Foundation models, or large AI models, are crucial for advancing generative AI applications. However, existing large language models (LLMs) struggle with understanding observability metrics, which require processing numerical time series data and identifying trends, seasonality patterns, and anomalies. Dedicated foundation models for time series and structured data have the potential to complement general-purpose LLMs. An example of such a model is Toto, developed by Datadog, which achieves top performance on several open time series benchmarks and consistently outperforms existing models in key accuracy metrics. Toto is trained on nearly a trillion data points, including 750 billion fully anonymous numerical metric data points from the Datadog platform and time series datasets from Large-scale Open Time Series Archive (LOTSA). It excels in zero-shot forecasting and generalizes well to other time series domains. Toto outperforms existing foundation models and full-shot models on several benchmarks, showcasing its reliability and precision in forecasting capabilities. The model is still early in its development but holds promise for improving AI, ML, anomaly detection, and forecasting algorithms within the Datadog platform and powering products such as Watchdog and Bits AI.