Nvidia reserves AI accelerator capacity through 2028

By Amara Osei

Nvidia announced multi-billion-dollar agreements with six suppliers on October 21 at GTC Berlin, reserving AI-accelerator capacity across substrates, advanced packaging, semiconductor testing and co-packaged optics through 2028.

The partners are Ibiden, Unimicron, Samsung Electro-Mechanics, SPIL, Amkor and KYEC. Expanded supply begins in 2027, while the commitments and production roadmaps run through 2028. The agreements reserve capacity and add redundancy around bottlenecks constraining current and future AI infrastructure.

Nvidia leads the AI chip market by a considerable margin, and its accelerators sit in a vast majority of global AI data centers, making secured component supply a competitive necessity.

Nvidia CEO Jensen Huang said: “We are in the midst of the largest infrastructure buildout in human history. Demand is extraordinary. Capacity is the constraint.”

He added: “Our supply chain partners are the world’s best manufacturers. I’m grateful they are expanding capacity around the globe to scale production and help us deliver the AI supercomputing systems powering this new industrial revolution.”

The result is a supply-chain hedge spanning 2027-2028, rather than a single capacity increase.

Why do substrates and optical engines need more capacity?

Nvidia plans to secure 2.5 times its dedicated substrate capacity over the next three years. It has signed agreements with Ibiden, Unimicron and Samsung Electro-Mechanics, three of the world’s five largest substrate manufacturers.

A substrate connects the accelerator’s silicon chips to its board. Only a handful of manufacturers make them, and rising component density now consumes more material and capacity in the X-Y and Z directions. Blackwell combines two reticle-sized GPU dies and eight HBM3E stacks on one package; Rubin and Rubin Ultra add more components.

Fubon Research projects Nvidia will supply around 12.4 million data-center AI GPUs in 2028, up from a projected 8.2 million in 2026, with one substrate layer per accelerator. Earlier this year, Nvidia agreed to provide a multibillion-dollar capacity-reservation advance for Ibiden’s Ono Plant in Japan, targeting three times the plant’s current individual capacity.

Samsung Electro-Mechanics will upgrade existing operations and build new plants in Busan and Sejong; Nvidia is projected to account for a significant share of its output by 2028. Unimicron will expand substrate production in Taiwan, with capacity arriving in the second quarter of 2027 and reaching full operation by mid-2028.

At speeds above 224 Gbps per lane, copper traces consume too much power even over short board distances. Nvidia therefore plans to move optical engines into accelerator packages through co-packaged optics. SPIL will expand optical-engine capacity by 1,400%, or 15 times, by mid-2027, then double it again by the end of 2028.

The practical result is more package capacity with less dependence on any single substrate or optics supplier.

How much packaging and testing capacity comes online?

The world’s largest semiconductor company, Nvidia, owns no fabrication plants. TSMC fabricates logic chips and binds them with memory on silicon interposers; Amkor then combines the module, substrate and protective casing into the finished accelerator package, which increasingly includes co-packaged photonics.

Nvidia and Amkor will lift interposer-wafer capacity by more than 300%—about four times current output—by the end of 2028 through advanced-packaging expansion in South Korea. Amkor values the agreement at $1.3 billion, taking Nvidia’s total investment across its Arizona and South Korean facilities to $3 billion.

KYEC, the world’s largest chip-testing company, handles nearly all Nvidia chip tests. Nvidia will double testing capacity from calendar 2026 to 2027 through expansions in the United States, Taiwan and Singapore.

On October 13, Nvidia agreed to pay KYEC $1.1 billion to reserve capacity at its new $1.4 billion US facility, where high-volume production is due in April 2027. The US project will build domestic testing capacity, while Singaporean high-volume production starts in February 2027.

These additions place contracted packaging and testing capacity in place before the next accelerator ramp.

What does the buildout mean for competitors?

Nvidia reported $279 billion in outstanding supply and manufacturing-capacity commitments by July 2026, according to its August 2026 disclosures. Together, those commitments and recent agreements cover semiconductor fabrication, HBM memory, substrates, advanced packaging, optical interconnects, final assembly and testing.

• AMD announced more than $10 billion in Taiwan investments on May 21, 2026, targeting advanced packaging and manufacturing partnerships. Lisa Su said on October 6 that AMD would increase those investments while working with TSMC, packaging, substrate and memory suppliers to raise 2027 production. Google and Amazon are also expanding their custom TPU and Trainium programs.
• Intel is expanding EMIB-T packaging after reports that Google reserved capacity for more than three million TPUs in 2028; Intel also advances its 18A and 14A processes.
• Samsung, with about 5.9% of the global foundry market by revenue in Q2 2026, is ramping 2nm production and preparing its Taylor, Texas, facility for volume output in 2027.
• Tesla, SpaceX and xAI’s Terafab aims eventually to manufacture one terawatt of annual AI-computing capacity. Intel joined in April, and Elon Musk confirmed plans to use Intel’s 14A process. Terafab targets experimental chips in 2027 and high-volume production in the third or fourth quarter of 2028.

Nvidia has contracted capacity through 2028; rivals now face longer lead times and heavier capital commitments before products can reach neoclouds and data centers.