Nvidia và Intel khoe chuỗi cung ứng chip Mỹ, nhưng vẫn có khoảng trống

Nvidia shouted proudly in a recent blog post that its network of American manufacturing partners and suppliers now spans 43 states, that TSMC's Phoenix plant is producing Blackwell wafers at volume, and that it plans to produce up to $500 billion of AI infrastructure in the U.S. over four years with partners including TSMC, Foxconn, Wistron, Corning, Coherent, and Amkor. Intel has made its own case in an America 250 post presenting end-to-end U.S. capabilities across design, manufacturing, and advanced packaging. Both accounts hold up at the wafer stage but omit the same downstream step: every Blackwell die that leaves TSMC's Arizona fab still crosses the Pacific to be packaged, no HBM is manufactured or packaged on U.S. soil, and the facilities intended to close those gaps won’t start production until 2028 at the earliest. Lofty projects Foxconn is building a Houston factory to produce GB300 tray modules for Nvidia, and Wistron will assemble and test Nvidia AI systems at a new facility in Fort Worth, Texas. Coherent broke ground in June on an expanded Sherman, Texas, plant that the company describes as the first volume-production 6-inch indium phosphide fab, supplying the lasers and optical components that link AI systems together. Corning is adding more than 3,000 jobs across optical manufacturing sites in North Carolina and Texas. The post also cites an estimate from consultancy Public First that Nvidia-driven AI demand will contribute $485 billion to U.S. GDP in 2026 and support over 100,000 jobs. “AI is driving a once-in-a-generation opportunity to reinvigorate American manufacturing and supply chains,” said Nvidia’s Jensen Huang in the post. Meanwhile, Intel's post lists R&D and manufacturing across Oregon, Arizona, New Mexico, and California, describes Ohio as “a planned site,” and devotes most of its length to workforce programs, K-12 AI education, and the company’s America250 partnership. Neither post addresses where the most advanced AI processors are actually assembled into finished chips. The Pacific round-trip Nvidia and TSMC produced the first Blackwell wafer at Fab 21 near Phoenix last October, and the site has since moved to volume output of Blackwell silicon on TSMC's 4NP node, the custom 4nm-class process built for Nvidia. On the other side of the Phoenix metro area, Intel's Fab 52 became fully operational in the same month as the first high-volume home of Intel 18A, and Naga Chandrasekaran, Intel's chief technology and operations officer, told CNBC in December that the fab is capable of more than 10,000 18A wafer starts per week. Panther Lake reached broad availability in January, Clearwater Forest is due in the first half of this year, and 18A yields are expected to reach industry-standard levels in early 2027, which I covered in my examination of Intel's fab roadmap . This ultimately means that leading-edge logic wafers are now being fabbed in the U.S. by two companies on two competing nodes. That’s a genuine change and, by any measure, a monolithic achievement when compared to the start of the decade, and neither company overstates that in their corporate blogs. However, a Blackwell data center GPU pairs two reticle-sized compute dies with eight stacks of HBM3e on a silicon interposer using TSMC's CoWoS-L packaging, and all of TSMC's CoWoS capacity is located in Taiwan. TSMC’s U.S. facilities currently send 100% of their chips to Taiwan for packaging, including wafers fabbed in Phoenix. A Blackwell die fabbed in Arizona therefore travels roughly 7,000 miles to be diced, stacked, and mounted, then travels onward through system assembly before any of it returns to a U.S. data center. As for HBM, every stack in production today comes out of SK hynix and Samsung facilities in South Korea or Micron's fabs in Taiwan and Japan, and the ABF substrates beneath the interposer are similarly concentrated in Japan and Taiwan. No U.S. facility currently manufactures or packages HBM. The one company running advanced…