What is the value proposition of "RL environment" companies? [D]

Recently I've been getting bombarded with reach outs from companies that build RL environments and sell to frontier labs like Anthropic, OAI, etc. I'm not really understanding how and why these companies exist. Lets say you somehow create an environment that no other model is able to perform well on. Your second criteria would be that this environment is suitable for RLVR, perhaps you have some metric of gradient information per example in the dataset. So you have to prove to the lab you are selling this on, that 1) their model cannot do this task currently and 2) this task is solvable via RL. At that point, why couldn't you just approach a VC (most of these companies are startups with a good amount of funding anyways), show them scaling laws + your current RL env and just raise money to train your own model? We already see open source models tend to lag behind frontier ones by ~3 months or so, it sort of implies the frontier is mostly new RL environments and stuff (or distillation methods are really good). I'm kinda operating under the impression that pre-training/architecture/infra for large models is largely much easier nowadays, but maybe this is just not the case? I guess an answer to my question would be either 1) I am totally wrong regarding my assumptions on the difficulty of pre-training + infra (which from my experience in industry does seem to be much easier) or 2) actually collecting all of these other, past, RL environments is still hard and only certain labs have access to all of them or 3) all of these companies are or eventually will train their own models?