Can GPU Telemetry Verify LLM Training? Evidence from Temporal Structure in a Black-Box Setting

Many proposals for governing advanced AI systems rely on the ability to monitor or verify large-scale training runs, yet such verification often depends on access to logs, software instrumentation, or trusted reporting. This paper investigates whether temporal patterns in hardware-level GPU telemetry can provide workload-blind evidence for distinguishing LLM training from other computational workloads. We collect high-frequency telemetry traces in a controlled single-GPU setting across Llama and Qwen autoregressive LLM training and inference, BERT-base MLM training, and ResNet-50 CNN training. We analyze power signals using window-based features and a one-class verification approach. We find that autoregressive LLM training traces exhibit a distinctive temporal structure in GPU power telemetry in this setting. A verifier trained only on Llama-family training traces identifies held-out Qwen training runs as close to the reference distribution, while inference and ResNet-50 training are separated at larger distances. BERT-base remains closer to the boundary, suggesting graded similarity among transformer-based training workloads. These results show that hardware-originated telemetry can contain workload-relevant information even without access to model internals, logs, or datasets. They suggest a promising complementary pathway for verification in compute-based AI governance.