Rohan Paul (@rohanpaul_ai)T2·Infra
71 pts
Positron's Titan servers abandon scarce HBM for cheap smartphone-style DRAM, letting a single machine hold models of up to 32 trillion parameters and making memory capacity per watt the key metric. Inference cloud Parasail will be the first to deploy Titan at scale, starting at 5MW, claiming 4X better cost-per-token than next-gen NVIDIA/AMD systems and targeting trillions of cost-optimized tokens per day by late 2027.