Nvidia secures $1.5B in packaging from Amkor: the TSMC squeeze spreads

Context

Nvidia signs a $1.5 billion advanced packaging agreement with Amkor Technology, the second-largest OSAT (outsourced assembly and test) player in the world. The announcement follows a 10-year partnership between Amkor and TSMC to expand advanced packaging capacity in the United States, particularly at its Arizona site.

The Data

• Agreement amount: $1.5 billion (Tech in Asia, July 24, 2026).
• Amkor Technology: #2 worldwide in assembly/test behind ASE.
• Parallel Amkor TSMC partnership: ten years, expansion of advanced packaging in the United States, including Arizona.

Analysis

The fact that a hyperscaler client directly contracts the OSAT—rather than letting TSMC manage the back-end—speaks volumes about the compute bottleneck. TSMC's CoWoS has been a documented bottleneck in the AI supply chain since 2024. Nvidia secures its queue upstream of the queue by attaching a second assembly flow geographically aligned with its future US fabs.

Scenarios

• Short term (3-6 months): Other similar agreements, particularly from AMD and Broadcom, are likely.
• Medium term (12-18 months): US packaging capacity increases enough to absorb a significant portion of Blackwell/Blackwell Ultra deliveries without going through Taiwan.
• Risk of the opposite scenario: The bottleneck shifts to HBM memory (SK Hynix, Micron), which is not keeping up at the same pace.

Implications for the Professional

For a CTO or CFO building a 2027 compute plan, the announcement is clear: Nvidia's package delays become structurally more flexible in the United States. Budgetary corollary: The critical constraint shifts to HBM and to data center energy/land.

Signals to Monitor

• The next OSAT contracts (ASE, SPIL, PTI) signed directly by hyperscalers.
• Any mention of Arizona and commissioning dates in Amkor and TSMC guidance.
• The level of HBM tension (SK Hynix semi-annual results, August 2026).

Our Take

The TSMC bottleneck is not easing—it is being redistributed. Good news for Nvidia, bad news for those betting on compute delays being resolved by simply adding wafers. The next wall is already in sight.