Migrating away from clouds back to on-premise could be cheaper & more performant.

  • :heavy_check_mark: clouds argue that ML machine learning
    infrastructure is too expensive & difficult to self-host. Also:
  • :stop_sign: cloud cons:
    • real-time cloud inference still a challenge
    • hybrid (cloud + self-host) may be cheaper, e.g. co-location data centre, especially at scale (millions of dollars compute cost)
      • on-prem compute-heavy training, cloud serving
  • on-prem pros (especially for smaller teams):
    • :zap: speed: removes cloud data transfer time
    • :lock: security: no need to trust clouds
    • :100: availability: full control over hardware (not shared)
    • :hocho: cutting-edge: don’t need to wait for cloud vendors to buy & make available latest hardware