Domestic substitution index

The Chinese line is the interesting one, and the usual telling of it is wrong in a specific way. Tianhe-1A took the No. 1 position in 2010 and is routinely described as a Chinese supercomputer, which it was, but its compute was American, Intel Westmere CPUs and NVIDIA Fermi accelerators. What was domestic was the interconnect: NUDT's own fabric, at a layer most observers assumed was import-dependent. On this index that machine scores as partially, not fully, domestic.

The step change arrives in 2016 with Sunway TaihuLight, which is the first system in this dataset to score 1.0: a domestically designed processor, a domestic fabric, a domestic operating system, no Western silicon anywhere in the compute path. It reached No. 1 six years before the substitution question became a mainstream policy topic.

The counter-motion is Tianhe-2, and it is the cleanest natural experiment this dataset contains. It shipped in 2013 with Intel accelerators. In 2015 the United States added NUDT and the operating centres to the Entity List, blocking the planned Xeon Phi upgrade. The 2018 rebuild replaced the American accelerators with NUDT's own Matrix-2000, and the machine got faster. An export control intended to constrain capability produced a domestic substitution instead, and this index is where that shows up as a number rather than an anecdote.

The American line sits high throughout and should be read with the fabrication caveat above firmly attached. It says that US systems are designed and supplied by US companies. It does not say they could be built without Taiwan.

The European line is the one that surprises people, and it sits at zero. Not one European system in this dataset is fully domestically supplied, even treating the EU as a single home country. LUMI runs on AMD silicon, Leonardo and MareNostrum 5 on Intel and NVIDIA, JUPITER on NVIDIA Grace Hopper. What Europe supplies is integration (Eviden's BullSequana platform, and the Bull lineage behind it), and integration is the one role in the node that does not require a fab. Eviden's BXI fabric is the exception that proves it: a European interconnect that exists, is carried in our parts taxonomy, and appears in none of the European systems we hold.

That is a sharper finding than the China line, and it is invisible in any system-centric view. A ranking shows Europe with three machines in the top ten. The supply-chain view shows that essentially all of that capability is bought, and the sovereignty programme that funded it has so far moved the assembly layer rather than the silicon layer.