AWS Launches S3 Files: Mount S3 Buckets as File Systems with 1ms Latency and 90% Lower Cost
Amazon Web Services has launched Amazon S3 Files, a new capability that makes S3 object storage buckets directly mountable as file systems — eliminating one of the most persistent pain points in cloud-based AI and machine learning infrastructure. The service, which reached general availability on April 7, allows any application built on POSIX file system interfaces to read and write directly to S3 without code changes, data migration, or the cost of maintaining a separate file system layer. For AI teams running large-scale training workloads on AWS, it removes the tradeoff that has forced years of costly workarounds. This is part of the broader AWS infrastructure expansion underway as Amazon races to build the compute and storage backbone for the AI era.
What Amazon S3 Files Does and How It Works
S3 Files creates a file system layer on top of an existing S3 bucket, making it accessible via standard POSIX file system protocols — the same interface used by Linux applications, ML frameworks like PyTorch and TensorFlow, and containerised workloads on Kubernetes. Applications that previously required data to be copied into a separate file system (EFS or FSx) can now mount an S3 bucket directly and access data as if it were a local file system.
The service caches actively used data on high-performance storage for approximately 1 millisecond latency, with data automatically expiring from the cache after a configurable inactivity period of 1 to 365 days. According to the AWS announcement, S3 Files supports 10 million-plus IOPS per bucket, more than 4 TB/s of aggregate read throughput, and up to 25,000 concurrent compute resources mounting the same file system simultaneously.
Why This Matters for AI and ML Workloads
The problem S3 Files solves has plagued ML teams for years. S3 is the default storage layer for data lakes and training datasets on AWS — cheap, durable, and scalable — but ML frameworks require POSIX file systems. The standard workaround was copying training data from S3 into Amazon EFS or Amazon FSx before starting a training job, then syncing results back afterward. This doubled storage costs, introduced sync complexity, and added latency to every iteration of the training loop.
S3 Files eliminates this by keeping the authoritative data copy in S3 while presenting it to compute resources as a file system. Unlike EFS or FSx, there is no data duplication — the object and file interfaces remain synchronised automatically. AWS estimates up to 90% lower costs compared to the traditional S3-plus-file-system pattern, with no provisioned capacity and no minimum commitments. For teams running large-scale model training on Amazon EKS or SageMaker, this is a material reduction in both cost and operational overhead. It connects to the $200 billion AWS infrastructure investment outlined in Jassy's shareholder letter — building the storage layer to match the scale of AI demand.
Availability and Use Cases
S3 Files is available in all 34 commercial AWS Regions at general availability. It supports EC2 instances, Amazon ECS containers, Amazon EKS Kubernetes pods, AWS Lambda functions, and any Linux-based AWS compute resource. The use cases span ML training pipelines that need file system access to large datasets, AI agents that must read and write files at scale, data lake analytics workloads using file-based tools, and multi-tenant applications where thousands of containers need simultaneous access to the same large dataset.
The 25,000 concurrent mount limit is particularly relevant for distributed training across large GPU clusters — a common configuration for foundation model training where hundreds of instances simultaneously read training data from a central store.
Frequently Asked Questions
What is Amazon S3 Files and how is it different from Amazon EFS?
Amazon S3 Files makes S3 buckets mountable as file systems — but unlike EFS, the authoritative data copy stays in S3 rather than being duplicated into the file system. S3 Files is ideal when your data already lives in S3 and you need file system access without migration or sync. EFS is better suited for workloads that primarily generate and consume data via file interfaces from the start.
Can I mount an S3 bucket as a file system on EC2?
Yes. Amazon S3 Files lets you mount an S3 bucket directly on EC2 instances, as well as on Amazon ECS, EKS, Lambda, and any Linux-based AWS compute. No code changes are required — applications use standard POSIX file system calls and S3 Files handles the translation to S3 object operations transparently.
How much does Amazon S3 Files cost compared to EFS?
AWS estimates S3 Files can reduce costs by up to 90% compared to maintaining a separate file system alongside S3. Pricing is pay-as-you-go based on active cache usage, with no provisioned capacity or minimum commitments. Idle data reverts to standard S3 pricing after the configurable inactivity period expires.
The Bottom Line
Amazon S3 Files solves a long-standing infrastructure problem that has added cost and complexity to AI workloads on AWS for years. By making S3 directly mountable as a file system — without duplicating data or changing application code — AWS has removed one of the key friction points in cloud-based ML infrastructure. For teams training large models or running AI agents at scale, it is a meaningful reduction in both operational overhead and storage cost.