Point taken.

It’s not a “government it farce”, it’s educational institutions and the science community. And we cannot afford *not* to do it.
The taxpayer doesn’t pay for anything except government debt, all government spending is newly created money, the only limit on government spending is what the markets think we can bear, capex doesn’t upset markets, free money in tax breaks does.

The UK’s national supercomputing facilities are far removed from the kind of systems that end-up as “government IT farce”; perhaps you’re thinking of things like Post Office Horizon, 2000s-era NHS spine and various other government/local government business systems.

Successive generations of facilities have largely been successful, productive, and cost-effective in the forty-plus years with which I’m familiar. As for cost to the taxpayer, the return on investment has been demonstrated to be many times the cost of building and and operating the systems — see, for example, the London Economics study: https://www.ukri.org/wp-content/uploads/2022/07/EPSRC-050722-ImpactEPSRCInvestmentsHighPerformanceComputingInfrastructure.pdf

The UK can’t afford not to remain competitive in this capability.

Some of that comparative efficiency is probably due to research being cross-funded by student fees (remember international student fees?)

Whether having a large supercomputer should lead to increased GDP would be for the Industrial Strategy to determine.

That’s the on-CPU NUMA region–totally different.

I’m talking Non-Uniform Memory Access across multiple nodes.

Surely you are talking UMA, rather than NUMA? The Cray EX nodes each have 128 cores and 8 NUMA regions (so 16 cores sharing L3 memory per NUMA region)

Not by any definition of NUMA that I know. It uses a two-socket node with AMD 32-core Epycs, with the nodes linked by two-port Slingshot Ethernet adapters over a Slingshot fabric. Slingshot is Ethernet, not NUMA. There is RDMA for sure, but not memory coherency over its 5,860 nodes.