Grid operators need to know how much power an AI data center can shed, how fast, and at what cost to service quality, but no public dataset answers this. The team will run LLM and multimodal inference workloads on JLSE GPU nodes while sweeping GPU clock frequency, GPU/CPU power caps, batch size, and request concurrency, recording node-level GPU/CPU/DRAM power and energy time series, clock, utilization, and temperature, together with per-request QoS (time-to-first-token, latency, tokens/second). From this, they will fit a power-flexibility function mapping control action to achievable power range and ramp rate, and a performance-flexibility function mapping control action to QoS degradation, then embed both in an optimization model for data center participation in electricity markets and demand response.