TSMC's Phoenix fab is now producing wafers for Nvidia's Blackwell AI chips on American soil, but every one of those silicon dies still ships back to Taiwan for the advanced packaging step — the chip stitching process that turns a bare die into a
TSMC's Phoenix fab is now producing Nvidia Blackwell wafers at volume on American soil. The harder step, the one that turns a bare die into a working GPU, is still happening 8,000 miles away in Taiwan (Tom's Hardware).
That step is called advanced packaging, and for Blackwell it has a specific name: CoWoS-L. It is the process TSMC uses to stitch two reticle-limited compute dies together with a silicon interposer and to wire them up to high-bandwidth memory stacks. Without CoWoS-L, a Blackwell chip is not a chip. It is two pieces of silicon with no way to talk to each other. Every Blackwell die that leaves TSMC's Fab 21 in Arizona today still flies back to Taiwan for this step (The Register, TweakTown, iConnect007).
This is not a rounding error. Nvidia has reportedly reserved the majority of TSMC's latest-generation CoWoS-L capacity (CNBC), which means the entire US AI chip pipeline, the one Nvidia says will deliver roughly $500 billion of US-built AI infrastructure over four years through partners including TSMC, Foxconn, Wistron, Corning, Coherent, and Amkor (Tom's Hardware), depends on a packaging line that exists in exactly one country. The 43-state US supplier network Nvidia has assembled is real. The wafer step is real. Foxconn's new Houston factory, which will produce GB300 tray modules, is real. The advanced packaging step is not.
The memory stack is the second half of the dependency. A Blackwell GPU also needs HBM, high-bandwidth memory, and no HBM is currently manufactured or packaged on US soil. The stacks come from Korea (SK hynix and Samsung) and, eventually, from Micron's planned Idaho ramp, which remains forward-looking rather than operational (Tom's Hardware). Even if the Arizona wafer and the Arizona packaging landed tomorrow, the memory would still arrive by ship from East Asia.
What about Intel? The company's "America 250" pitch leans on EMIB, Intel's competing advanced-packaging technology, and on Intel Foundry's domestic capacity. EMIB is real, but it is Intel's lane, not Nvidia's, and it does not absorb any Blackwell volume today (TechPowerUp). Intel's domestic advanced-packaging footprint is still ramping, not relieving the Nvidia bottleneck.
The window when this might change is 2027 to 2028. Amkor Technology is building an advanced packaging plant in Peoria, Arizona, intended to take on US-made CoWoS-style work, with production targeted for that two-year window. If Amkor hits its ramp, and if Micron's Idaho HBM reaches volume, the gap narrows and the "Made in America" framing starts to mean what it claims. Until then, the load-bearing step of every US-built Blackwell, the one that decides whether the AI buildout's most-cited number ($500 billion) gets delivered on schedule, is still running through a single Taiwan packaging line that the US has no domestic substitute for.
The next milestone to watch is Amkor's Peoria tool-in date. When the first CoWoS-S or CoWoS-L equivalent line moves from construction to qualification in Arizona, the story changes. Until that happens, the "Made in America" label on a Blackwell GPU describes where the wafer was grown, not where the chip was finished.