Nvidia backs AI inference optimisation startup Simplismart - Global Venturing

Nvidia's bet on inference optimization is not a portfolio move—it is a strategic repositioning. As training compute dividends plateau, the moat shifts from "faster" to "cheaper per token." Inference will consume over 70% of total AI compute within 18 months, and once optimization layers embed into mainstream frameworks, the switching cost for AMD's ROCm or Intel's Gaudi escalates exponentially. On the supply chain, inference is far more memory-bandwidth-sensitive than training, accelerating the HBM3E-to-HBM4 transition and intensifying the fight over advanced packaging capacity in Taiwan, China. Export-control logic also erodes: if per-token inference costs drop 40%, the "insufficient compute" narrative collapses, and compliance pressure migrates to the software layer. Within 24 months, inference optimization becomes existential rather than incremental. Expect two or three top startups to be acquired, and the window for hyperscaler in-house inference stacks to close. Nvidia's real play: define inference standards through software, demoting hardware competition to a secondary axis.