Deb Shoquist, Executive Vice President of Operations at NVIDIA, recently offered a rare, systematic look at the global manufacturing network behind the AI chip giant. The supply chain she oversees spans TSMC, Foxconn, Wistron, SPIL, KYEC, Unimicron, Amphenol, Coherent and Lumentum, among others, and her core mission is turning design blueprints into high-performance products that can be delivered at volume.

In Shoquist’s telling, the job is more like that of a “middleman”: finding partners whose capacity ramp cadence matches NVIDIA’s, confirming their technical precision is sufficient to handle highly complex designs, and requiring that suppliers have enough financial strength to invest ahead of expansion. She calls these three conditions indispensable pillars—if a company has only two of the three, NVIDIA will choose to inject capital or provide support to help it close the gap.

These deep ties did not begin yesterday. TSMC’s partnership with NVIDIA dates back to 1998, and SPIL and KYEC have performed chip packaging and testing for NVIDIA for more than two decades. Shoquist meets with major suppliers every quarter and spends six weeks a year visiting partners in Taiwan, mainland China, South Korea, Vietnam and the United States, with discussions often moving from the conference room to the factory floor. She recalled that one Taiwanese supplier called her at 3 a.m. with a solution, and that workers at a Vietnamese plant had only one request: more tools so they could do more work.

Rebuilding the supply chain from chip company to systems company

As NVIDIA shifted from selling chips to building data center systems and platforms, supply chain complexity rose markedly. Shoquist acknowledged that when the company was a chip vendor it was smaller, and the semiconductor industry itself already had scale—all it had to do was pick the best technology mix for a conceived product. As a data center systems company, it needs far more partners, and the supply base for many links in the chain is still maturing.

Citing optical components as an example, she noted that partners had the technology and the willingness but had not yet built the factory scale and capital needed to meet NVIDIA’s required volumes—which is precisely why NVIDIA entered investment partnerships with Coherent and Lumentum. On new products such as co-packaged optics (CPO), NVIDIA is building an entirely new supply chain together with a group of newcomers to optics that nonetheless have large-scale production capability.

Amphenol’s cable assembly is a typical example of this collaboration. The assembly contains 1,500 small cable branches, each of which must be precisely assembled to keep the defect rate below one per million units. Amphenol built over a hundred automated production lines in Vietnam to match the speed and scale NVIDIA requires. Shoquist said it is the most complex interconnect assembly Amphenol has ever manufactured, and the line is now in full production.

Weekly chip allocation, delivery cycles compressible to two weeks

NVIDIA’s logic for prioritizing resources is fairly direct: identify the most critical components first, then organize production around them. Shoquist said wafers are “as indispensable to the company as air,” and the most critical component is TSMC’s chips. Each week, after chips are packaged, the team decides which factory gets how many; if a plant’s output that week is low or its holding period falls short of expectations, it will not receive more chips that week.

Traceability thus becomes a threshold for suppliers. From the identification number on each die to the customer invoice, NVIDIA tracks every chip, and companies hoping to become its suppliers must support this process. When needed on either the supply or demand side, the team can shorten the cycle from chips leaving TSMC to systems shipping from Foxconn and Wistron from five weeks to two.

Shoquist describes the supply chain as a large circulatory system with pathways everywhere, and one must be able to push any link that needs accelerating.

A mass-production scare and “swarm” operations

In the interview she also revisited an instance when a plan fell short. NVIDIA was preparing to mass-produce compute boards with Foxconn and Wistron, with the final step being testing by automated handling equipment—a brand-new method for the company, aimed at meeting the high-throughput demands of complex products. But during the initial capacity ramp, actual throughput came in well below plan.

The manufacturing team’s operations director called to say that, to safeguard the product and shipments, the entire process needed to be paused and manual test loading restored. The manufacturer promptly deployed as many as several hundred operators to support the effort, and production continued—but it still could not fundamentally meet the throughput requirements for full mass production. Shoquist said the root causes involved multiple factors, including test tray assembly, tray design, automation hardware design, control software and product yield, and were difficult to untangle in a short time.

NVIDIA then launched an internal cluster operation code-named Swarm, rapidly deploying teams from design engineering, operations engineering, product and test engineering, and Foxconn’s automation engineering onto the line to complete a redesign and restore production before scaling. She summed up the approach as “the mission is the highest directive,” with teams unconstrained by organizational boundaries.

Implications for the industry chain

The disclosure sketches one reality of the AI infrastructure race: what determines delivery capability is no longer chip design alone, but the ability to orchestrate and deploy capital across dispersed links such as packaging, testing, interconnect and optics. By investing to close suppliers’ capacity gaps, trading chip allocation authority for line priority, and controlling cadence through end-to-end traceability, NVIDIA has effectively made supply chain management part of its product competitiveness.

For upstream manufacturers, entering NVIDIA’s system means access to expansion capital and long-term orders, but also means accepting stricter yield standards, automation investment and data transparency requirements. For investors watching AI compute supply, Shoquist’s remarks offer a window: at a stage when demand persistently exceeds supply, the bottleneck is often not in the most closely watched wafer fabrication link, but in easily overlooked supporting areas such as packaging and testing, cable interconnect and optical devices.

Shoquist recently received the 2026 Dr. Lisa Su Woman of Innovation Award. At the end of the interview she said operations is a place where people can fully apply their talents, and that the team continuously designs and implements supply chain changes to meet new product needs: “Between our goals and our success, the only variable is our own creativity.”