Initially, the focus was on creating a transpiler to transition CUDA workloads to other silicon targets like Inferentia, Cerebras, and Groq with equivalent or superior performance.