skinny.

Latest / SAGA — Success Stories | AGORA Intelligence

ZenteiQ 4x Synthetic Data Throughput with GKE Inference Gateway

ZenteiQ says it gets 4.07 times more synthetic tokens per 1,000 accelerators after replacing round-robin with routing that reads queue depth and cache state. Running those numbers, generating 100 billion tokens would drop from about 68 days to 17, freeing roughly 51,000 accelerator-days for pre-training. Every figure comes from a post co-written with Google, the vendor.

The skinny

The skinny isn't ready yet — notes appear once the transcript is processed.