Ah I see. I still think if you're measuring that way, there's no reason languages like Python et al. can't accomplish similar too. And if they can't, well, there's always adding more machines. Horizontal scaling tends to occur no matter what you're using, so is the argument just that with less productive (probably a matter of taste for many at this point) languages you'll have to scale out later? That's a tradeoff to consider, but there are other tradeoffs too.
https://www.techempower.com/benchmarks/#section=data-r14&hw=... has some interesting benchmarks, apparently we should be using Dart or C++. Go does slightly better than JS but not a lot. Some newer Python frameworks aren't on there yet. None of them reach near 50k, but I don't know the details of the benchmark and they aren't all using the same DB. Certainly you can get crazy numbers depending on what you're testing. e.g. https://github.com/squeaky-pl/japronto gets one million requests per second through async and pipelining but those requests aren't doing much.
True, horizontal scaling will always save you no matter how slow the front end is. Cost becomes significant at a certain level too though. For example, Google estimates each search query goes to over a thousand machines. If you need 100x 1000 machines to serve a query because the back end is PHP it adds up.
And you can make Python or even PHP fast if you try hard enough.
My argument is that the engineer overhead for Go and new breeds of Java frameworks are small enough that it makes no sense to use anything else if you're planning on scaling for real.
If you start with something else the cost of making a slow language fast and the multiples of extra machines you need costs far more than just using the faster language to start with
For the benchmark you posted, take a good look at the "realistic" vs "stripped" implementations and whether the test used an ORM. You'll quickly see that the realistic test applications with any kind of ORM are exclusively C#, Java, Go, and C++
Java/Go with pretty much any database can manage about 50,000 individual queries+REST calls a second.