CUDA Lock-In Under Pressure from Multiple Directions
Two threads today hit the same structural problem in AI infrastructure from different angles. Zluda 6 (story 48730713) is a project that lets unmodified CUDA applications run on non-Nvidia GPUs. The LongCat-2.0 thread (48727116) includes a candid comment from the team that their non-Nvidia ASIC clusters have immature software ecosystems compared to Nvidia's, which is a real operational cost even when the hardware is competitive.
The pattern: Nvidia's moat is not just the hardware, it's the decade of software ecosystem that makes CUDA the default. Every serious AI workload is written for CUDA. Zluda is one path around this, letting AMD or Intel GPUs run existing code. The LongCat team is betting on proprietary ASICs but paying a software maturity tax.
The Zig SPIR-V backend progress (story 48679762) fits here too. SPIR-V is the intermediate representation that lets shader code run on non-Nvidia hardware. There are multiple independent bets being made on breaking CUDA's monopoly, but none of them are close to interchangeable yet.
So what?
If you're building AI infrastructure that needs to scale, you're still essentially required to use Nvidia. That's a supply constraint and a cost constraint that multiple serious projects are trying to break. Watch Zluda and AMD's ROCm ecosystem specifically. If CUDA compatibility improves on AMD hardware, the compute cost curve for AI startups changes meaningfully within 12 to 18 months.