Nvidia Commits Billions to Six Suppliers to Lock AI Chip Capacity Through 2028

Nvidia will spend billions of dollars across six supply chain partners to reserve manufacturing capacity from 2027 through 2028, the company announced on October 21 at GTC Berlin. The deals span substrates, advanced packaging, semiconductor testing, and optical engines. They target the bottlenecks now constraining AI infrastructure buildout.

"We are in the midst of the largest infrastructure buildout in human history. Demand is extraordinary. Capacity is the constraint," CEO Jensen Huang said. He called the partners "the world's best manufacturers" and credited them with expanding capacity globally. That is a supplier's promise, not a shipped product. Every roadmap here stretches years into the future.

What did Nvidia actually sign?

The agreements cover two categories: hardware interconnects and OSAT, outsourced assembly and test.

• Substrates. Nvidia signed deals with Ibiden, Unimicron, and Samsung Electro-Mechanics, three of the world's five largest substrate makers, and plans 2.5x dedicated substrate capacity over three years.
• Co-packaged optics. A deal with SPIL expands optical engine capacity 15x by mid-2027, with a further doubling planned by end of 2028.
• Advanced packaging. Amkor targets a 300% increase in interposer wafer capacity by end of 2028, backed by expansions in South Korea.
• Testing. Nvidia will pay KYEC $1.1 billion to reserve capacity, doubling testing capacity from 2026 to 2027.

Amkor confirmed its deal is worth $1.3 billion. That brings Nvidia's total investment in Amkor's Arizona and South Korea facilities to $3 billion.

Why substrates first?

The substrate connects the silicon above to the board below. Only a handful of companies make it. Each Blackwell package carries two reticle-sized GPU dies and eight HBM3E stacks on one substrate. Rubin and Rubin Ultra mount more. Each substrate now consumes more material even as unit demand climbs.

Fubon Research estimates Nvidia will ship about 12.4 million data-center AI GPUs in 2028, up from a projected 8.2 million in 2026. Those are analyst projections, not Nvidia guidance. The company is nonetheless building ahead of them.

Ibiden leads the substrate push. Nvidia signed a multibillion-dollar capacity reservation advance earlier this year for Ibiden's Ono Plant in Japan, targeting 3x the plant's current capacity. Samsung Electro-Mechanics adds capacity from existing operations plus upgrades and new plants in Busan and Sejong. Nvidia is projected to take a significant share of its output by 2028 — a projection, again, not a contract volume. Unimicron's Taiwan expansion starts supplying capacity in Q2 2027 and reaches full capacity by mid-2028.

What about optics and packaging?

Beyond 224 Gbps per lane, copper traces burn considerable power moving data even a few inches. The industry's answer is co-packaged optics, placing the optical engine inside the accelerator package. Nvidia has long flagged silicon photonics as the standard for future packages. The SPIL deal expands optical engine capacity 1,400% by mid-2027. Whether demand materializes on that schedule is a separate question.

Nvidia owns no fab. TSMC makes the logic and bonds it with memory on a silicon interposer. Amkor then assembles that module with the substrate and casing into the final package. The Amkor deal expands interposer wafer capacity more than 300% by end of 2028 through South Korean buildout.

After assembly, packages go to KYEC for extreme testing. KYEC already handles nearly all of Nvidia's chip testing. On October 13, Nvidia agreed to pay $1.1 billion to reserve capacity at KYEC's new $1.4 billion U.S. facility, where high-volume production starts in April 2027. KYEC's Singapore operations begin high-volume production in February 2027.

How large is the total commitment?

Per its August 2026 financial disclosures, Nvidia had accumulated $279 billion in outstanding supply and manufacturing-capacity commitments by July 2026. That figure covers component purchases and capacity reservations, particularly memory and chip production, and the newer deals add to it. Nvidia is now securing nearly every stage: fabrication, HBM, substrates, packaging, optical interconnect, assembly, and test.

Where do competitors stand?

The rest of the industry is not standing still.

• AMD announced over $10 billion in investments across Taiwan's semiconductor ecosystem on May 21, 2026, targeting packaging and manufacturing. CEO Lisa Su confirmed on October 6 that AMD will increase those investments, working with TSMC, packaging partners, substrate makers, and memory suppliers to raise output substantially in 2027.
• Google and Amazon are aggressively expanding custom TPU and Trainium accelerator programs.
• Intel is positioning its foundry and packaging as alternatives to constrained TSMC capacity. Reports indicate Google has reserved Intel EMIB-T packaging capacity for more than three million TPUs in 2028. Intel is also advancing its 18A and 14A processes.
• Samsung, holding roughly 5.9% of global foundry revenue in Q2 2026 per the source, is ramping 2nm technology and preparing its Taylor, Texas facility for volume production in 2027.
• Terafab, the Elon Musk joint project of Tesla, SpaceX, and xAI, targets one terawatt of annual AI computing capacity. Intel joined in April and will supply 14A technology. First experimental chips are slated for 2027, with high-volume production by Q3 or Q4 2028.

Compute does not come online overnight, and fabs are not built in a day. With nearly every link in the supply chain still constrained, heavy investment from the major players will continue, and capacity — not demand — will decide who ships accelerators into data centers through 2028.