The Operating System
16 Platform Play

Never Break Compatibility

Every Nvidia accelerator runs the same software, and Huang treats that as the one rule that is never up for negotiation.

that's the only unnegotiable rule in our company. Everything else is negotiable.

Asked on the Acquired podcast (October 2023) what Nvidia will not trade away, Huang pointed to compatibility. His claim: “We are the only accelerator on the planet where every single accelerator is architecturally compatible with the others.” Code written for a GeForce card years ago still runs on a data-center GPU today, so developers never have to start again.

The mechanism. Compatibility is what turns an install base into a promise. Every new chip inherits the software of every old one, and each developer who writes for CUDA is writing for hardware that has not shipped yet.

The critique. The same discipline looks different from outside. Analysts cited by Digitimes describe CUDA as the core of Nvidia’s switching costs. Tom’s Hardware and Digitimes reported that the CUDA 11.6 license terms barred running CUDA through translation layers on other hardware, and France’s competition authority was reported to have named CUDA among its concerns about dominance.

Doesn’t transfer for free. Backward compatibility is a tax on every future design, and it only pays when you already have developers worth keeping. At Nvidia’s scale, the discipline that protects developers also locks them in, which is why competition regulators have started asking about CUDA.