A modder rebuilt NVIDIA's DLSS 5 neural rendering pipeline to run on Intel's Arc 140V integrated graphics via XMX units, with no CUDA involved. It works, technically, but frame rates crater to 3-5 FPS.
The headline reads like a jailbreak win; the real signal is the opposite. Within weeks the modding community moved from older RTX cards to AMD's RX 9000 series and now Intel integrated graphics, which tells us the algorithm itself is portable. What does not port is performance. The gap between 'runs' and 'runs playably' is precisely NVIDIA's competitive advantage: years of hardware-software co-design, tensor throughput tuned to specific kernels, and a driver stack that assumes the silicon underneath. CUDA is not a fence around the idea. It is a fence around the execution. For AMD and Intel, this is a double-edged result. It proves their hardware can host frontier rendering workloads, but it also quantifies how far behind the tuning is when the same code drops from playable to a slideshow. Feature parity on paper does not close a moat built on optimization.
The broader lesson applies well beyond gaming. Every enterprise weighing a move off NVIDIA to cut inference costs is running a version of this experiment. Portability of the model or algorithm is rarely the blocker; the collapse in throughput on alternative accelerators is. That is the hidden switching cost buyers underestimate.
For Japanese companies the implication is direct. Game studios and hardware groups such as Sony and Nintendo's ecosystem partners have long optimized around fixed, known silicon, and this episode validates that discipline over chasing vendor-agnostic flexibility. More consequentially, Japanese enterprises and SIers building AI inference platforms face the identical CUDA lock-in calculus. Proposals that assume smooth migration to cheaper Intel or AMD accelerators to ease GPU procurement pressure should budget for a real performance penalty and heavy re-tuning, not a drop-in swap. Firms tied to NVIDIA supply constraints may find the exit door narrower than the spec sheets suggest.
For SIers advising clients, the honest framing is that hardware diversification is a multi-quarter engineering project, not a purchasing decision. The modder's 3-5 FPS is a useful benchmark for how much optimization work sits between owning alternative silicon and actually using it in production.
