Training and inference workloads
Distinction between the intensive learning phase of an AI model and the operating phase producing responses.
Training and inference workloads correspond to the two main usage regimes of artificial intelligence infrastructure. Training is the model's learning phase, extremely compute-intensive, concentrated in long sessions mobilizing thousands of processors in parallel, with very high thermal density. Inference is the operating phase, where the model produces responses to queries, less intensive per query but massively distributed and continuous. This distinction shapes the infrastructure risk profile: training concentrates value and heat in large specialized centers, while inference often deploys at the edge close to users. For insurance, the two regimes call for distinct analyses: training follows a critical high-density site logic, inference a distributed fleet and service-availability logic. The split between the two determines the nature and location of exposure.
Hypothetical case: a training-dedicated center concentrates peak thermal risk, while a network of inference servers spreads service-availability risk across many sites.
Follow the research
One letter a week. First, the week's underwriting note, when one is published: a dated, argued decision, signed by Paul Bernardski. Then what the observatory published on cyber risk, political risk and insurance. No account, no tracking.
No spam, one-click unsubscribe, no data ever sold.