Nine supercomputers, six countries, one foundry

This catalog’s /supercomputers section now covers nine flagship machines, built for six national or regional programs across six countries, using compute silicon from four different designers. Across all 57 company namings on their 111 build-out lines, one company shows up on every single machine.

The list

Every row names a different buyer and, in six of the nine rows, a different chip designer. Leading-edge foundry work for every one of them lands at the same company:

The part that makes this more than “Nvidia and AMD both use TSMC”

Two of these machines exist specifically to avoid depending on a US GPU vendor’s chip design. Fugaku runs on the A64FX, an Arm-based processor Fujitsu designed itself rather than buying a merchant GPU. JUPITER’s cluster module runs on SiPearl’s Rhea1, an 80-core Arm chip designed by a French-German startup for exactly this program. Neither is an Nvidia or AMD part in any sense — the instruction set, the core design and the customer are all different from the other seven machines on this list.

Both still fab at TSMC. SiPearl’s own site states Rhea1 “will equip JUPITER” and is built at TSMC; Fujitsu’s A64FX documentation names TSMC’s 7 nm process and CoWoS packaging. Designing your own processor changes who draws the chip. It has not, in either of these two cases, changed who manufactures it.

How solid the citation is

TSMC’s name is attached to at least one line on all nine machines, and on eight of the nine — every machine except El Capitan — that citation rests on a vendor’s own published specification page (Nvidia’s Hopper write-up, AMD’s MI250X/MI300A pages, Intel’s Ponte Vecchio material, SiPearl’s and Fujitsu’s own product pages), not on secondhand reporting. El Capitan’s single TSMC mention, sourcing the MI300A die split, is tagged reported: trade-press coverage of AMD’s packaging rather than a line on AMD’s own site naming the foundry for that specific part. Worth knowing before treating “all nine” as nine identical-strength citations — eight are disclosed, one is reported.

The next tier down

Below TSMC, the pattern is a drop-off, not a second monopoly. HPE (as system integrator, mostly under its Cray HPC line) and AMD each appear on six of the nine machines; Nvidia on five; Arm’s instruction set underlies four, through three different implementers (Nvidia’s Grace, Fujitsu’s A64FX, SiPearl’s Rhea1) who compete with each other everywhere else. Thirteen of the catalog’s 25 distinct supercomputer suppliers appear on exactly one machine. The foundry layer concentrates in a way the rest of the stack, including the switch fabric and the rack integration, does not.

What this is not

This reads the nine machines this catalog currently covers, not the world’s full supercomputer fleet — leading-edge foundry capacity exists outside TSMC (Samsung, Intel’s own fabs), and a machine built around one of those would change the count. Nine for nine is this catalog’s current sample, not a claim that no other foundry has ever built a flagship system.