Nvidia reserves chip capacity with six suppliers, from substrates to optical engines

Nvidia said on October 21, 2026 that it is working with six suppliers to expand capacity in substrates, advanced packaging, testing and optical engines, including a $1.1 billion advance to testing firm KYEC and a multibillion-dollar reservation with substrate maker IBIDEN. The release, issued at GTC Berlin, is a rare look at where the physical limits on AI chips actually sit: not only in wafers, but in the parts around them.

What Nvidia announced

According to Nvidia's press release, the partners are Amkor, IBIDEN, King Yuan Electronics (KYEC), Samsung Electro-Mechanics, Siliconware Precision Industries (SPIL) and Unimicron. Founder and CEO Jensen Huang said: "We are in the midst of the largest infrastructure buildout in human history." He added: "Capacity is the constraint."

The numbers, all from Nvidia's own release and all targets or plans, not delivered output:

• Substrates: Nvidia plans a 2.5x increase in substrate capacity for its products over three years.
• IBIDEN: a multibillion-dollar capacity reservation advance for its Ono Plant in Japan, which targets 3x current plant capacity, with additional capacity ramping from Q1 2028.
• Unimicron: a capacity reservation agreement for its KF substrate expansion in Taiwan, with additions starting Q2 2027 and full capacity by mid-2028.
• Amkor: a greater than 300% increase in interposer wafer capacity in South Korea by the end of 2028.
• Testing: total semiconductor testing capacity expected to double from calendar 2026 to 2027, with buildouts in the U.S., Taiwan and Singapore. KYEC receives a $1.1 billion advance for new U.S. facilities, with high-volume production targeted for April 2027.
• Optical: SPIL targets a 15x increase in optical-engine capacity in Taiwan by mid-2027 and another doubling by the end of 2028.

Why these parts matter

A frontier accelerator is not just a die. It needs a package substrate to carry signals and power, an interposer to join the processor to its memory, and a test step before shipping. Each is a place where supply can pinch. When Nvidia pays advances for reserved capacity, it is buying priority on those steps years before the output exists.

The pattern is not new. In July 2026 Reuters reported that Amkor entered a multi-year agreement with Nvidia worth $1.5 billion to expand advanced packaging and test capacity in the U.S., with Nvidia making a prepayment that includes capacity in Arizona. Amkor shares rose 17% in extended trading on that news. The October release extends the same logic across a wider supplier list.

What the timeline says about the race

Read the dates, not the multiples. IBIDEN's extra capacity begins ramping in Q1 2028. Unimicron reaches full capacity in mid-2028. Amkor's interposer target is end of 2028. Those are the years when today's training plans would need to be fully fed. The headline "largest buildout in human history" sits next to a schedule that mostly lands in 2027 and 2028.

That fits the demand signals elsewhere, including the foundry side, which we covered in TSMC's third-quarter report. Nvidia's supplier list is the same story told from the back end of the line.

How this connects to the rack

The other half of Nvidia's October 21 news shows where the finished product goes. Nvidia and Google Cloud announced general availability of Vera Rubin infrastructure the same day, with Google Cloud planning to scale to as many as 960,000 Rubin GPUs across multiple sites. A figure like that only holds if substrates, packages and test slots exist for every one of those parts. The supplier release is the plumbing behind that headline.

What the release does not say

It does not say how much any of this adds up to in finished accelerators, and it gives no chip counts. It does not state the size of most advances: IBIDEN's is described only as "multibillion-dollar," and the Singapore testing operator is not named. It does not say who bears the risk if demand softens. A reservation advance is a bet by Nvidia, but a capacity expansion is a bet by the supplier, who has to build the factory.

Every figure comes from Nvidia, a seller with an interest in showing a confident supply chain. We found no independent confirmation of the capacity targets as of this writing, so treat them as plans.

Power is the other constraint

Nvidia's same-day blog on its DSX platform puts the second bottleneck bluntly: "Power is now the biggest bottleneck in infrastructure." It cites more than $300 billion in equipment backlogged and over 200 gigawatts of projects waiting in U.S. interconnection queues. Packaging and substrates decide how many chips can be built. Power decides where they can be plugged in. Both constraints are binding at once.

The hype check

The claim: capacity expansion will let AI infrastructure scale. The evidence: real prepayments with named suppliers and dated ramps, but targets only, from the party that benefits. Our view is that the dollar advances are the credible part, because money paid in advance is hard to fake. The multiples are the soft part.

Our take

We would watch two things. First, whether KYEC's U.S. testing line and the Singapore operation reach high-volume production on the stated dates of April and February 2027. Those are the earliest checkable milestones. Second, whether suppliers confirm the targets in their own filings. If they do, the constraint on frontier compute shifts later in the decade from packaging to power. If they slip, the race is slower than the capex headlines imply.