General Purpose Buckets: The Original Standard
This is the bucket type everyone knows, the original S3 standard dating back to 2006.
The architecture: It uses a flat namespace. Unlike your computer's file system, there are no actual directories. A file named /images/2026/logo.png is not physically stored inside a 2026 directory. It is simply an object whose full name includes the slashes.
The trade-off: This model offers virtually unlimited scalability, supporting billions of objects without reaching a saturation point. Data is natively replicated across at least 3 Availability Zones (AZs), providing exceptional resilience.
The limitations: Renaming a parent folder containing 100,000 files technically requires copying all 100,000 objects individually, consuming substantial compute resources and API requests (Class A). In addition, latency is measured in milliseconds, which can be too slow for certain AI applications.
Directory Buckets: The Era of Sub-Millisecond Performance
Introduced alongside the S3 Express One Zone storage class, Directory Buckets represent a fundamental shift in AWS storage architecture.
The architecture: They move away from the flat namespace in favor of a genuinely hierarchical structure, similar to a traditional file system. They are also tied to a specific Availability Zone (AZ), bringing storage closer to clusters of EC2 GPU servers.
The objective: Minimize latency. They deliver consistently low, single-digit millisecond response times, with access speeds up to 10 times faster than conventional S3.
Use cases: Machine learning model training, 3D rendering, real-time video processing, and in-memory databases that require ultra-fast temporary storage.
Secure all your S3 buckets, from standard to high performance
tamper detects, qualifies, and protects your cloud data, completely agentless and with zero latency across any bucket type.
Table Buckets: Native Big Data Analytics
Modern Data Lakes rely on open table formats such as Apache Iceberg, Hudi, and Delta Lake. Historically, managing these tables required manually handling millions of Parquet files and complex metadata across standard buckets.
The architecture: AWS introduced table buckets (S3 Tables) to integrate the Apache Iceberg format natively at the storage layer. Instead of interacting with the bucket by sending individual files, you can work directly with transactional tables.
The objective: Delegate maintenance tasks, including small-file compaction and cleanup of obsolete snapshots, to S3 itself. The bucket automatically optimizes data layout to accelerate SQL queries executed through AWS Athena, EMR, or external engines such as Snowflake and Databricks.
Vector Buckets: The Engine Behind Generative AI
The architecture: These buckets index embeddings, which are mathematical representations of text or images, directly within object storage using algorithms such as HNSW (Hierarchical Navigable Small World).
The objective: Allow applications to send a vector directly to the S3 API and retrieve semantically similar documents without moving outside the storage infrastructure. This significantly simplifies the architecture of AI agent applications.
The Ultimate Innovation: AWS S3 Files, the File System
The latest major development, recently introduced to bridge the gap between S3 and Amazon EFS/FSx, is called AWS S3 Files. For a long time, traditional applications designed to read disks using standard POSIX requests needed complex third-party software, such as FUSE, to access S3. These solutions often suffered from serious performance and compatibility issues.
S3 Files adds a genuine, AWS-managed file system layer on top of your buckets. It allows you to mount an S3 bucket directly on an EC2 instance or an EKS container as if it were an ultra-fast local disk. It handles file locks, directory-level permissions, and random writes within a file, which standard S3 General Purpose buckets do not support natively and which otherwise require uploading the entire file again.
Conclusion
The idea that S3 is merely a “cloud FTP folder” is now completely outdated. In 2026, cloud architecture requires carefully selecting the bucket type that best fits each workload. Using a Directory Bucket for archival storage would be a major FinOps mistake, just as using a General Purpose Bucket to train a large language model (LLM) could significantly slow down GPU clusters. Choosing the right bucket is the first step in designing your architecture.