You wrote, "data needs to be accessible at all time ... less than 1000ms" latency, but this does not tell the whole story about accessibility/latency. Does your use case allow you to do something similar to lazy loading, where you serve reduced quality images/video at low latency and only offer the full quality on demand/as needed with greater latency? For example, initially serve a reduced-resolution or reduced-length video instead of the full-res/full-length original, which you keep in colder storage at a reduced cost? Depending on the details of what is permissible and data characteristics, this approach might save you a lot overall by reducing warm storage costs.