if Cerebras's architecture is the future, 10x Nvidia replacement shouldn't at all cost all port all resource to Cerebras instead Nvidia, accelerate development and resolve constraint
why not intentionally sacrifice short-term Nvidia scale to accelerate Cerebras long-term?
Anyone with technical knowledge can explain
base on cerebras current constraint and strength, if constraint was solved practically, wouldn't that pick them no brainer?
if reward clear why don't at all cost resolve the constraint Cerebras face, then nvidia can gone for good with groq.
base on company ranking cycle, we would sense: who lead wrong direction was the one have most resources and wasted most, and who lead right dir access to little or none.
cerebras 10× intelligent mass per time passes for both training and inference (if you total the time u run on nvidia minus cerebras, that's the number u just wasted waiting on nvidia, and eventually shipped 10x slower, overtime 10x less usable intelligence).
for user experience, here is the story. even 4G to 5G most barely notice diff(battery drains was the one much noticeable but like nobody know it), but still rest people choose 5G think is upper replace and worth the higher price, dont even talked initial 5g infra cost. who one bought ergonomic keyboard from logi tech, and they not even sell split keyboard
and when come to the real innovation side here, it become inverse. ppl keep port their $ to legacy gpu co, instead of exponentially accelerate Cerebras.
other some few downside i know like: 3× to 11× higher price on api (that's ceil but i still unsure the expense cost) and dont support img input.