⚠ Translation LayersModerate threat

The most direct attack on switching friction is the translation layer — software that lets code written for CUDA run, unmodified or nearly so, on a competitor's hardware. Efforts in this vein, from AMD's compatibility tooling to independent projects¹, aim to let a customer keep their existing CUDA software while swapping the Nvidia chip beneath it for a cheaper rival, which would demolish the single largest barrier to switching.

The danger is that translation, if it worked seamlessly, would turn Nvidia's greatest asset into a commodity socket. If a customer could run their whole CUDA codebase on a rival chip at comparable performance, the accumulated software investment would no longer bind them to Nvidia at all, and the competition would collapse back to price and raw silicon — precisely the fight Nvidia most wants to avoid.

What has protected Nvidia so far is that translation is fiendishly hard to do well. CUDA is a large, fast-moving target, its performance depends on deep hardware-specific behavior, and a translation layer tends to lag Nvidia's newest features and to give up meaningful performance in the process. Nvidia has also moved to discourage such layers², and a customer betting production workloads on an unofficial bridge takes on real risk.

Call it moderate. Translation layers are the most direct assault on the switching cost and they improve steadily, but they remain perpetually behind Nvidia's frontier and lossy where it matters most — and running one's most important workloads on a fragile, unsupported bridge is a gamble few large customers will take. The friction is lowered at the edges, not yet dissolved at the core.