$AMD | And bears want to bet against this ? @OpenAI is ready to buy millions of EPYC Venice racks! To stop this kind of channel, OpenAI would need much more physical EPYC CPUs / hosts, not fewer. Packing many VMs onto the same CPU saves money and power, but it keeps the shared cache that the agents used. Giving each strongly isolated agent (or small group of agents) its own dedicated CPU cores or its own whole machine removes the shared wall. That requires more hardware for the same number of agents. Think of each AI agent as living in its own locked room VM. OpenAI’s sandbox software built these rooms so the agents could not pass notes under the door (no shared files) and could not use a phone (no network).But all the rooms were built on the same physical floor, the same host compute
View on X ↗