Amkor's $7 Billion Arizona Bet Is Sold Out Before It Builds. That Is Also the Risk.

The standard description of Amkor's new campus outside Phoenix is that the company is expanding to meet AI demand. That is true, and it is backwards. Amkor is not building capacity on the hope that the customers arrive; the customers have already committed to paying for it before a single part is produced. The question the stock forces is a different one: whether a structurally thin-margin packaging business can carry $7 billion of construction through the years when that committed demand has not yet turned into paying, profitable utilization.

A factory sold before it is built

Amkor broke ground in Peoria, Arizona in October 2025 on an advanced packaging and test campus it expects to be the first high-volume outsourced assembly and test (OSAT) facility in the United States. The investment was expanded during the planning to $7 billion across two phases — more than $5 billion above the original figure — for over 750,000 square feet of cleanroom space, with the first building slated for completion in mid-2027 and production in early 2028, supported by up to $400 million in proposed federal CHIPS funding and an advanced manufacturing investment tax credit.

What makes the campus legible as committed capacity rather than speculation is who is standing behind it. On the day of the groundbreaking, both Apple and NVIDIA were named as customers — Apple's packaging and testing of the Apple silicon built at TSMC's Arizona fab, NVIDIA's onshoring of its AI stack. The follow-through came in July 2026: a multi-year NVIDIA partnership carrying a $1.5 billion prepayment for advanced packaging and test, plus a ten-year advanced packaging agreement with TSMC. Management said Phase 1 of the Arizona facility is fully committed.

This is the memory-supply playbook run at the assembly level. The prepayment and the long-term agreements are the analogue of customer-committed foundry capacity: Amkor does not fund the build until the demand is contracted, and the customer carries part of the capital risk. The "expansion to meet demand" of the headlines is really pre-sold supply.

The two OSAT markets

The catch is what these economics look like once the demand is booked. Amkor spans two businesses. Advanced products — 2.5D, high-density fan-out, system-in-package sold into computing and AI — generated 82% of second-quarter revenue. The mainstream legacy packaging business is the commoditized remainder, and it exerts mechanical drag on the blended margin.

Even in the strongest quarter in company history, that blend was modest. Revenue rose 26% year over year to $1.90 billion, gross margin reached 16.8% and operating margin 10.5%, and Amkor beat the consensus on bottom line by nearly half. A packaging company's ceiling is simply below a foundry's or an equipment maker's, whatever the demand.

That is the structural tension the Arizona bet inherits. Financing a $7 billion build at OSAT gross margins means the payout depends on utilization arriving close to plan, because there is very little profit buffer to absorb a slow ramp. Amkor itself flags the consequence: the Arizona expansion is expected to create margin headwinds in 2027 and 2028 from underutilization during the ramp. Phase 1 being fully committed lowers the demand risk; it does nothing for the cost risk.

Funding the build against a negative-FCF clock

The financial identity of this trade is what bears scrutiny. Amkor guided 2026 capital expenditure to $2.5–3.0 billion, about double last year, with roughly two-thirds directed at facility expansion including Arizona. Trailing free cash flow is already negative — capex has outpaced operating cash flow for the trailing year — and the company recently raised $1.15 billion in long-term debt to fund the build, ending June with cash and short-term investments of $2.5 billion against $2.5 billion of debt.

So the margin story and the cash story point in opposite directions. Utilization across the network climbed into the 70% range and management called growth in computing structural, and that is the bull case. But a good quarter still leaves the company investing beyond its operating cash flow into a fixed asset that produces nothing until 2028, on top of a separate disruption: migrating system-in-package production from Korea to Vietnam, which management says will muddy results into the first half of 2027.

The market's reaction has already split the difference. Amkor beat Q2 estimates decisively and still fell 6.5% on the softer third-quarter outlook and the migration noise. The stock, up roughly 96% over the past year from a low near $26, trades at about $50 today — down by nearly half from its 52-week high and at a market valuation near $13 billion.

The key issue is not demand, which is contracted. It is whether the committed capacity converts into paid utilization while Amkor carries roughly $2.5–3 billion a year of construction cost at OSAT margins, through 2027 and into the 2028 ramp. NVIDIA's prepayment begins to be returned as delivered service over the coming years, and that payment stream — plus discipline on the Arizona spend target and the 2027 completion date — is the variable that determines whether the financing math holds. Phase 1 being fully booked removes the excuse for a weak outcome; it does not remove the year of negative free cash flow and margin pressure that sits between now and production. A well-run company can still lose money making the transition, and Amkor is not exempt from that arithmetic.