Roland Berger

Transformation as Strategy, Ep. 14 Smoothing the Surge: Powering the Next AI Data Center Generation

The podcast explores how AI data centers, especially those training large language models, differ from traditional cloud facilities in their electricity demand. Training workloads generate highly synchronous, sawtooth‑shaped power consumption as GPUs cycle through forward, backward, and optimizer phases, while inference workloads resemble conventional cloud patterns with diurnal peaks and valleys. Spencer Gore explains that this variability creates two major challenges: rapid, gigawatt‑scale ramp‑up and ramp‑down events during training cycles and checkpoint operations, and the need to meet utility‑mandated ramp‑rate and low‑frequency ripple limits. Conventional countermeasures—throttling GPU clocks or fabricating dummy work—either slow training or increase heat, making high‑performance batteries the preferred solution for absorbing excess energy and dispatching it during drops. Key examples include checkpoint periods where power can fall from a gigawatt to a few hundred megawatts within minutes, and utility requirements that prohibit more than 10‑30 MW per minute ramp rates. Batteries also enable fault‑ride‑through, allowing data centers to sustain 90 % of load during grid voltage dips, and they protect on‑site generators from fatigue caused by erratic AI loads. The implications are clear: AI‑focused data centers will increasingly embed high‑rate, long‑duration battery systems, ranging from five‑minute UPS‑style buffers to multi‑hour storage that can replace diesel generators. Battery manufacturers targeting this fast‑growing niche must prioritize cycle life, power density, and grid‑compliance features to capture gigawatt‑hour contracts that could define the next wave of AI infrastructure investment.