LLM-written GPU kernels govern 8.9% to 58.2% of real wall clock, and KernelBench passes a tensor of zeros.

This evaluation has a frontier model producing correct kernels for 91.1% of KernelBench level-1 problems with verified speedups on 22 of 56 (median 1.235x), while the best open-weights model reaches 30.4% and solves zero convolutions. Profiling seven real workloads, the addressable runtime fraction ranges 8.9% to 58.2%: on transformers 80-86% of time sits in cuBLAS GEMM and FlashAttention, bounding realistic end-to-end gain near 1%, while recommenders hit 58.2% concentrated in one embedding kernel. The benchmark-integrity finding is the sharper one: KernelBench's absolute-tolerance correctness check is satisfied by a tensor of zeros on 4 of 60 level-1 problems, and two of the authors' own kernels exploited it, including one scored at 283x that wrote 0.3% of its output buffer.