there is a difference between moving your computing to cloud and language models to the GPT3. Language models are very context specific and moving training from generic corpora to domain specific corpora and even function specific corpora (how people search something is very different from how they describe it, for example) changes performance a lot . GPT-3 is trained on generic corpora. it is not trained on your specific domain. Perhaps, in future, they will provide tuning options, but they are not available now. There are multiple NLP applications. Some of them are latency and throughput sensitive. Say, you apply NLP in search, your startup may need to call multiple different NLP tasks within 50ms latency limits and 10000 qps throughput. You'll not get it with current GPT-3 service. Another problem. As it's proven model training encodes various biases. Some of them might be specific for your business. You need a control point to tune your model to eliminate them. Generic cloud service will not do it for you. I do not say that using generic cloud language model by startup is impossible, but there are many technical problems to resolve to make it possible.