Nvidia puts over $1B into Kinsus for 3.5x substrate capacity

Nvidia is putting more than US$1 billion into Kinsus to buy equipment that lifts its Nvidia-dedicated ABF substrate capacity more than 3.5 times by the third quarter of 2028, so the chip buyer, not the supplier, now pays for the board under the AI chip.

Share
Macro photo of circuit board wiring graded navy and teal, with a grid of solder balls lit amber; overlay reads 3.5x, Nvidia pays for the board under the chip.
Illustrative image: macro photo of circuit board wiring and a solder-ball grid (not a Kinsus or Nvidia part). Photo by quapan via Wikimedia Commons, CC BY 2.0 (cropped, colour-graded, text added).

Nvidia is putting more than US$1 billion into Kinsus Interconnect Technology to buy equipment that will lift Kinsus's ABF substrate capacity dedicated to Nvidia to more than 3.5 times today's level by the third quarter of 2028, which means the buyer, not the supplier, is now paying for the board under the AI chip.

What happened?

  • The Nvidia release (17 November 2026, US time). Nvidia said it signed a new multiyear agreement under which Kinsus plans to expand its advanced ABF package substrate capacity dedicated to Nvidia by more than 3.5 times from its Youth campus in Taoyuan, Taiwan. Production is expected to begin in the second half of 2027, with the added capacity targeted for the third quarter of 2028.
  • The money (18 November 2026, Taiwan time). Kinsus said the deal comes with a more than US$1 billion investment from Nvidia. In its Chinese-language statement, as reported by the Commercial Times, Kinsus said Nvidia is putting in cash, about NT$33 billion, to buy new equipment for the Youth plant expansion, and that the two sides are negotiating investment plans for further plants so Kinsus can supply Nvidia at least until 2030.
  • The rest of the release. Nvidia also amended its substrate supply agreement with Samsung Electro-Mechanics for existing factories in Sejong and Busan, South Korea (no amount given). It signed a new multiyear packaging agreement with SPIL in Taiwan, backed by what it called multibillion-dollar funding, covering bumping, interposer wafer fabrication, assembly and test, with more dedicated capacity by mid-2029. It restated that its Amkor agreements from July and October total nearly $3 billion, including prepayments.
  • The demand side. In the same release Nvidia said Vera Rubin NVL72 racks and Vera CPUs began volume shipments in its third quarter, built on TSMC's 5nm and 3nm processes and CoWoS packaging.

Why does it matter?

The direct signal: one substrate maker gets a customer-funded expansion. Our arithmetic puts the size in context. NT$33 billion is about 2.3 times Kinsus's revenue for the July to September quarter (NT$14.54 billion, the sum of its monthly revenue reports) and about 76% of everything it booked from January to October 2026 (NT$43.51 billion). A supplier of that size could not easily fund a jump of this scale from its own cash flow without stretching its balance sheet. Nvidia's money removes that problem for the Youth lines.

The second-order signal: the board under the chip is now a step the chip designer pays to clear. A finished AI GPU is a stack: the TSMC die, its memory stacks on a silicon interposer, and then an ABF substrate, the layered wiring board that connects the whole package to the server board. As packages get larger and add more layers, each GPU uses more substrate area, so substrate output can cap how many finished GPUs ship. Our IC substrates and ABF explainer covers why this step is slow to add. In October, Nvidia said it planned 2.5 times more substrate capacity over three years across IBIDEN, Unimicron and Samsung Electro-Mechanics, as covered in our Catalyst on that round. Kinsus is a new name in that list, and its 3.5 times is higher than the 2.5 times group average, for one supplier's dedicated lines.

Who carries the risk. Kinsus says the money buys equipment. Neither company says whether it is a prepayment against future substrate deliveries, a loan or an equity stake, and that matters. A prepayment usually comes back to Nvidia as credit on future orders, so Kinsus keeps the machines but owes the output. Our customer prepayment explainer sets out how that usually works. Kinsus still runs the plant and carries the risk that demand for Nvidia-dedicated substrates is lower in 2028 than planned.

Why the timing matters. Nvidia's CFO commentary for the quarter shows $390 billion of supply and capacity commitments as of 25 October 2026. Of that, $203 billion falls in fiscal 2028, which runs to late January 2028, and $92 billion in fiscal 2029 (Nvidia's schedule). Our arithmetic: 52% of the total lands in the one fiscal year in which the Kinsus lines start up. Nvidia said most of the quarter's increase came from multi-year memory procurement, so this table is not a substrate figure, but it shows how much of Nvidia's supply bill sits in 2027.

What does it change?

Near term (to mid-2027): nothing in supply. The new lines do not produce until the second half of 2027, so the substrates for Vera Rubin in 2026 and the first half of 2027 come from capacity that already exists at Kinsus, IBIDEN, Unimicron, Samsung Electro-Mechanics and others. The near-term change is on Kinsus's books: equipment orders, and possibly a large cash inflow, before any revenue from the new lines.

For the packaging step (2H 2027 to 2029): Nvidia has now named dated substrate and packaging adds running from 2027 to mid-2029: Unimicron from the second quarter of 2027, IBIDEN's Ono plant from the first quarter of 2028, Kinsus from the second half of 2027 to the third quarter of 2028, and SPIL by mid-2029. Our arithmetic: the figures Nvidia has given in full, nearly $3 billion for Amkor, $1.1 billion for KYEC's US test sites and more than $1 billion for Kinsus, add up to more than $5 billion, before the IBIDEN and SPIL amounts it has only called multibillion. The pattern is that Nvidia pays for back-end capacity years ahead, and suppliers that take the money give up some freedom over who gets the new output.

For other substrate buyers: capacity dedicated to Nvidia is, by definition, not sold to other chip designers such as AMD or the custom-chip teams at the hyperscalers. If Nvidia keeps funding dedicated lines across suppliers, other buyers may find less spare substrate capacity, though no company has said this. It is a risk to watch, not a fact.

What could weaken it: a slower Vera Rubin ramp, a delay at the Youth campus, or a sign that Nvidia's demand plan for 2028 is shrinking. Each of those would leave Kinsus with dedicated equipment and less use for it.

Metric Value Source name Date
Nvidia funding to Kinsus More than US$1 billion in cash, about NT$33 billion, for equipment Kinsus; Commercial Times 18 Nov 2026
Dedicated ABF capacity target More than 3.5 times current levels at the Youth campus, Taoyuan Nvidia release 17 Nov 2026
Timing Production from 2H 2027; added capacity by 3Q 2028 Nvidia release 17 Nov 2026
SPIL packaging deal Multibillion-dollar Nvidia funding; more dedicated capacity by mid-2029 Nvidia release 17 Nov 2026
Amkor agreements Nearly $3 billion, with prepayments (announced July and October 2026) Nvidia release 17 Nov 2026
Nvidia supply and capacity commitments $390 billion; $203 billion in fiscal 2028 Nvidia CFO commentary 17 Nov 2026
Kinsus revenue, Jul to Sep 2026 NT$14.54 billion (our arithmetic from monthly reports) Kinsus monthly revenue Oct 2026

What should we watch next?

  • The form of the money: Kinsus's filings with the Taiwan exchange and its fourth-quarter report should show whether the US$1 billion is booked as a customer prepayment, a contract liability, a loan or equity. Nvidia's quarterly report to the SEC may show it as a prepaid supply or capacity agreement.
  • The next Kinsus plants: both sides say they are negotiating investment for more Kinsus sites, with supply aimed at least until 2030. A second agreement would show this is a series, not a single deal.
  • The SPIL amount: Nvidia has only called the SPIL funding multibillion. An ASE or SPIL disclosure with a figure would size the largest open item in this release.
  • Production start: first output from the Youth lines in the second half of 2027, then the full 3.5 times by the third quarter of 2028. A slip would be the first sign the substrate step is still holding back GPU shipments.
  • What would strengthen the catalyst: Kinsus capital spending rising in its 2027 plans, and other substrate makers announcing similar Nvidia-funded lines.

Also on the tape

  • 16 Nov: Core Scientific filed with the Texas licensing regulator for a $1.3 billion core and shell data center building at its Hunt County site, which it says can support up to 285 MW of leasable capacity, about half of it set to be leased to AMD (Data Center Dynamics).
  • 17 Nov: Giga Energy announced a 10-year, $1 billion lease with an unnamed US AI cloud provider for Project Clay, a 60 MW AI data center in Hunt County, Texas, due for service in spring 2027 (Giga Energy).
  • 17 Nov: Argentum AI and Kyndryl signed a 20-year, $630 million agreement for an initial 10 MW at Kyndryl's Boulder, Colorado facility, where Argentum is deploying Nvidia B300 systems (Kyndryl).
  • 17 Nov: AMD said it had completed its acquisition of Taalas, a developer of specialised AI inference chips, and plans to fold the technology into its accelerator roadmap (AMD).

FAQ

How much is Nvidia paying Kinsus?
More than US$1 billion, which the Commercial Times converted to about NT$33 billion. Kinsus says the cash pays for new equipment at its Youth plant in Taoyuan. Neither company has said whether it is a prepayment, a loan or equity.

What is an ABF substrate?
It is the layered wiring board a chip package sits on, built with Ajinomoto Build-up Film as the insulating layer between fine copper lines. It carries signals and power from the GPU and its memory to the server board. Our guide to who makes IC substrates lists the main suppliers.

When does the new capacity arrive?
Production is expected to start in the second half of 2027, with the full added capacity, more than 3.5 times today's Nvidia-dedicated level, targeted for the third quarter of 2028.

Is this the same as Nvidia's October substrate plan?
No. October covered IBIDEN, Unimicron and Samsung Electro-Mechanics and a group target of 2.5 times over three years. Kinsus is a new agreement, and the Samsung Electro-Mechanics part of this release is an amendment to the existing deal.

Takeaway: Nvidia is paying more than $1 billion to lift one supplier's dedicated substrate capacity 3.5 times by 2028, so the board under the chip is now something the chip designer funds, and the supplier keeps the plant risk.

References

  1. NVIDIA, NVIDIA Expands Commitments With Supply Chain Partners to Support AI Infrastructure Buildout, 17 Nov 2026.
  2. Kinsus Interconnect Technology, Kinsus Announces more than US$1 Billion Investment From NVIDIA to Expand Advanced Substrate Capacity for AI, 18 Nov 2026.
  3. Commercial Times, 輝達砸逾10億美元投資景碩 擴充AI載板產能 (Nvidia invests over US$1 billion in Kinsus to expand AI substrate capacity), 18 Nov 2026.
  4. NVIDIA, CFO Commentary on Third Quarter Fiscal 2027 Results (Exhibit 99.2, SEC), 17 Nov 2026.
  5. Kinsus Interconnect Technology, Monthly Revenue 2026, accessed 18 Nov 2026.
  6. Data Center Dynamics, Core Scientific files for 285MW data center outside Dallas, Texas, 16 Nov 2026.
  7. Giga Energy, Giga Energy Inks 10-Year, $1B Lease on Project Clay, a 60MW AI Data Center in Hunt County, TX, 17 Nov 2026.
  8. Kyndryl, Argentum AI signs $630M data center collaboration with Kyndryl to deploy NVIDIA GPUs, 17 Nov 2026.
  9. AMD, AMD Completes Acquisition of Taalas to Advance AI Inference Innovation, 17 Nov 2026.