Strange. Make sure that your thread actually destroys the context before exiting, and that you don't have some kind of weird race condition that somehow leads to context (or DC) leakage. There might be an internal limit on the number of open contexts that can be hardware accelerated.

What you're trying to do there is a very uncommon (and extremely inefficient) usage scenario. So I wouldn't be surprised if NVidia didn't test their drivers in such a situation.

The usual way to handle this scenario would be through a permanent worker thread, that would be assigned a single GL context. If you don't need the thread, put it to sleep, but don't destroy its context. Once you need it, wake it up, let is to its work, and let it sleep again. Constantly creating and destroying threads is already bad enough. But doing this with GL contexts is even worse, from a performance point of view.