NVIDIA’s next AI bottleneck is being negotiated in substrates, packaging and testing

The factories that determine how many AI systems can be delivered do not all manufacture the main processor. NVIDIA’s supplier announcements on 21 October put package substrates, assembly, test capacity and optical components alongside advanced chips in the expansion plan. The company is trying to enlarge several connected production stages rather than solve a single shortage.[1]

That is a useful corrective to a common way of discussing AI infrastructure: treating a chip order as if it were a finished computing installation. Components still have to be interconnected, assembled and checked. Increasing one stage’s capacity faster than the next can move a queue without shortening the whole delivery process.

Three commitments, three different meanings

Amkor says a new $1.3 billion agreement supports expansion in South Korea, taking its multi-year agreements with NVIDIA across South Korea and Arizona to $3 billion. The supplier describes prepayments that align manufacturing capacity with NVIDIA’s growth plans. Those are supplier commitments, not a reported acquisition of Amkor or a statement that all the additional facilities are operating.[2]

Samsung Electro-Mechanics separately describes joint investment in dedicated ABF package-substrate production at Sejong and Busan. Its multi-year supply arrangement begins in 2026 and covers both computing chipsets and high-speed networking switches. The company’s account matters because it identifies the actual manufacturing locations and product categories behind the broader capacity headline.[3]

The percentages cannot be added together to derive a growth rate for finished AI systems. They refer to different activities and starting points. An increase of more than 300% means more than four times the starting level, whereas “2.5 times” means a 150% increase. Neither number, by itself, tells a customer how many additional working servers will be available.

Capacity finance is also supply-chain strategy

Amkor’s description of prepayments reveals the financial mechanism. A customer can support an expansion before the resulting production is ready. From an analytical perspective, that can help a supplier commit to specialised capacity while giving the customer a stronger claim on future output. The exact allocation of risk depends on the contracts, whose full commercial terms are not supplied in these announcements.[2]

The plans also join compute and networking demand. Samsung’s agreement covers substrates for both, underscoring that an accelerator’s usefulness depends on more than its own arithmetic performance. A cluster cannot be evaluated solely by counting processors if the connections and packaging required to use them arrive on a different schedule.[3]

For buyers, the central information is the staggered timetable. Capacity coming into service during 2027 or 2028 cannot be treated as available inventory today. Nor do these manufacturing commitments establish that a specific data centre has power, cooling or installation capacity. They address one layer of the buildout.

The commercial signal is therefore more precise than “AI demand is booming”. NVIDIA is committing support across multiple specialised suppliers, and those suppliers are publishing multi-year manufacturing plans. The next informative milestones will be facilities entering production, qualified output and delivery performance—not another multiplication of unrelated capacity percentages.

Report an error or suggest a correction