Chip supply fragility

The machines that train artificial intelligence are built from parts that cross the Pacific more than once, and this week two separate stories showed how exposed that route still is. Super Micro Computer opened a board-led investigation after U.S. prosecutors charged three people tied to the company in an alleged scheme to divert Nvidia server systems to China. (theregister.com) (justice.gov) The indictment unsealed on March 19, 2026 named Yih-Shyan “Wally” Liaw, Ruei-Tsang “Steven” Chang, and Ting-Wei “Willy” Sun, and said they used false documents, dummy servers, and transshipment routes to hide the real destination of restricted artificial intelligence hardware. The U.S. Department of Justice said the alleged scheme involved high-performance servers assembled in the United States and sold to buyers in China in violation of export-control law. (justice.gov) Supermicro said it was told about the alleged misconduct on March 19 and said the company itself is not accused of taking part. The internal probe is being led by independent directors Scott Angel and Tally Liu, with help from Munger, Tolles & Olson, AlixPartners, and auditor BDO USA. (theregister.com) That case is about finished servers, not loose chips. A graphics processing unit from Nvidia only becomes a usable artificial-intelligence server after it is mounted, wired, cooled, tested, and shipped inside a much larger box built by companies like Supermicro. (justice.gov) (theregister.com) The second weak point sits earlier in the chain, at a step called advanced packaging. That is the stage where separate pieces of silicon and stacks of memory are bound into one finished part, like turning an engine, transmission, and frame into a car that can actually drive. (latercera.com) (cnbc.com) Taiwan Semiconductor Manufacturing Company now makes advanced chips in Phoenix, where high-volume production on its N4 process started in the fourth quarter of 2024. But chips from Arizona still have to go to Taiwan for final assembly because the most advanced packaging capacity remains concentrated in Asia. (tsmc.com) (latercera.com) (cnbc.com) That packaging step became critical when artificial-intelligence chips stopped being one slab of silicon and started being clusters of compute chips plus high-bandwidth memory in one package. Taiwan Semiconductor Manufacturing Company’s main method for this is called Chip on Wafer on Substrate, and CNBC reported that demand for it is growing at an 80% compound annual rate. (latercera.com) (cnbc.com) Nvidia has reserved most of Taiwan Semiconductor Manufacturing Company’s top packaging capacity, which means the bottleneck is no longer only the factory that etches the chip. The choke point is now also the factory that bolts the memory and processor together into the final artificial-intelligence part. (cnbc.com) (latercera.com) Taiwan Semiconductor Manufacturing Company says its Arizona plan has grown from $12 billion to $165 billion and now includes six wafer plants, two advanced-packaging facilities, and a research center. The catch is timing: the fabrication plant is already producing, while the packaging build-out is still part of a longer expansion. (tsmc.com) So the same supply chain now has two different failure modes. One is legal and political, where export controls can be evaded through middlemen and rerouted shipments; the other is industrial and physical, where even “made in America” chips still depend on Taiwan for the step that makes them usable. (justice.gov) (theregister.com) (latercera.com) (tsmc.com) The result is a strange map for the most valuable hardware in technology. A chip can be etched in Arizona, packaged in Taiwan, installed in a server in the United States, and then become the subject of a criminal case if prosecutors believe that server was quietly steered to China. (tsmc.com) (justice.gov) (theregister.com)