Table of Contents
Six suppliers take Nvidia’s prepayments, but the 2028 ramp stays unconfirmed
In sum – what to know:
- Six partners, four chokepoints – Nvidia’s agreements with Amkor, IBIDEN, KYEC, Samsung Electro-Mechanics, SPIL and Unimicron span substrates, advanced packaging, testing and co-packaged optics.
- Front-loaded commitments – Supply and capacity commitments on Nvidia’s books reached $279 billion as of July 26, up from $119 billion one quarter earlier.
- Where the risk sits – Most targets land in 2027 or 2028, and Nvidia’s 10-Q notes some agreements can be cancelled or rescheduled before firm orders are placed.
Nvidia has announced capacity-expansion agreements with six manufacturing partners, spanning package substrates, advanced packaging, semiconductor testing, and optical engines for co-packaged optics. The partners are Amkor Technology, IBIDEN, King Yuan Electronics (KYEC), Samsung Electro-Mechanics, SPIL, and Unimicron. Together, the deals underpin a supply-chain buildout that Nvidia frames as both a volume play and a geographic-diversification strategy — resiliency and redundancy, in the company’s language. Nvidia says it plans to raise substrate capacity for its products by 2.5x over three years, with “more underway.” “Demand is extraordinary,” CEO Jensen Huang said at the keynote. “Capacity is the constraint.”
Supply and capacity commitments on NVIDIA’s books reached $279 billion as of July 26, 2026, up from $119 billion just one quarter earlier. That figure covers memory, manufacturing facilities, and reservations tied to current and future chip architectures. The company holds roughly half of TSMC’s total CoWoS advanced-packaging capacity, and TSMC said on October 15 that demand for its chips still outpaces what it can supply. CoWoS bottlenecks are expected to persist through at least 2027.
Most of the targets Nvidia disclosed are multipliers off undisclosed baselines — 2.5x substrate capacity, greater-than-300% interposer wafer growth, 15x optical engine output — without absolute unit volumes or fully disclosed dollar amounts. Nvidia’s own release classifies the expansion figures, timelines, and expected benefits as forward-looking statements. The financial commitments are real. Whether the capacity materializes on schedule is a separate question.
Substrate expansion leads the buildout
Substrates are at the center of Nvidia’s expansion push. Three of the six partnerships target package substrates directly, and together they’re meant to deliver on the stated 2.5x capacity increase over three years.
The largest commitment involves IBIDEN, the Japanese manufacturer that supplies substrates for more than 80% of cutting-edge GPUs. Under an agreement signed earlier in 2026, Nvidia is providing a multibillion-dollar capacity reservation advance for IBIDEN’s Ono Plant in Gifu Prefecture, targeting 3x the plant’s current capacity with additional ramp beginning in Q1 2028. Japanese outlet Bunshun reported the advance at roughly ¥280 billion (about $1.8 billion) — a figure neither company has confirmed. IBIDEN CEO Koji Kawashima has acknowledged that a “substantial portion” of the company’s ¥500 billion three-year capex plan is funded by customer prepayments, a practice IBIDEN says it adopted only in the last few years.
IBIDEN has been here before, though. Capacity that came online in 2007 saw utilization slump after the financial crisis, and an Intel-oriented facility announced in 2021 and completed in FY2023 ran near-unused when Intel’s business deteriorated. The current ¥500 billion bet assumes AI-server demand keeps expanding for three to five years. That may prove right, but the company’s own history shows how quickly capacity investments can go sideways.
Samsung Electro-Mechanics is the newest addition to Nvidia’s substrate supply chain, bringing immediate capacity from existing South Korean operations. The multi-year ABF substrate agreement runs from 2026, with deliveries already underway and volumes expected to grow at a double-digit annual rate. Samsung’s Sejong and Busan lines will supply substrates for both Nvidia compute chipsets and high-speed networking switches.
Unimicron rounds out the substrate trio. The Taiwanese manufacturer will expand KF substrate production under a capacity reservation agreement, with capacity expected online in Q2 2027 and full production by mid-2028. No dollar figure has been disclosed. Goldman Sachs estimates that ABF substrate supply gaps could reach about 42% by 2028, a projection that suggests the shortage won’t close even with these investments in place.
All of the current spending targets ABF organic substrates, but the industry could shift toward glass-core substrates and TSMC’s CoPoS technology around 2028–2029. If that transition accelerates, some of the organic-substrate equipment being built now risks going underutilized. IBIDEN is reportedly a TSMC CoPoS partner, which partly hedges against that scenario. It doesn’t eliminate it.
Packaging, testing and optics capacity
Beyond substrates, Nvidia is scaling advanced packaging, testing, and optical-engine production — three areas where capacity is arguably just as constrained.
Amkor, the world’s largest U.S.-headquartered outsourced semiconductor assembly and test company, is expanding its partnership with Nvidia in South Korea, targeting a greater-than-300% increase in interposer wafer capacity by the end of 2028. That’s specifically interposer wafer capacity, not Amkor’s total advanced packaging output. Amkor said a new $1.3 billion agreement announced at GTC Berlin brings its multi-year Nvidia deals across Arizona and South Korea to a combined $3 billion. Amkor’s Arizona campus, where it broke ground in 2025, is already moving to Phase 2, which raises total planned investment to roughly $12 billion. Shares rose about 4.9% on the announcement, though the stock remains well below its moving averages and Amkor reports Q3 earnings on October 26.
Testing is a less visible but increasingly critical bottleneck. Next-generation AI chips require over 70% longer test times than prior models, so the same floor space produces fewer finished parts per day. Nvidia says it expects total testing capacity to double from 2026 to 2027, supported by buildouts in the U.S., Taiwan, and Singapore, with Singapore expanded testing operations expected to begin high-volume production in February 2027. The company is providing KYEC a $1.1 billion capacity reservation advance to support new U.S. testing facilities through its subsidiary AeternZ, with high-volume production targeted for April 2027. Fort Worth, Texas, has reportedly emerged as the likely site, though KYEC hasn’t formally confirmed the location. KYEC’s own filing, however, disclosed a prepayment cap of up to NT$45 billion — roughly $1.4 billion. Neither party has reconciled the gap with NVIDIA’s stated $1.1 billion figure. KYEC’s general manager has noted that U.S. construction timelines have stretched from under 12 months to at least 24, and surging demand may require two or three new plants a year.
SPIL, a subsidiary of ASE Technology Holding, handles the optical side of the expansion. The company started low-volume co-packaged optics production in Q3 2026 and is targeting a 15x increase in optical engine output by mid-2027, with another doubling by end-2028. Co-packaged optics, or CPO, places optical components physically close to switch silicon, shortening electrical paths and reducing power loss. NVIDIA has said its next-generation rack-scale platforms will rely on the technology. The ramp targets are aggressive, and CPO production is still in its early stages, where yield challenges in high-precision optical assembly could slow things down.
Risks behind the $279bn commitment
The $279 billion in supply commitments is a striking figure, but it’s forward-looking by design. The gap between financial commitments and delivered capacity is where the real risk lives.
Most of the production targets announced sit in 2027 or 2028. Equipment needs to be installed, processes qualified, and customers certified before any of this yields finished chips. Nvidia’s 10-Q notes that some agreements can be cancelled or rescheduled before firm orders are placed, sometimes at additional cost. If demand diverges from projections, the company may not be able to reduce its commitments in time. There’s precedent for that exact scenario. Nvidia took a $4.5 billion charge in the April 2025 quarter for excess H20 inventory and purchase obligations, followed by a further $0.4 billion charge on excess H200 inventory in the first half of fiscal 2027.
Free cash flow fell to roughly $21.3 billion in the July quarter from $48.6 billion the quarter before, driven partly by a surge in accounts receivable — days sales outstanding rose to 60 from 45. Nvidia also took in $24.9 billion of net proceeds from new debt during the same period, taking senior notes outstanding to $33.5 billion, while repurchasing $19.7 billion of stock. None of that signals distress on its own. But rising receivables, growing debt, and massive supplier prepayments, all moving in the same direction at the same time, have drawn scrutiny.
The sharpest criticism — made by analysts and outside observers, not by Nvidia — frames the company’s prepayments as a modern echo of the vendor-financing arrangements that fueled telecom buildouts in the Lucent and Nortel era, and that collapsed when demand didn’t materialize. Management has pushed back on that framing rather than adopting it. Addressing investor concern about Nvidia’s financial support for its customers on the Q2 FY2027 call in August, CFO Colette Kress said: “We recognize the scale of this support, and we know some will call this circular financing. We see it differently.” The core question is whether Nvidia’s money creates supplier capacity that feeds demand back toward Nvidia, and who bears the risk if AI capital spending slows. Revenue concentration makes that question more pointed. One direct customer accounted for 16% of Q2 revenue.
Execution risk is already visible in places. Nvidia’s Kyber rack design was reportedly pushed from 2027 to 2028 over a circuit-board manufacturing issue, a reminder that upstream component problems can shift Nvidia’s own roadmap regardless of how much capacity it has locked in.
