John Carmack
@ID_AA_Carmack · Jul 6, 2026
Memory cost and capacity are significant issues for AI accelerators.
Unlike game rendering, model inference can have a deterministic memory access pattern. You don’t need “random access memory” at all for model weights, and you could tolerate cold-start latencies in the multiple
Elon Musk
@elonmusk
Yeah
06:11 AM · July 7, 2026 · 153.9K views
143
91
1.5K