Rohan Paul (@rohanpaul_ai)T2·71 pts
- A single Titan server can hold models up to 32 trillion parameters, shifting the key metric from raw FLOPS to memory capacity per watt
- Parasail is the first at-scale deployer, starting at 5MW and targeting cost-optimized inference
- Positron claims 4X better cost-per-token than next-gen NVIDIA/AMD systems, aimed at large-scale agent swarms