Native NFS v4.1 access directly on S3 buckets — backed by EFS-powered caching — erases the decade-long divide between object storage and high-performance file systems.
For over a decade, cloud architects lived in a bifurcated world: S3 for massive, cost-effective object storage, and EFS or FSx when applications demanded a real file system. The tax of bridging them — DataSync jobs, S3 File Gateways, pre-hydration pipelines before every training run — was simply accepted as the cost of doing business at scale.
Amazon S3 Files changes that entirely. By embedding native NFSv4.1 support directly into S3 buckets, backed by EFS-powered caching, AWS has unified object and file storage into a single layer. The same bucket can now serve REST API consumers and NFS-mounted ML training jobs simultaneously — no synchronization, no duplication.
Adjust the sliders to match your workload. Costs update instantly and the lowest-cost option is highlighted automatically.
Calculating…
S3 buckets can now be mounted as POSIX file systems using standard NFS v4.1 clients — no gateway appliance, no data duplication, no synchronization delay. The S3 bucket is the file system.
The EFS cache layer is the key technical detail. AWS maintains an active working set of metadata and small files (<128 KB) in a high-performance cache, delivering sub-millisecond latency for hot data. Large files stream directly from S3, bypassing the cache — which is why large-file NFS reads are essentially free.
PyTorch and TensorFlow data loaders work unmodified. GPUs stay saturated without pre-hydration waits. Throughput scales to multiple TB/s in aggregate.
AI agents using file-based Python tooling can share state and persist memory across distributed pipelines without custom S3 API wrappers.
Apps that assume a POSIX file system — CMS platforms, ERP systems, monolithic backends — can move to S3 without code changes.
Eliminate separate EFS volumes for configs and S3 buckets for assets. One bucket serves NFS-mounting services and REST clients simultaneously.
HPC workloads can use S3 as shared scratch space with NFS semantics, without a dedicated parallel file system like FSx for Lustre.
ETL pipelines writing intermediate results use standard file operations against S3 — no choosing between object semantics and temporary EFS volumes.
From "Copy-to-Compute" to "Compute-on-Data." The GPU cluster comes to the data, not the other way around. For large-scale ML where dataset sizes dwarf GPU memory by orders of magnitude, this is a fundamental change in training economics.
You pay standard S3 storage rates for all data at rest, plus data-access surcharges for the caching and movement overhead introduced by NFS operations. The large-file free-read rule is the detail that makes the economics work for ML workloads.
| Component | Rate | Notes |
|---|---|---|
| Data at Rest (S3 Standard) | $0.023 / GB‑mo |
Standard S3 pricing. S3-IT tiers apply as normal. |
| Cache Storage (EFS layer) | $0.30 / GB‑mo |
Only charged for the active working set — typically 1–5% of total bucket size. |
| NFS Writes | $0.06 / GB |
All NFS writes go through the cache first, then sync to S3. |
| NFS Reads — Small (<128 KB) | $0.03 / GB |
First-read "import" costs $0.06/GB; subsequent cached reads $0.03/GB. |
| NFS Reads — Large (>128 KB) | FREE |
Streams directly from S3. Standard GET request fees only (~$0.0004/1000 GETs). |
ML datasets are almost entirely large files — multi-GB Parquet shards, TFRecords, checkpoints. Under S3 Files pricing, large-file reads are essentially free, making it dramatically cheaper than EFS ($0.30/GB-mo) for the massive, read-heavy access patterns that define model training at scale.
The $0.06/GB NFS write fee adds up fast for write-heavy workloads — log rotation, frequent checkpointing, code compilation scratch space. For those patterns, pure EFS or local NVMe is likely cheaper. Use the estimator above to check your specific numbers.
| Feature | Amazon S3 | Amazon EFS | Amazon S3 Files |
|---|---|---|---|
| Core architecture | Object storage | Distributed file system | Object storage + EFS cache layer |
| Access protocol | REST / HTTPS | NFSv4.1 (POSIX) | NFSv4.1 and REST simultaneously |
| Latency | ~1–10 ms (TTFB) | Sub-millisecond | Sub-ms (cached); ~ms (cold miss) |
| Throughput ceiling | Multiple GB/s | Multiple GB/s | Multiple TB/s (aggregate) |
| Storage pricing | $0.023/GB-mo | $0.30/GB-mo | $0.023 + $0.30 (hot data only) |
| Large-file read cost | Near zero | Included | FREE (bypasses cache) |
| Write cost | Near zero | ~$0.06/GB (Perf-Optimized) | $0.06/GB |
| Consistency | Strong (read-after-write) | Close-to-open NFS | Close-to-open NFS |
| Single source of truth | Yes (object API only) | Yes (NFS only) | Yes — both access modes, one bucket |
| Best for | Cold lakes, archives, web | Shared configs, legacy apps | ML training, agentic AI, data prep |
Amazon S3 Files is the most significant change to AWS storage architecture in years. It doesn't just add a feature to S3 — it eliminates an entire category of infrastructure complexity that teams have been building workarounds for since 2006.
For AI and ML teams, the value is immediate: no pre-hydration steps, no managing separate EFS volumes per training cluster, no data duplication costs. The GPU cluster talks directly to S3 via NFS, and the free large-file read path makes it the obvious default for any read-heavy training workload.
For platform engineers, the convergence of object and file APIs on a single data layer enables architectures that weren't practical before — truly unified data planes where the same dataset is queryable by S3 Select, accessible via NFS by legacy apps, and consumed by REST APIs, all without a line of synchronization code.
The open question is whether this makes FSx for Lustre obsolete for extreme-scale HPC. For most teams — even large GPU clusters — S3 Files will be sufficient. The edge cases that need striped, sub-millisecond access to petabytes of scratch space still have a home with Lustre. But that edge just got much smaller.
Official announcement: aws.amazon.com/about-aws/whats-new/2026/04/amazon-s3-files/ · Pricing: aws.amazon.com/s3/pricing/