Most of the time on the backend is spent waiting for network requests to complete. With LLMs, the inference cost is even longer. You don't need much compute.

With dynamic languages like Python or TS/JS, you lose money in hardware cost, but you gain 10 fold in not having to pay your developers as much because everything gets done faster.

Strong typing is also not a guarantee of correctness. In fact, anything but fully strict dependently typed language is a waste of time honestly. Because with the former, you only write code, and if your program compiles, its guaranteed to be 100% correct, so you don't need to write unit tests.

If you don't have that, you have to write unit tests, at which point you may as well spend time writing the unit tests to prove correctness independent of the language being used.

This is why TS is winning. Agent frameworks largely replace backend logic, and they have their own quirks with runtime costs. Most agent use cases look like chron job scripting or chatbots right now, and I expect we'll see far more use cases with proactive scripts. All of these avoid traditional bottleneck of runtime because the AI is the bottleneck.

Yeah I'd use something different for sure. If bang per buck is needed. Since an API call to AI can cost like 1c or more maybe it doesn't matter if you waste 100ms of cpu time.

There is value in fast iteration and delivery. These are early stage startups, the clock is ticking. I wouldn’t do it with JS but I understand why it would be used.

Now we can argue how good or bad the nodejs (or any alternative JS implementation) is. At the end of the day the objective is most likely to rock the boat enough to be acquired and have a payday. With enough discipline you can get far before the footguns show up.

It depends. There is no right or wrong.