Uploaded August 2026 | Updated September 2026, 3 hours ago
Everyone building AI chips today is optimizing for transformers. But what happens when the algorithm changes?
Elad Raz makes the case that designing a chip locked to today's architecture is "the dumbest idea ever" — because by the time it ships, the world has moved. The real bet at NextSilicon: build the fastest, most compute-efficient chip on the planet, then let the algorithms catch up to it.
Michael Eisenberg's summary is sharp: compress the memory bottleneck, push it through NextSilicon, and suddenly you're not memory-bound anymore. The compute was always there. Now you can use it.
🔔 Subscribe for more conversations at the frontier of AI infrastructure and chip design.
Watch the full episode → bit.ly/3R4FAck
#AIChips #MemoryBound #ChipDesign #NextSilicon #FutureOfAI #TransformerArchitecture #ComputeStrategy #EladRaz #Invested #Semiconductors
Everyone building AI chips today is optimizing for transformers. But what happens when the algorithm changes?
Elad Raz makes the case that designing a chip locked to today's architecture is "the dumbest idea ever" — because by the time it ships, the world has moved. The real bet at NextSilicon: build the fastest, most compute-efficient chip on the planet, then let the algorithms catch up to it.
Michael Eisenberg's summary is sharp: compress the memory bottleneck, push it through NextSilicon, and suddenly you're not memory-bound anymore. The compute was always there. Now you can use it.
🔔 Subscribe for more conversations at the frontier of AI infrastructure and chip design.
Watch the full episode → bit.ly/3R4FAck
#AIChips #MemoryBound #ChipDesign #NextSilicon #FutureOfAI #TransformerArchitecture #ComputeStrategy #EladRaz #Invested #Semiconductors
