vLLM reportedly gets within 6% of DeepInfer’s published decode throughput on GB200

The vLLM project says that, with Wide-EP and EPLB, its GB200 throughput came within 6% of DeepInfer’s published decode results across batch sizes 8–256 under DeepSeek’s configuration. On the AgentX workload, it reports up to 1M tokens/s/GPU on GB300 and estimates $38.96 per GPU-hour in revenue at 196 P90 TPS using DeepSeek’s off-peak API pricing. It also reports up to 826 TPS per user on AgentX with extended low-latency kernel support.