Announced in full production at Hot Chips on August 24, 2026, LPX targets the decode-phase bottleneck in agent loops where users feel streaming latency during tool use and replanning. Artificial Analysis measured ~3,400 output tokens per second on Gemma 4 31B at 100K input context. Nebius Token Factory was the first cloud adopter, exposing LPX through existing APIs — datacenter rack hardware, not a consumer SKU.