Nvidia backs open standards for AI factory builds - IT Brief UK

Nvidia's endorsement of open standards for AI factory construction is not generosity—it is a standard-setting power play. History repeats: Intel's 2003 push for PCIe, ARM's 2015 v8 opening—each "openness" calcified the proponent's architecture into the industry default. The technical cascade is precise: once interconnect layers (NVLink, UALink, CXL) converge on a common spec, NVSwitch transitions from proprietary interface to de facto standard, structurally raising rivals' switching costs. Broadcom's custom-ASIC thesis—built on "non-standard, deeply tailored" silicon—loses its differentiation premium. AMD's UALink coalition and Intel's CXL camp face dilution: customers no longer pay an openness premium when the dominant player has already "opened." On supply-chain risk, open specs objectively reduce single-vendor lock-in, but in the accelerating US-China tech decoupling, cloud operators in mainland China and packaging capacity in Taiwan, China may leverage this window to accelerate proprietary interconnect protocols, creating a parallel standards regime. 12–24 month outlook: AI factory procurement shifts from "buying Nvidia" to "buying Nvidia-compatible factories." The CUDA moat survives hardware standardization. The real inflection arrives when inference workloads exceed 60% of total compute—software switching costs finally drop below hardware costs, and Nvidia's "openness" becomes its exposure.