New analysis out: “AI Hardware Choices are Highly Variable and Sparsely Disclosed” (with Francisco Ríos, Ian Reynolds, Robert Praas, and Irene Solaiman)! We find the hardware landscape is both diversifying and poorly documented: on training, NVIDIA still dominates but Chinese developers increasingly mix in domestic accelerators like Huawei’s Ascend, while inference options keep expanding, especially for locally deployable models. Because hardware choices materially shape both model performance, this sparse disclosure undercuts reproducibility and informed policy. We call for more transparent reporting from model trainers and inference providers alike.