· Compute Atlas
Europe's domestic substitution index is zero
Not one European system in our dataset is fully domestically supplied, even counting the EU as one home country.
We built a domestic substitution index to watch the China line, because that is the one everybody expects to be interesting. The China line is interesting. The European line is the one that stopped us.
It is zero. Not low: zero, in every year we have data for, even after we relaxed the test to count the entire EU as one home country rather than scoring a German system against German suppliers.
What the test actually asks
A system counts as fully domestic if none of its four recorded supply roles (CPU, accelerator, interconnect, operating system) has a supplier headquartered outside the system’s own country, and if at least the CPU and interconnect are recorded.
An absent role is not a foreign role. Fugaku has no accelerator at all, and scoring that as a gap would rank Japan’s most sovereign machine as partially foreign. So a missing accelerator is fine. A foreign one is not.
By that test the United States sits around 94% and Japan around 96%. China moves from zero in 2010 to a step change in 2016. Europe never leaves the floor.
Why
Run down the European systems in the dataset and the pattern is the same every time.
LUMI is AMD compute on HPE Slingshot. Leonardo is Intel hosts with NVIDIA A100s on NVIDIA InfiniBand. MareNostrum 5 is Intel and NVIDIA. JUPITER, the first European exascale-class machine, is NVIDIA Grace Hopper. JUWELS Booster, Vega, Karolina: AMD or Intel hosts, NVIDIA accelerators, NVIDIA fabric.
What Europe supplies is integration. Eviden’s BullSequana platform, and the Bull lineage behind it, appears as the integrator on a substantial share of European installed FLOPS. That is a real capability and a real business. It is also the one role in the node that does not require a fab.
The exception that proves it
Eviden has a fabric. BXI was developed explicitly so that European HPC would not depend on American interconnect vendors. It exists, it works, and it is carried in our parts taxonomy.
It appears in none of the European systems we hold.
That single fact is the sovereignty programme’s problem in miniature. Building a European alternative is not the hard part; getting it bought, in preference to the incumbent that every code is already tuned for, is. The programming chapter explains why that inertia is so heavy: the switching cost is not the hardware, it is the years already spent tuning to the hardware you have.
What this is not
It is not a claim that European HPC is a failure. The systems are real, the science is real, and three of them have been in the top ten. EuroHPC has built a fleet from nearly nothing in under a decade.
It is a claim that a ranking-shaped view of that achievement is misleading. A ranking shows Europe with several machines near the top. The supply-chain view shows that essentially all of that capability was purchased from American suppliers, and that the layer Europe controls is the one furthest from the silicon.
It is also not a complete measurement, and the limitation cuts against the American number rather than the European one. This index measures design and supply nationality, not fabrication. Essentially every leading-edge accelerator in this dataset is manufactured by TSMC, including all the parts scoring as domestic for the United States. A country can score 94% here and still be entirely dependent on one foundry in one place. That caveat is on the analysis page in bold, and it should travel with any number taken from it.
Check it yourself
The per-system breakdown scores every machine role by role, and the derived CSV carries both the index and the per-system scoring the index is computed from. If you think a row is wrong, that table is where to settle it: every edge behind it carries its own source.