MachGen

MachGen AI builds a high-performance inference and fine-tuning stack purpose-built for diffusion models, covering image, video, and emerging world models. The company optimizes open-source generation models at the kernel level - attention, caching, kernels, and parallelism - so they run at a fraction of the usual latency and cost while producing the same output from the same weights. Published benchmarks claim roughly 4-6x faster image generation and about 6x faster video generation, at 2-4x lower cost than incumbent providers. It was founded in 2025 by Kismat Singh and Manoj Krishnan, engineers who previously worked on TensorRT inference at Nvidia, datacenter AI software at Intel, vLLM at Google, and PyTorch training infrastructure at Meta.

A diffusion-specific inference stack rather than an LLM stack retrofitted for images and video - kernel-level optimization that keeps outputs bit-for-bit equivalent to the original weights while cutting latency several-fold, which independent benchmarking placed in the fastest and lowest-cost quadrant.