Amkor is a "greater than 300% increase"). Neither Nvidia nor Amkor publishes absolute wafer or unit counts, so the multiples cannot be converted into accelerators per quarter. A reader should also separate announced targets from delivered capacity: ramps are scheduled for 2027 and 2028, and the same release warns that construction timing and cost can change.
What is not stated
| Gap | Why it matters |
|---|---|
| Where the extra $0.2B in Amkor's total comes from | The headline does not match the itemized deals |
| Exact prepayment terms and clawbacks | Determines risk if demand falls |
| Baselines for the capacity multiples | A 300% increase from a small base may be a modest absolute number |
| Delivery slippage risk | Amkor's release warns the Arizona campus may not be built on timeline, at cost or to specification |
| Price effects | None of the announcements discuss pricing |
What this means for what you buy or build
If you plan GPU capacity for 2027, assume packaging and testing remain constrained through at least the first half of the year, and watch the dates in these announcements as the earliest points for relief: KYEC's April 2027 start, Singapore testing in February 2027, Unimicron's second quarter 2027 ramp, and SPIL's mid-2027 optical engine target. Cloud buyers should expect providers to keep selling reserved capacity and to favor long-term contracts; independent teams should favor smaller models and efficiency work that reduce dependence on the newest accelerators. Our NVIDIA Europe coverage shows how providers are already marketing capacity.
Related reading
- Nvidia GB300, TSMC Arizona and Amkor packaging (July)
- AI server MLCC shortage and 16-month lead times
- Samsung HBM4 yield ahead of schedule
- Michael Dell on the memory shortage
- Nvidia DSX pod: claimed vs verified
- Nvidia Vera CPUs in Europe
- Sources: Amkor investor release · Semiecosystem
Figures are from Amkor's October 21, 2026 release and Semiecosystem's summary of Nvidia's announcements; the July Arizona amount comes from press coverage of that announcement.