NVIDIA prepays $1.1 billion to tester KYEC as substrates rise 2.5x in three years and optical engines 15x, moving AI's bottleneck past the chip

NVIDIA announced that it is working with supply chain partners to expand production capacity for AI infrastructure (NVIDIA Newsroom, primary source). Wafer fabrication is not the focus. The expansion covers substrates, packaging, testing and optical engines. Substrate capacity will grow 2.5x over three years, and optical engine capacity 15x. NVIDIA also used its own money to prepay $1.1 billion to KYEC (King Yuan Electronics), a Taiwanese semiconductor testing company.

The announcement shows that the constraint on AI supply has moved. Being able to make GPU dies is not enough to ship racks. Each die needs a substrate to sit on and a package that combines it with HBM. It has to be tested before it ships, and racks need optical components to connect them. A shortage in any one of these means no finished product. NVIDIA is now stepping into those downstream steps itself.

What is being expanded

The announcement lists four steps.

• Substrates: Capacity for the substrates that carry GPUs and memory will grow 2.5x over three years.
• Packaging: More capacity for the step that combines dies and HBM into a single component.
• Testing: The step that checks whether finished components work correctly. NVIDIA prepaid $1.1 billion to KYEC.
• Optical engines: Capacity for components that convert electrical signals into light for transmission will grow 15x.

None of the four involves chip design or lithography. All of them turn chips into products that run in data centers.

Why NVIDIA put its own money into testing

Of the four, testing at KYEC is the only one where the announcement says NVIDIA prepaid with its own money. When a customer pays a supplier in advance, the supplier does not carry the risk of adding equipment, and the customer locks in capacity. $1.1 billion is a large sum to pay a single testing company.

What follows is this reporter's interpretation. Testing gets little attention among semiconductor steps, but it has become a heavier job for AI chips. The dies are large, HBM is stacked many layers high, and each component is expensive. Because a single defect is costly, testing takes more time and more equipment. Adding test equipment and facilities takes time. If testing backs up, shipments stop there, no matter how much the earlier steps produce. NVIDIA's choice of a prepayment suggests it did not want to leave this step to a scramble in the open market.

What the 15x for optical engines signals

Among the four figures, the 15x for optical engines stands out. Compared with the 2.5x for substrates, it points to a step that is ramping quickly from a small base.

This is also the reporter's interpretation, but the figure likely anticipates a shift from copper to optics for connections inside and between racks. The more tens of thousands of GPUs are linked to run as one system, the more the bandwidth and power of those links matter. The 15x target suggests NVIDIA expects interconnect components to be the next bottleneck.

The constraint moves from GPU counts to what surrounds them

Set alongside recent months of news, the trend is clear. Microsoft has said it will deploy Maia 300, its in-house inference chip, at gigawatt scale starting in 2027, and it has reportedly secured more than 300,000 units of TSMC production capacity (unconfirmed reports). The fight over the chips themselves has become a race to lock up foundry capacity.

Even with foundry capacity secured, products cannot ship if the downstream steps can't keep up. NVIDIA's announcement is a case of the largest buyer stepping into those steps itself. For substrate, testing and optical companies, visible demand and a buyer covering part of the equipment cost are what persuade them to add capacity.

There is another way to read it. By pushing suppliers to expand and even paying in advance, NVIDIA is betting heavily that demand will last. If AI capital spending slows, the added capacity will sit idle. The $1.1 billion prepayment also means NVIDIA has taken on part of that bet itself.

What to watch next

• What baseline year do the 2.5x for substrates and 15x for optical engines use, and which suppliers will deliver them?
• What are the terms of the $1.1 billion prepayment to KYEC (duration, pricing, volume)? Check them against KYEC's own disclosures.
• Is HBM supply growing in step with packaging capacity?
• Will Huang say more about supply in his GTC Berlin keynote (October 21)?