NVIDIA's $1.1 Billion Test Prepayment Puts the AI Buildout on Its Suppliers' Balance Sheets

NVIDIA stopped waiting for its back-end suppliers to finance the AI buildout. On October 21, 2026 it announced six capacity agreements spanning substrates, advanced packaging, chip test and co-packaged optics, including two explicit cash advances: $1.1 billion to King Yuan Electronics to reserve U.S. test capacity, and a multibillion-dollar advance to fund IBIDEN's Ono Plant in Japan. The King Yuan line targets high-volume production in April 2027, with expanded Singapore test operations due to start high-volume output in February 2027.

The advance to the test line is the small, visible end of a much larger shift. NVIDIA's supply and capacity commitments rose from $119 billion at the end of April — of which $95 billion was due to be paid in the remainder of fiscal 2027 — to $279 billion at the end of July, and the increase was reported as driven mainly by memory procurement. That is what makes the test advance interesting: the same structure quietly funding memory is now reaching the back end of the line, where installation timing rather than design decides how fast chips ship.

Balancing that is a cost NVIDIA has already shown it can absorb. A $0.4 billion charge tied to H200 excess inventory and purchase obligations landed in the first half of fiscal 2027, and $985 million of inventory provisions in the June quarter already show how the structure fails when committed product does not sell. Deferred revenue on NVIDIA's balance sheet rose from $3.1 billion at the end of April to $9.6 billion at the end of July, though the filings do not break out how much of that relates to capacity prepayments.

King Yuan Electronics is the clearest case, and it is not a healthy one. The Taipei-listed test house raised its 2026 capital budget to NT$50 billion from NT$39.4 billion in May, weighted toward high-power burn-in and advanced packaging test, and on a trailing basis it spends 74% of revenue on capacity: NT$29.98 billion of capital expenditure against NT$16.75 billion of operating cash flow, leaving free cash flow at minus NT$13.2 billion and NT$16.5 billion of new net debt issued over the period. A capital budget that exceeds a full year of revenue is what customer prepayments look like from the inside.

The pattern repeats elsewhere at different intensities. ASE Technology Holding, whose wholly owned packaging unit Siliconware Precision Industries NVIDIA named for the optical-engine buildout, runs trailing capital spending at 28.8% of revenue, with trailing free cash flow to the firm of minus NT$28.3 billion and net debt at 1.33x EBITDA; the group said in June 2026 it was expanding capacity to support AI demand. Samsung Electro-Mechanics spent ₩1.62 trillion against ₩1.83 trillion of operating cash flow over the trailing twelve months, and IBIDEN reinvested essentially all of its fiscal 2026 operating cash flow into capacity while still holding net cash.

Every tester, handler and prober bought for those lines is a separate revenue stream, and the equipment makers' order books confirm the capacity buildout is real rather than announced. Advantest raised its fiscal 2027 forecast on July 29, 2026 to sales of ¥1,714 billion from ¥1,420 billion and operating income of ¥846 billion, said test demand for inference AI semiconductors is now expected to significantly exceed its own April projections, and is expanding production capacity ahead of schedule. Teradyne's semiconductor test revenue of $1,122 million in the June quarter more than doubled year over year, lifting group revenue 104% to $1,329 million at a 59.8% gross margin.

The demand filling NVIDIA's commitment book comes from a very small set of buyers with enormous capital budgets. Microsoft's fiscal 2026 capital spending reached $115.9 billion, up from $64.6 billion a year earlier; Amazon spent $173.0 billion over the trailing twelve months to June 2026, more than its $161.4 billion of operating cash flow, pushing trailing free cash flow to minus $11.6 billion. Neither can fund that pace indefinitely from current cash flow, which is why Microsoft and Amazon are the ultimate financiers of the AI buildout and the vendors have become the intermediate lenders.

The geography carries the same message. At GTC Berlin, NVIDIA said it was committing nearly $10 billion to the European technology ecosystem, while the announced buildout places test capacity in the United States and Singapore, advanced packaging in Korea and Arizona, and substrate capacity in Japan. Capacity is no longer only a cost line — it is industrial policy, and the suppliers who own it gain bargaining weight that chip designers do not have.

This is where the structure is most fragile. A prepayment is a promise to buy, and if the buildout slows, the obligation does not shrink with the slowdown: NVIDIA's own first-half fiscal 2027 results include a $0.4 billion charge for H200 excess inventory and purchase obligations and $985 million of inventory provisions in the June quarter. Apply the same treatment to capacity commitments and the vendors who absorbed the financing absorb the write-down, on plants built for a single customer.

Fact: NVIDIA has committed $279 billion of supply and capacity, prepaid $1.1 billion to reserve U.S. test capacity and told six suppliers to expand. Fact: those suppliers, not NVIDIA, are carrying the fixed assets, with trailing capital spending at 25% to 74% of revenue across the test, packaging and substrate names. Inference: the AI buildout has quietly become a supplier-financed infrastructure project, with NVIDIA as anchor tenant and funder, which relocates the durable profit from the chip designer to the owners of scarce capacity.

Speculation, flagged as such: if commitments keep compounding while test pricing holds, the owners of test and substrate capacity earn a financing spread that did not exist in the previous cycle — customer cash funds assets that keep earning for years. The counter-case is equally concrete: prepayments are contractual obligations against future services, NVIDIA's own results show what happens when committed product does not sell, and a vendor spending 74% of revenue on capacity has no cheap exit if the order book stalls.