Vibe Coding: easy to say, difficult to survive
Disclaimer: The views and opinions expressed here are my own and do not necessarily reflect those of my employer (NTT DATA).
Vibe coding is an artificial intelligence-assisted software development technique popularized by Andrej Karpathy in February 2025. from wikipedia
When K8s started spreading, there was already solutions in the area. AWS has cloud formation, Docker has swarm etc. But no one managed the exact spot addressed by K8s (read: every cluster need with the required features). In this phase, a lot of companies proposed ‘managers’ to manage early clusters, then started to propose ‘managers’ for k8s, then started simply to provide accessories services around K8s (like proper installation and so on). In the meantime K8s expanded its scope to provide stronger orchestration support which now we give for granted
Azure was the last company to provide K8s support around 2017 side by side with its propietary managers: and so k8s won.
Now vibe coding is in a similar situation because there are plenty of startup offering more or less the same thing: Replit, Lovable, Cusror… there are at least 8+ big tools out of there.
OpenAI and Anthropic are offering similar tools too.
I have the chance to use OpenAI Codex and Github Copilot deeply (thank to NTT Data’s Enterprise license we have).I have a short experience with Replit and Lovable.
OpenAI Codex works very well for simple tasks (like refactoring small code portions). It seems like having all StackOverflows team at your command :) Adding a github actions to enable CI/CD, or converting an application.properties to yaml are all simple tasks Codex can do very well. But if you ask Codex to write documentation or update a README, small error arise, like:
- adding not-pertinent statement to the readme (mission sometimes the point)
- failing to detect the correct License (confused MIT with Apache 2.0 for instance, but okey they are similar)
If you remember AppFuse the output is similar but more powerful because you can also refactor existing code. Also I exepct good perfromance on spaghetti-Java code, but I do not have a bad example to test it.
Github Copilot seems a bit less powerful, but the results can be similar. The difference is the Senior Developer which reviews the produced code, because asking is easy, but reviewing is not.
The hard quesiton is: how much time do you gain? I have not precise numbers, but for sure if (and only if) the answer speed of the GenAI increase, it can boost your productivity. If the answer speed is around 1-2 minutes like now, a good Spring boot template with well strutured code and a clean idea, can bet it for a while.
Missing revenues
OpenAI and Anthropic (+ MistralAI, and the other) needs to monetize and offer more and more services. OpenAI is pushing on Codex for this reason in my huble opinion: they released on September 15th a new ChatGPT-5 Codex refresh.
My prediction is that is unlikely that a Company focusing only on vibe coding will survive easily, at least not with the price tag I see now.
Vibe coding require a lot of trial and error from the user side, so their user base will not be happy to pay multiple services in a month if the total sum exceed a Netflix-like subscription (around $15-$20).
A lot of these startup seems to resell OpenAI and Anthropic GenAI services: it is not clear to me how much value they add to them, to justify the price.
Last but not least, the revenue pie seems very tiny right now: over 90% of ChatGPT users are using the free tier. These services must boost productivity for real, to justify their introduction (and the security risk of sharing your company code with them, in the cloud).
And you what do you think? Feel free to add a comment below!