VLSI KoreaNVIDIA's Substrate Bet: 2.5x Capacity, a USD 1.1 Billion Testing Advance, and Korea's New Seat in the AI Supply Chain
On October 21 at GTC Berlin, NVIDIA struck capacity deals with six partners — Samsung Electro-Mechanics, IBIDEN, Unimicron, Amkor, KYEC and SPIL — for 2.5x substrate capacity, 300%+ more Korean interposer packaging and 15x co-packaged optics.
🆓 Free article · Anyone can read this article for free
The bottleneck in AI chips has moved downstream — and on October 21, 2026, NVIDIA paid to move it. At GTC Berlin, NVIDIA announced capacity agreements with six manufacturing partners — Amkor Technology, IBIDEN, King Yuan Electronics (KYEC), Samsung Electro-Mechanics, Siliconware Precision Industries (SPIL) and Unimicron — spanning package substrates, advanced packaging, semiconductor testing and co-packaged optics. The headline numbers: substrate capacity up 2.5x over three years, interposer packaging capacity in South Korea up by more than 300% by the end of 2028, testing capacity doubling between 2026 and 2027, and optical-engine capacity up 15x by mid-2027. Two of the six partners are expanding in South Korea, and NVIDIA is putting cash upfront — a multibillion-dollar advance to IBIDEN and USD 1.1 billion to KYEC — to make sure the capacity actually gets built.
This is a company-radar deep research piece based on NVIDIA's October 21, 2026 announcement (NVIDIA Newsroom / GlobeNewswire), cross-checked against Samsung Electro-Mechanics' own partnership release and wire coverage of the same event. Market data referenced below is tracked on VLSI Market's company wire and stock pages.
What happened: the October 21 announcement, fact by fact
NVIDIA framed the announcement in one sentence from founder and CEO Jensen Huang: "We are in the midst of the largest infrastructure buildout in human history. Demand is extraordinary. Capacity is the constraint." The agreements are structured by supply-chain segment:
| Segment | Partner(s) | Committed target | Timing |
|---|---|---|---|
| Package substrates | IBIDEN, Samsung Electro-Mechanics, Unimicron | Aggregate substrate capacity 2.5x over three years | 2026–2028 |
| Substrates — Japan | IBIDEN (Ono Plant) | Capacity to 3x current levels; NVIDIA provides a multibillion-dollar capacity reservation advance | Additional capacity ramps from Q1 2028 |
| Substrates — Korea (new entrant) | Samsung Electro-Mechanics (Sejong, Busan) | Joins NVIDIA's substrate supply chain; existing Korean capacity immediately, plus facility upgrades and new factories | Multiyear supply from 2026; by 2028 NVIDIA is projected to take a significant portion of Samsung Electro-Mechanics' total annual package substrate capacity |
| Substrates — Taiwan | Unimicron (KF expansion) | Capacity reservation agreement supporting KF substrate expansion | Capacity starts Q2 2027; full capacity by mid-2028 |
| Advanced packaging | Amkor Technology (South Korea) | Interposer wafer capacity up by more than 300%, alongside new Korean facilities | By end of 2028 |
| Testing | KYEC (United States), plus Taiwan and Singapore buildouts | Total testing capacity doubles from 2026 to 2027; NVIDIA pays a USD 1.1 billion capacity reservation advance for new US testing facilities and equipment | Singapore high-volume production February 2027; KYEC US high-volume production April 2027 |
| Co-packaged optics | SPIL (Taiwan) | Optical-engine capacity 15x, then a further doubling | 15x by mid-2027; doubling by end of 2028 |
Samsung Electro-Mechanics issued its own release the same day, calling the deal a strategic business partnership to strengthen the global AI semiconductor supply chain. It will supply package substrates for NVIDIA's AI compute chipsets and for the networking (switch) chips that link chips inside data centers, under a multiyear agreement beginning in 2026. Korean coverage of the announcement noted this is the first time NVIDIA has formalized a strategic partnership with Samsung Electro-Mechanics, and that NVIDIA-bound substrate volumes are expected to grow by double digits each year.
Technical anatomy: why substrates, interposers, test and optics — all at once
Substrates are the quiet constraint of the AI era. Every large AI accelerator sits on an organic package substrate (in high-end AI packages, an ABF-based FC-BGA substrate) that fans out thousands of connections from the package to the board. As dies grow toward and beyond reticle limits and packages add more HBM stacks, substrates get larger, gain layers, and become harder to build without warpage or yield loss. Fewer suppliers can build the largest formats at yield, and capacity additions take years — which is why substrate lead times, not wafer starts, have repeatedly set the pace for AI accelerator shipments. For background, see our deep anatomy of ABF substrates.
The interposer is the second constraint. Advanced packaging for AI accelerators routes die-to-die and die-to-HBM traffic through a silicon interposer. Interposer wafer capacity is concentrated and has been the pacing item of advanced packaging lines for several years. NVIDIA and Amkor's plan to raise interposer wafer capacity in South Korea by more than 300% by the end of 2028 is therefore not a routine OSAT expansion — it is a geographic diversification of the single most contested packaging step, into the same country that supplies most of the world's HBM.
Testing scales with units, not with wafers. Every finished accelerator must be tested, and test time rises with device complexity. NVIDIA's plan to double total testing capacity between calendar 2026 and 2027 — with KYEC building new US facilities on a USD 1.1 billion advance and Singapore lines reaching high-volume production in February 2027 — shows final test being treated as a strategic capacity pool rather than a commodity service bought on the spot market.
Optics is the forward bet. SPIL's 15x optical-engine expansion targets co-packaged optics (CPO), where optical engines move next to the switch or accelerator package to cut the power and reach limits of copper. Notably, NVIDIA is funding both paths at once: on October 12 it introduced a new pluggable optics form factor with Cisco, and nine days later it committed to a 15x CPO optical-engine buildout with SPIL. That is a hedge, not a contradiction — pluggable optics serves today's installed base, CPO serves the next platform generation.
The financing structure is the real story. In two of the six deals NVIDIA is explicitly paying capacity reservation advances — multibillion-dollar money to IBIDEN and USD 1.1 billion to KYEC — so suppliers can build ahead of demand. The buyer with the strongest balance sheet in the chain is now financing its suppliers' capex. That shifts construction risk from supplier to buyer, locks in allocation, and tells you the bottleneck is physical, not financial: money is available, cleanrooms and equipment slots are what is scarce.
Competitive landscape: who held the bottleneck before this deal
| Segment | Before October 21 | What changes |
|---|---|---|
| High-end substrates | Supply concentrated in a small set of Japanese and Taiwanese makers (IBIDEN, Unimicron among them); Korean makers had limited presence in top-tier AI substrates | Samsung Electro-Mechanics formally enters NVIDIA's substrate chain with dedicated Sejong and Busan expansion — a third geography for the most constrained input |
| Advanced packaging / interposer | Interposer capacity concentrated in Taiwan, tied to one dominant foundry ecosystem | Amkor expands interposer capacity 300%+ in South Korea, creating a second hub next to HBM supply |
| Final test | OSAT test capacity largely in Taiwan; bought as a service | NVIDIA prepays USD 1.1 billion to put new KYEC test capacity on US soil, with Singapore ramping first |
| Optics | Pluggable optics dominant; CPO in early deployment | SPIL optical-engine capacity 15x by mid-2027 underwrites CPO at platform scale, while the Cisco pluggable path continues in parallel |
The pattern across all four rows is the same: NVIDIA is converting single-geography, single-supplier dependencies into financed, multi-geography capacity pools — and paying advances to accelerate the conversion. This is the same logic as its rack-scale DSX POD push for faster AI factory deployment, which we covered in our analysis of the Vera Rubin DSX POD announcement: compress every timeline that sits between a wafer and a running AI factory.
📩 Following this story?
Get the follow-up analysis — earnings, reports and HBM supply — in your inbox daily at 10:30 KST. Subscribe free →
What it means for Korea and the regional supply chain
South Korea is the only country that appears twice in this announcement — Samsung Electro-Mechanics in substrates and Amkor in advanced packaging — on top of its existing position in HBM through Samsung Electronics and SK hynix. Three implications stand out.
First, Samsung Electro-Mechanics moves from component supplier to named AI infrastructure partner. The company already supplies FC-BGA substrates and has been investing through a difficult yield and utilization cycle — context we laid out in our October analysis of Samsung Electro-Mechanics' FC-BGA capex. A multiyear NVIDIA agreement, with NVIDIA volumes projected to reach a significant portion of its total substrate capacity by 2028, changes the demand side of that investment case from merchant-market hope to contracted anchor demand.
Second, Korea gains an advanced-packaging beachhead outside the memory stack. Amkor's Korean interposer expansion puts a US-headquartered OSAT's most advanced packaging growth in Korea, adjacent to the HBM supply base. If interposer, HBM and substrate capacity co-locate, Korea's AI supply-chain role widens from "memory supplier" to "packaging hub" — a materially different negotiating position.
Third, the customer-funding model reaches Korean capex. The announcement does not disclose a Samsung Electro-Mechanics advance figure, and none should be assumed. But the structure NVIDIA used with IBIDEN and KYEC — buyer money first, capacity second — is now the template every Korean supplier negotiating an AI capacity expansion will be measured against.
Risks and what to watch next
- Timing risk: Most of the new capacity lands in 2027–2028 (Unimicron from Q2 2027, IBIDEN ramping Q1 2028, Amkor and SPIL targets at end-2028). It relieves the next platform cycle, not this quarter's shortage.
- Execution risk at Samsung Electro-Mechanics: Top-tier AI substrates are a yield game. The Sejong and Busan expansion must qualify large-format, high-layer substrates at NVIDIA's standards; the ramp, not the announcement, is the proof point.
- Concentration risk in reverse: By 2028 a significant portion of Samsung Electro-Mechanics' substrate capacity is projected to serve one customer. Anchor demand de-risks utilization — and concentrates customer risk.
- Demand risk: Capacity reservation advances are sunk commitments against AI infrastructure demand holding its current trajectory through 2028. If hyperscaler capex growth slows, the advances, not the suppliers' balance sheets, absorb the first loss — which is precisely why suppliers accepted the structure.
- Watch points: KYEC's US high-volume production start (April 2027), Singapore test ramp (February 2027), SPIL's mid-2027 optics milestone, IBIDEN's Ono ramp (Q1 2028), and Samsung Electro-Mechanics' quarterly substrate revenue mix for the first visible NVIDIA contribution.
Frequently asked questions
Q: What did NVIDIA announce on October 21, 2026?
A: At GTC Berlin, NVIDIA announced capacity expansion agreements with six partners — Amkor, IBIDEN, KYEC, Samsung Electro-Mechanics, SPIL and Unimicron — covering substrates, advanced packaging, testing and co-packaged optics, including a 2.5x substrate capacity target over three years and capacity reservation advances to IBIDEN and KYEC.
Q: Why are package substrates a bottleneck for AI chips?
A: Large AI accelerators need large, high-layer ABF substrates that few suppliers can build at yield. Substrate capacity takes years to add, so it can pace accelerator shipments even when wafer and HBM supply are available.
Q: What will Samsung Electro-Mechanics supply to NVIDIA?
A: Package substrates for NVIDIA's AI compute chipsets and data-center networking switch chips, under a multiyear agreement from 2026, with dedicated line expansion in Sejong and Busan. By 2028, NVIDIA is projected to account for a significant portion of Samsung Electro-Mechanics' total annual package substrate capacity.
Q: What is a capacity reservation advance?
A: Upfront money a buyer pays a supplier to reserve future manufacturing capacity and help fund its construction — here, a multibillion-dollar advance to IBIDEN and USD 1.1 billion to KYEC. It shifts build-out risk to the buyer in exchange for guaranteed allocation.
Q: When does the new capacity come online?
A: In stages: Singapore testing high-volume production in February 2027, KYEC US testing in April 2027, Unimicron substrate capacity from Q2 2027, SPIL optical engines at 15x by mid-2027, IBIDEN's Ono expansion ramping from Q1 2028, and Amkor's 300%+ Korean interposer increase plus SPIL's further optics doubling by the end of 2028.
Chase's Take
From a design and timing background, the most telling detail in this announcement is not any single multiplier — it is which steps NVIDIA chose to finance. Substrates, interposers, final test and optical engines are the unglamorous back half of the flow, the steps that never appear in a keynote render but decide whether finished systems ship on schedule. Anyone who has waited on a packaging or substrate slot during a tape-out crunch recognizes this pattern: the critical path moves, and it usually moves to the step nobody was watching.
NVIDIA is effectively admitting that its constraint for 2027–2028 is no longer GPU architecture or even HBM — it is the physical back end, and it is buying its way out with advances, second geographies and a Korean packaging hub. For Korea, the strategic read is bigger than one substrate contract: memory, substrate and interposer capacity in one country is the beginnings of a full AI back-end stack. The open question is execution — yields at Samsung Electro-Mechanics and construction pace at Amkor Korea will decide whether October 21, 2026 is remembered as an announcement or as the quarter Korea's role in the AI supply chain actually changed.
Sources: NVIDIA announcement, "NVIDIA and Supply Chain Partners Expand Capacity to Accelerate AI Infrastructure Buildout," October 21, 2026 (NVIDIA Newsroom / GlobeNewswire); Samsung Electro-Mechanics strategic partnership release as reported by Yonhap News Agency, October 21–22, 2026; Chosun Biz coverage of the GTC Berlin announcement, October 21, 2026. Forward-looking capacity targets are the companies' stated plans, not guarantees. This article is analysis, not investment advice.
📩 Missed today's analysis? Get tomorrow's in your inbox
Daily Silicon at 10:30 KST plus earnings and analyst-report breakdowns, first. Samsung Electronics, SK hynix, TSMC and HBM supply, read by practicing engineers. Join 300+ subscribers.
Subscribe free →
Free · No spam · Unsubscribe anytime
