Whenever someone says something like "they've tested it with X throughput", it usually means that they're not considering security properties. Sure, you can push throughput through on test machines in a test setting and it will happen. But in an adversarial environment, systems pushed to those limits can often be attacked. Its not sufficient to test throughput in normal conditions. You have to go through every attack vector and test the throughput during attacks and ensure the throughput you're pushing through isn't making those attacks easier to pull off.
Centralized systems can push plenty of throughput. The question is not: "How fast can we make this system run?", the question is "how fast can we make this system run WHILE keeping its properties of decentralization and resilience against attacks?".
They claim they have managed security, scalability and decentralization. I'm in no position to evaluate this. However they've got Sequoia onboard, plus a Turing award winner, plus Shanghai gov and team is full of PhDs...so I figured good enough for me.
Not Eth, but you're missing one option - tree graph based. They've pushed 3000+ TPS through it already in testnets...and that's on layer 1. High level description:
https://confluxnetwork.medium.com/tree-graph-structure-of-co...
>or layer 2 on BTC, aka lightning network?
Indeed - that could work. I do think Eth has the edge though. Throughput only matters if you have an eco-system that needs it ultimately.