The bottleneck moved: AMD Venice, Fractile's $7B bet, and the memory wall

In three days, the AI hardware race changed its shape. AMD put a 256-core server CPU on general sale. Google widened its memory supply to Samsung and SK hynix as its custom TPUs close on Nvidia's share of the world's high-bandwidth memory. A British startup that has never shipped a chip was valued at about $7 billion for stacking memory on top of logic. A Singapore startup came out of stealth with light built into the chip itself. The thread is not more compute — it is the wall between compute and memory.