Nvidia Expands AI Accelerator Supply Chain Capacity Through 2028

| 2 Min Read
Nvidia is reserving more substrate, packaging, optics and testing capacity to support rising accelerator production and reduce supply bottlenecks.

Nvidia’s latest supply agreements reach well beyond chip fabrication. The company is reserving more capacity for the substrates, optical engines, advanced packaging and testing that turn AI silicon into deployable accelerators. The commitments extend through 2028 and point to a practical constraint on growth: production depends on a chain of specialized suppliers, not just access to leading-edge chips.

Announced October 21 during GTC Berlin, the agreements involve Ibiden, Unimicron, Samsung Electro-Mechanics, SPIL, Amkor and KYEC. Nvidia’s plans include 2.5 times more dedicated substrate capacity over three years, a 15-fold expansion in optical-engine manufacturing by mid-2027, more than four times Amkor’s interposer wafer capacity by the end of 2028, and twice the semiconductor testing capacity by 2027. The company says it is building capacity and redundancy across its accelerator supply chain.

“We are in the midst of the largest infrastructure buildout in human history. Demand is extraordinary. Capacity is the constraint,” Nvidia CEO Jensen Huang said. That framing explains why Nvidia is making deals across several manufacturing stages at once: a shortage at any one of them can limit the number of complete systems it can ship.

Substrates are becoming a larger production constraint

A substrate sits between the silicon components in an accelerator package and the circuit board beneath it. It has to connect increasingly complex groups of processors and memory, and only a small number of manufacturers can produce these specialized layers at scale. The demands on substrates are growing in both surface area and complexity as chip designers bring more components into a single package. A substrate manufacturing roadmap details how that pressure is building as AI packages expand.

Nvidia’s Blackwell generation combines two reticle-sized GPU dies with eight HBM3E memory stacks on one package. Rubin and Rubin Ultra are expected to pack in still more components. Each larger, denser assembly places greater demands on the substrate underneath it, so rising accelerator shipments require more than a proportional increase in chip output. The limits of 2.5D packaging are part of the same manufacturing challenge.

Nvidia is spreading its substrate supply across three major manufacturers. It has agreements with Ibiden, Unimicron and Samsung Electro-Mechanics, which together cover three of the world’s five largest substrate makers. The company plans to secure 2.5 times its dedicated substrate capacity over the next three years, with Ibiden leading a significant part of the expansion.

Earlier this year, Nvidia agreed to provide Ibiden with a multibillion-dollar advance to support the company’s Ono Plant in Japan. The facility is targeting three times its current individual capacity. Samsung Electro-Mechanics will contribute capacity from its existing operations, along with upgrades and new plants in Busan and Sejong. Nvidia is expected to account for a significant share of the supplier’s output by 2028.

The Unimicron agreement adds another source of supply in Taiwan. Its expansion is expected to begin providing capacity in the second quarter of 2027 and reach full capacity by mid-2028. That timing gives Nvidia additional supply as it scales, while the broader supplier mix reduces reliance on any one manufacturer.

The demand outlook helps explain the scale of these plans. Fubon Research estimates Nvidia could supply about 12.4 million data-center AI GPUs in 2028, compared with a projected 8.2 million in 2026. Each accelerator requires a substrate. If shipments approach those estimates, material availability and factory output will have to rise alongside chip production.

Optical links and packaging add capacity upstream

As data rates climb beyond 224 Gbps per lane, copper traces can consume substantial power moving data across even short distances. Co-packaged optics place optical engines close to the processor, offering a way to manage higher-speed connections within AI systems. Nvidia has described co-packaged optics plans as part of its future accelerator packaging approach, and the wider silicon photonics market is expected to grow alongside demand for faster data-center connections.

To prepare for that shift, Nvidia has signed an agreement with Taiwan-based SPIL to expand optical-engine manufacturing capacity 15-fold by mid-2027, with a further doubling planned by the end of 2028. The figure shows how far production needs to scale if optical engines become a standard part of high-volume accelerator packages.

Packaging is another point where limited capacity can hold up finished hardware. Nvidia does not operate its own fabrication plants. TSMC makes the logic chips and connects them with memory on a silicon interposer; companies such as Amkor then assemble the integrated module with its substrate and protective casing. As packages grow more complex, these final manufacturing stages need to expand with them.

Nvidia and Amkor are targeting an increase of more than 300% in interposer wafer capacity by the end of 2028, supported by expanded advanced packaging operations in South Korea. Amkor says the agreement is worth $1.3 billion. It brings Nvidia’s total investment in Amkor’s Arizona and South Korea facilities to $3 billion, including the Arizona packaging campus.

Testing capacity is also part of the plan

Completed accelerator packages still need extensive testing before they can be deployed. KYEC, or King Yuan Electronics Co., currently handles nearly all of Nvidia’s chip testing. Nvidia expects testing capacity to double between calendar years 2026 and 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 a new $1.4 billion U.S. facility. High-volume production there is scheduled to begin in April 2027. KYEC is also expanding in Singapore, where it expects high-volume production to start in February 2027. The agreement ties Nvidia’s capacity plans to specific sites and production schedules, while adding testing options across several regions.

Supply commitments are becoming a competitive tool

Nvidia has already made large commitments across its manufacturing network. In financial disclosures from August 2026, it reported $279 billion in outstanding supply and manufacturing-capacity commitments as of July 2026. Those commitments include future component purchases and reserved production capacity, particularly for memory and chip manufacturing. The newer deals extend that approach across substrates, optical interconnects, packaging, assembly and testing.

Competitors are making their own moves to secure production. AMD announced more than $10 billion in investments across Taiwan’s semiconductor ecosystem on May 21, 2026, targeting advanced packaging and manufacturing partnerships. CEO Lisa Su said on October 6 that the company planned to increase those investments. AMD is working with TSMC, packaging firms, substrate suppliers and memory manufacturers to raise output in 2027.

Google and Amazon are expanding their in-house accelerator programs, including Google’s TPUs and Amazon’s Trainium. Their plans add to demand for the same broad set of manufacturing resources, even as they offer customers alternatives to Nvidia hardware.

Intel is positioning its foundry and advanced packaging operations as another route for companies facing limited capacity at TSMC. Its EMIB-T packaging technology is attracting attention, with reports that Google has reserved Intel capacity to package more than three million TPUs in 2028. Intel is also developing its 18A and 14A manufacturing processes to attract outside customers.

Samsung is expanding its foundry business alongside its role as a substrate supplier. It held about 5.9% of global foundry revenue in the second quarter of 2026, is ramping up 2nm manufacturing and is preparing its Taylor, Texas, facility for volume production in 2027. The company’s foundry roadmap reflects an effort to grow its manufacturing operations and narrow the gap with TSMC.

A much larger, longer-term plan comes from Terafab, a semiconductor manufacturing project involving Tesla, SpaceX and xAI. Its stated ambition is to produce the equivalent of one terawatt of AI computing capacity each year. Intel joined the project in April, and Elon Musk later said it would use Intel’s 14A process. Terafab aims to produce experimental chips in 2027 and reach high-volume production in the third or fourth quarter of 2028, according to reported project targets.

These plans share a basic challenge: adding compute takes time, and capacity must be built across multiple stages before finished systems can ship. Nvidia’s new agreements show how a market leader can turn demand forecasts into manufacturing reservations years ahead. The next measure of success will be whether suppliers bring that capacity online on schedule as accelerator designs, packaging needs and data-center demand continue to grow.

Source: Etiido Uko · www.tomshardware.com

Comments

Please sign in to comment.