What AI architectures are required/supported

Can the design be more generalized or is there a specific architecture that it makes sense to support?

Is it okay to just have "raw processing power" or does it need to be more specialized and does this question make sense?

Sparked from a video (Untether I beleive?) talking about flexibility of computing power. VERY loosely paraphrased from memory: "If you try to spin up hardware for CNNs four years ago for image processing and now you find everyone is using transformers instead, your hardware is now useless/less effective."

Are there pitfalls for general computing vs specific architectures? Is the performance too slow to be useful? Is there a significant increase in design complexity?