S3 at 20: The 5 Updates That Change What AWS Storage Even Means

Amazon S3 turned 20 this year. What launched in 2006 as a simple way to store and retrieve files now holds 500 trillion objects, handles 200 million requests per second, and costs 85% less than it did on day one.

But the updates AWS shipped in the last few months aren't incremental. They redefine what S3 is. Object storage? That's just the starting point now.

Here are the 5 updates that matter — and what they mean for your architecture.

1. S3 Files: Mount S3 as a File System

Launched: April 7-8, 2026 — GA in 34 regions

This is the big one.

S3 Files lets you access your S3 buckets via NFS v4.1 and v4.2. Not through a FUSE adapter. Not through a third-party gateway. Native NFS mount, managed by AWS.

Any application that reads or writes to a file system can now use S3 as its backend. Legacy apps, shared file workflows, media processing pipelines — no code changes needed. Just mount the bucket.

The numbers

• Latency: 1ms for cached reads (cache sits on local SSD on the mount endpoint)
• Connections: Up to 25,000 simultaneous NFS connections per mount
• Storage: $0.30/GB/month for cached storage, standard S3 pricing for cold data
• Regions: Available in 34 AWS regions at launch

Why it matters

EFS (Elastic File System) has been the go-to for shared file storage on AWS. It costs $0.30/GB/month for infrequent access. S3 standard is $0.023/GB/month.

If your workload doesn't need POSIX-strict semantics, S3 Files gives you a 13x cost reduction with the same NFS interface. For workloads that do need strong consistency, the cached storage tier provides it.

Werner Vogels wrote a blog post calling this "the changing face of S3." When the CTO writes about it personally, it's not a minor feature.

Who should care

• Teams running EFS for shared file access across EC2 instances
• Any workflow that pipes data through NFS before landing it in S3 anyway
• ML teams with training pipelines that read from mounted file systems
• Legacy applications that can't be rewritten to use the S3 API

2. S3 Vectors: Native Vector Storage

GA December 2025, expanded to 31 regions March 2026

S3 now has a native vector storage and retrieval engine. Not a separate service. Not a marketplace add-on. Built into S3.

How it works

You create a "vector index" within an S3 bucket. Each index supports up to 2 billion vectors. You insert vectors with metadata, and query them using approximate nearest neighbor search. Results come back in under 100ms.

No servers to provision. No clusters to manage. No capacity planning. You pay for what you store and what you query.

The pricing play

AWS claims S3 Vectors is 90% cheaper than standalone vector databases for large-scale storage. The economics make sense: S3's storage layer is already the cheapest durable storage in the cloud. Adding a vector index on top avoids the compute overhead of running dedicated vector DB instances.

The implications

If you're running Pinecone, Weaviate, Qdrant, or Milvus on AWS to store embeddings for RAG pipelines, semantic search, or recommendation engines — your cloud provider just built the same capability into their cheapest storage layer.

This doesn't replace every vector database use case. If you need sub-10ms queries, complex filtering, or real-time index updates at high throughput, a dedicated vector DB still wins. But for the majority of use cases — batch-loaded embeddings queried at moderate frequency — S3 Vectors is hard to argue against.

Who should care

• AI/ML teams storing embeddings for RAG applications
• Anyone paying for managed vector database services on AWS
• Teams building semantic search without wanting to manage infrastructure

3. Account Regional Namespaces: The End of Bucket Squatting

Launched: March 2026

For 20 years, S3 bucket names have been globally unique. If someone created my-app-data in us-east-1, nobody else on the planet could use that name. Ever. In any region.

That's over.

The new format

Buckets can now use account-scoped names: {prefix}-{account-id}-{region}-an

The suffix ties the bucket to your account and region. Two different AWS accounts can now have buckets with the same prefix, because the full name is unique to the account.

Why this matters more than you think

• No more bucket squatting: Bad actors can't camp on common bucket names to intercept traffic or confuse deployments.
• Clean Terraform modules: Infrastructure-as-code templates no longer need random suffixes or globally unique name generators. The account ID provides uniqueness automatically.
• Multi-account organizations: Enterprise orgs with hundreds of accounts can use consistent naming conventions without collisions.
• Security: Eliminates the class of attacks where S3 bucket names are guessable and can be pre-created by attackers (bucket namespace hijacking).

Who should care

• Platform/DevOps teams managing multi-account AWS organizations
• Anyone who's ever added -v3-final-FINAL to a bucket name
• Security teams concerned about bucket takeover attacks

4. 50 TB Maximum Object Size

Launched: December 2025

The maximum size of a single S3 object went from 5 TB to 50 TB. A 10x increase.

Why it matters

5 TB was fine for most workloads. But machine learning model weights are getting larger. Genomics datasets routinely exceed 5 TB per file. High-resolution media archives hit the limit.

The workaround was chunking: split the file, upload in parts, reassemble on retrieval. It works but adds complexity, increases error surface, and requires custom tooling.

50 TB eliminates the chunking requirement for most large-file workloads.

Who should care

• ML teams working with large model files
• Genomics and scientific computing
• Media companies with large archive files
• Anyone maintaining custom chunking logic for S3

5. SSE-C Disabled by Default

Launched: April 6, 2026

Server-Side Encryption with Customer-Supplied Keys (SSE-C) is now disabled by default on new S3 buckets.

Why it matters

SSE-C lets you provide your own encryption key when storing objects. In theory, this gives you full control. In practice, it created a ransomware vector.

The attack: a malicious actor with write access to your bucket re-encrypts your objects using their own key. You lose access to your own data. They demand payment for the key.

By disabling SSE-C by default, AWS blocks this attack pattern on every new bucket. Existing buckets aren't affected — this only applies to new ones. But the default change means new deployments are secure out of the box.

Who should care

• Security teams and compliance officers
• Anyone setting up new AWS accounts or environments
• Teams that don't explicitly need customer-managed encryption keys

The Bigger Picture

S3 launched on March 14, 2006, with 8 objects. Twenty years later, it holds 500 trillion.

But the quantity isn't the story. The capability shift is.

S3 started as object storage — a place to PUT and GET files. Over 20 years, it has become:

• A file system (S3 Files — NFS access)
• A vector database (S3 Vectors — native embedding storage)
• A data lakehouse (S3 Tables — Apache Iceberg support)
• A metadata engine (S3 Metadata — live inventory and journal tables)

The pricing has dropped 85% while the capabilities have expanded 10x.

For builders, the implication is simple: before reaching for a specialized service, check if S3 can do it natively. Increasingly, it can. And at S3 prices, the cost difference is significant.

---

I'm Anas. I build Fleetzz, TheVerifico, and other products. Follow me for more on building with modern cloud infrastructure.