• fruitycoder@sh.itjust.works
      link
      fedilink
      English
      arrow-up
      5
      ·
      9 hours ago

      Honestly trying to treat token like watt is pretty crazy. Like a watt of power truly is a scientific standard. A bit of information is closer but again pretty standard, you know what are sending when you choose to send or receive a bit of data. Trying to apply a standard utilization rate to a unit of information in exchange for a non-deterministic rate, size, or quality back is nuts.

      How do you meaningfully show an ROI on that?

      • mlg@lemmy.world
        link
        fedilink
        English
        arrow-up
        1
        ·
        7 hours ago

        I wonder how long they can keep the circus running though. They were obviously trying to vendor lock enterprise size businesses in before they eventually nuke the cost model, and Deepseek just preemptively slashed their cloud pricing probably knowing that everyone else will just migrate or even run on older cloud hardware which providers can sell by watt (and demand).

        China might be behind on hardware, but their biggest wrench in the gears will probably just be popularizing their published LLMs which don’t require multiple nuclear powerplants to run lol.