RouterPlus marketplace lets providers bid on custom inference requirements

Customers can post inference requirements and get bids from providers on precision, latency, geography, data retention, and usage duration. That gives bursty workloads like evaluations and backfills an alternative to annual capacity commitments or serverless rate limits. Buyers can choose BF16 over quantized FP8, US serving, zero data retention, and fast or low-cost batched tokens. RouterPlus says the marketplace is on track for 1 trillion tokens per month.