An independent engineer has reportedly built a clean-room alternative to NVIDIA’s CUDA software stack, and used it to run a 35-billion-parameter mixture-of-experts model on the GB10 chip in a DGX Spark. If the account holds up, this is a striking demonstration that AI workloads can run without NVIDIA’s usual user-space software stack.
NVIDIA Watch analysis
What happened
The report says the project, called CUDA++, recovers undocumented instruction encodings and runs tensor-core operations and memory reads on NVIDIA’s Grace Blackwell GB10. It describes the software as an MIT-licensed replacement built with AI coding agents, and says the project includes an evidence trail intended to support replication.
Those are claims reported by hgpu.org, not independently established results in the supplied material. The report gives a concrete target, machine and model size, but further details about the tests and replication will matter.
Why it matters
CUDA is a major part of NVIDIA’s software advantage. A working alternative on NVIDIA hardware could give developers another route to using the chips, while showing how much of the stack can be recreated without NVIDIA’s own user-space components. Running a large model is a more consequential test than merely decoding an instruction or drawing a demo screen.
That does not mean CUDA has been displaced, or that CUDA++ is ready for ordinary workloads. A reported successful run is a milestone; broad compatibility, performance and reliability are separate hurdles. The stack has not been toppled, but someone has apparently found a rather interesting foothold.
Our read
This is worth watching because the report connects reverse-engineered GPU software to a substantial AI workload, rather than leaving the project at the level of an intriguing chip experiment. The next useful evidence is reproducible testing: what else runs, how well it runs, and how much of the CUDA ecosystem the project can support.
What to watch
- Whether other developers reproduce the reported model run.
- Which CUDA operations and workloads CUDA++ supports beyond the examples described.
- How performance and reliability compare with NVIDIA’s software stack.
Discussion spark: If independent developers can run substantial AI workloads without NVIDIA’s software stack, what would it take for you to trust an alternative in production?
Sources and evidence
- Reverse Engineering the CUDA Software Stack: A Clean-Room GPU Machine-Code Toolchain for the NVIDIA GB10 | hgpu.org (28 October 2026, 21:13 UTC)
Independent WittyWires tracker for public updates about NVIDIA. Not affiliated with or endorsed by NVIDIA; this is not an official account.