A startup called MachGen emerged from stealth last week to address the slow speed and high cost of AI video generation. Founded by Kismat Singh and Manoj Krishnan, the company is building a high-performance inference and fine-tuning stack designed specifically for diffusion models. The founders’ collective resume includes co-building the TRT high-performance inference teams at Nvidia, leading Datacenter AI software at Intel, driving the vLLM inference effort at Google, and building the large-model training infrastructure behind PyTorch at Meta. MachGen aims to push diffusion latency and compute costs down by an order of magnitude to unlock interactive, real-time video generation. The core technical problem is that diffusion models are non-autoregressive and compute-bound throughout their entire “denoising” loop, unlike LLMs. To fix this, MachGen re-engineered the stack across four pillars: attention using lower-precision recipes and sparse attention techniques; proprietary caching recipes exploiting spatial and temporal redundancy; and highly developed kernels.