The last week of January 2026 has arguably been the most consequential week for AI infrastructure since the release of the first H100. Within days of each other, two major commercial entities emerged from the UC Berkeley ecosystem to stake their claim on the 'Inference Stack.' Inferact, the commercial vehicle for the vLLM team, launched with a massive $150M seed round led by a16z and Lightspeed, valuing the project at $800M. Simultaneously, the team behind SGLang officially spun out as RadixArk, securing a $400M valuation in a round led by Accel. These aren't just venture rounds; they are a declaration that the 'operating system' for the model'the layer that manages memory, scheduling, and kernels'is now the primary battleground for AI dominance. In 'LLM time,' we are moving from the 'show-and-tell' phase of R&D into the 'unit economics' phase of production. The problem is no longer just training the largest model, but serving it efficiently. Inference is an unpredictable,...
learn more
