Rohan Paul (@rohanpaul_ai)T2·72 pts
- Microsoft walked away from the Oracle contract because the hardware used older Nvidia chips that no longer matched its specs, leaving capacity that depreciated over time
- A Blackwell GPU produces several times the inference output of an H100, so per-GPU-hour pricing makes older cards buy far less work
- Frontier MoE models need large memory and fast interconnects; an H100 server links only 8 GPUs, while a Blackwell rack links 72 as one unit
- Oracle says AI infrastructure demand remains strong and renewed 4-year-old GPU contracts at a 20% premium, showing older capacity still has buyers