Datacenter GPUs pinned to 100% won't make it to their 3rd anniversary, models are getting larger and larger, they get smarter by running longer "reasoning" loop

Datacenter GPUs pinned to 100% won't make it to their 3rd anniversary, models are getting larger and larger, they get smarter by running longer "reasoning" loops, there is no indication that it'll get better soon. > Open source models are 3-6 months behind. On the benchmarks included in their training set yes, not in real life