The AI arms race is no longer about who can hoard the most silicon; it is about who can stop that silicon from idling. While the industry grovels for NVIDIA Blackwell allocations, the team at Cursor has proven that even the most expensive GPU is only as fast as the code that feeds it. With the release of Mixture-of-Kittens (MoK), Cursor isn’t just building an IDE—they are dictating how the NVL72 rack should actually work. By bypassing standard vendor libraries in favor of a specialized megakernel built from first principles, they’ve turned the Blackwell SM100 and SM103 architectures into something far more efficient than what comes out of the box.
The Architecture of the Megakernel
Technical efficiency in the Mixture-of-Experts (MoE) framework usually dies in the ‘plumbing’—the constant, friction-heavy data movement between GPUs and processors. Mixture-of-Kittens eliminates this by fusing all MoE computation and communication into a single, aggressive kernel. In a typical setup, the CPU and GPU are locked in a tedious back-and-forth, creating micro-delays that aggregate into massive performance leaks. MoK kills CPU-GPU synchronization entirely. By overlapping compute tasks with inter-GPU networking at a granular level, the system ensures that while one part of the NVL72 rack calculates token weights, another is already shoving the next batch across the wire. The result? The Blackwell chips never have to wait for instructions.
MoK fuses all MoE computation and communication into a single kernel, overlapping compute and inter-GPU networking at configurable granularity, and fully eliminates CPU-GPU synchronization.
For CTOs and infrastructure leads, the gains move from theoretical to transformative. Cursor’s benchmarks show MoK is up to 2.37x faster for MXFP8 forward passes and 1.92x faster for BF16 forward passes compared to the fastest existing baselines. This isn't a lab experiment; these kernels power the production training of Cursor’s Composer. By pairing BF16 with the high-throughput MXFP8 format, MoK provides a deterministic path that stabilizes enterprise-grade models without sacrificing the raw speed of the cluster.
Economic Realities of Optimized Training
The financial logic is brutal: when a single NVL72 rack costs millions, every second of training time is a line item on the balance sheet. Reducing training duration effectively doubles the ROI on hardware that most companies are still struggling to source. MoK requires a modern stack—Python 3.12, PyTorch 2.10, and CUDA 13.0—targeting the SM103 and SM100 architectures. By optimizing both forward and backward passes within the same fused structure, Cursor allows enterprises to slash their cloud compute bills and energy consumption simultaneously, proving that TCO is a software problem as much as a hardware one.
Compared with the fastest baseline, MoK is up to 2.37x faster for the MXFP8 forward and 1.58x for the BF16 backward.
This shift suggests that the primary competitive advantage in AI has moved from raw GPU count to the mastery of memory and communication buffers. Cursor’s functional layer manages workspace creation and scheduling, making these tier-one optimizations accessible to firms that don't have a small army of CUDA specialists on staff. This democratization of high-performance kernels through open repositories means R&D cycles are no longer tethered to vendor release schedules. If you can’t optimize your kernels, you are overpaying for your compute.
Software expertise is now the definitive bottleneck for AI performance, regardless of how many TFLOPS NVIDIA promises on stage. The Mixture-of-Kittens release demonstrates that a nimble engineering team can outrun general-purpose vendor libraries by designing specifically for the physical reality of the NVL72. As Blackwell becomes the enterprise standard, the winners won't be those with the biggest purchase orders, but those with the smartest code to run on them.