Data lakes promised enterprises a single, inexpensive place to keep all of their data in its original form and to analyze it with whatever tools suit the task. In practice, many lakes have become expensive. Storage grows without bound, small files overwhelm metadata and query engines, clusters sit idle between jobs, the same data is copied into warehouses and marts, queries scan far more bytes than they need, and nobody can say which team is responsible for which part of the bill. This paper presents a catalog of eight design patterns for building cost-optimized data lakes at enterprise scale. The patterns cover storage tiering and lifecycle management, columnar formats with partitioning and compaction, elastic and interruptible compute, query-in-place with scan reduction, zoned layers with retention, ingestion mode selection, selective materialization of marts, and governance with cost accountability. We position the patterns against the classical warehouse and lake architectures and against a cloud reference architecture, and we evaluate them in a simulation of an enterprise with 1.2 petabytes of data, 8 terabytes of daily ingestion and 120,000 queries per day. Together, the patterns reduced the monthly cost index of a naive lake from 100 to 41, cut median bytes scanned per query by 91%, improved median query time from 18 seconds to 4.2 seconds, reduced data duplication from 3.1 copies to 1.4, and raised the share of cost that could be attributed to an owner from 66% to 94%. The study also quantifies the trade-offs of each pattern, including retrieval charges, interruption overhead and data freshness, and offers a decision guide for practitioners.