Startup Modular raises $250M to challenge Nvidia's software dominance (www.ft.com)

Modular announced a $250 million funding round to build an alternative software stack aimed squarely at breaking Nvidia’s de facto control over AI tooling. The startup says it will invest in compilers, runtimes, optimized kernel libraries and developer tooling that let popular ML frameworks (PyTorch/TensorFlow) run efficiently across multiple accelerators—not just Nvidia GPUs. The pitch is portability: reduce CUDA lock‑in and give model builders the option to target AMD, Intel, AWS accelerators and other AI chips without rewriting kernels or accepting large performance tradeoffs. This matters because Nvidia’s ecosystem—CUDA, cuDNN and related tooling—has long been the performance and developer standard for training and inference. A well‑executed cross‑platform stack would lower costs, spur hardware competition, and accelerate innovation in model optimization (mixed precision, quantization, sparse kernels) and distributed training. The technical challenge is nontrivial: matching hand‑tuned vendor kernels requires sophisticated compiler IRs, autotuning, memory/scheduling optimizations and close co‑design with hardware partners. With $250M, Modular can hire engineering talent and build partnerships to tackle those problems; if successful, the effort could reshape deployment choices for researchers, cloud providers and chipmakers, and reduce single‑vendor dependency in the AI stack.