Gimlet Labs
Multi-silicon inference cloud targeting frontier labs and large data centers. Its software disaggregates agentic workloads into stages, routing compute-bound phases to GPUs and memory-bound phases to higher-bandwidth accelerators, delivering 5-10x throughput gains for the same power. Also ships kforge, which auto-generates optimized CUDA, ROCm, and Metal kernels from PyTorch using multi-agent search.