Amazon S3 Files delivers full NFS v4.1+ file system semantics directly on any general-purpose S3 bucket, accessible from EC2, ECS, EKS, Lambda, and Fargate with up to 25,000 simultaneous active connections. S3 is now the first cloud object store with native high-performance file system access — without data ever leaving S3. For enterprises running parallel S3 + EFS architectures, this eliminates a significant operational and cost layer. AI agents, legacy file-based applications, and HPC workloads can all connect to the same data without architecture bifurcation. GA in 34 AWS Regions.

AWS expanded S3 Vectors capacity forty-fold to 2 billion vectors per index in its GA release, introducing a “Storage-First” architecture that decouples vector compute from storage. By natively integrating vector search into S3’s storage engine, AWS claims up to 90% TCO reduction versus standalone vector databases for large-scale RAG workloads. The strategic move positions S3 as the default vector store for AWS-native AI pipelines, directly threatening Pinecone, Weaviate, and other vector database vendors competing in the retrieval layer.

Amazon S3 marked its 20th anniversary in March 2026, with AWS disclosing that “hundreds of exabytes” of customer data now reside in S3. While AWS does not break out storage revenue specifically, analyst estimates put storage at 20-25% of total AWS revenue — roughly $7-9B quarterly. The anniversary framing signals confidence: S3’s expansion from object store to AI data platform (S3 Tables, S3 Vectors, S3 Files) is the bet that storage economics will compound as AI workloads scale, not commoditize.

Microsoft’s 2026 Azure Storage vision explicitly separates frontier model training (25 PiB AMLFS namespaces, multi-TB/s sequential throughput) from agentic applications (thousands of concurrent agents generating an order of magnitude more I/O than human-driven systems). Elastic SAN becomes a core multi-tenant block storage pool shared across workloads with provisioned guardrails. Native Blob Storage integration into Foundry IQ makes Azure Blob the grounding layer for enterprise knowledge, fine-tuning, and low-latency context serving. Partnerships with ServiceNow, Databricks, and Elastic anchor the agentic scale narrative.

Microsoft launched Azure Storage Discovery as a fully managed service providing enterprise-wide visibility into Azure Blob Storage data estates. The zero-configuration dashboard surfaces cost anomalies, data governance signals from Security Command Center’s DSPM, and storage utilization patterns across billions of objects. 70% of Azure’s largest cloud storage customers already use Storage Intelligence — Discovery extends those capabilities with automated annotation and unified reporting. For cloud storage buyers managing multi-region or multi-subscription data estates, this addresses the operational blind-spot problem without custom tooling.

Azure Storage Mover gained blob-to-blob transfer support and enhanced scheduling capabilities in May 2026 updates, alongside a new Blob Storage SDK for Rust and Storage Action mock runs for testing automation before deployment. These are developer-experience improvements rather than pricing events, but they signal Microsoft’s intent to own the end-to-end migration and automation stack — reducing the need for third-party tools like Rclone or NetApp XCP for Azure-native data movement. The Rust SDK particularly targets performance-sensitive workloads and emerging AI inference frameworks.

Google’s Cloud Storage Rapid is now generally available, delivering 15+ TB/s of bandwidth, 20 million requests per second, and sub-millisecond latency from a single zonal bucket — leveraging Colossus, the distributed storage system powering Gemini and YouTube. Rapid Bucket delivers 50% reduced GPU blocked time and 2.5x faster data loading versus traditional object storage. Checkpoint restores are 5x faster and writes 3.2x faster, directly reducing wasted GPU time during training interruptions. Native integrations into PyTorch and JAX mean no code changes required. Rapid Cache (formerly Anywhere Cache) adds 2.5 TB/s aggregate read throughput with 2.2x faster checkpoint restores for existing buckets.

Google Cloud Managed Lustre now delivers 10 TB/s of throughput — 10x higher than a year ago and 4–20x higher than any competing managed Lustre offering. Powered by C4NX VMs and Hyperdisk Exapools, checkpoints write and restore 2.6x faster than other GCP storage solutions. The new Dynamic tier at $0.06/GB-month serves data from persistent disk rather than object-based caching, eliminating the performance cliff that causes accelerator stalls. A single predictable SKU removes billing complexity. Salesforce’s production deployment confirms B200 GPUs stay fully saturated, translating to faster LLM inference response for agentic workloads.

Google Cloud doubled CDN Interconnect peering egress rates effective May 1, 2026, with North America jumping from $0.04 to $0.08/GB, Europe from $0.05 to $0.08/GB, and Asia from $0.06 to $0.085/GB. Direct Peering and Carrier Peering rates are also increasing. For GCP storage customers using Cloudflare, Akamai, or Fastly as CDNs — or operating carrier-peered networks — this is an immediate cost increase that arrives automatically on May 2026 invoices. Google’s rationale: “significant investments in global infrastructure.” Cloud architects should model total egress exposure across CDN delivery, inter-region analytics, and multicloud data movement paths.

CoreWeave’s Zero Egress Migration (0EM) program covers the hyperscaler egress fees when migrating large-scale datasets to CoreWeave AI Object Storage — with typical savings up to $1M per migration. Once on CoreWeave, customers pay no egress fees regardless of where data is consumed. CoreWeave’s LOTA technology delivers 7 GB/s per GPU throughput. Customers retain active accounts at AWS, Azure, or GCP — no lock-in penalty. For AI labs and enterprises moving GPU workloads to neoclouds, 0EM removes the last major financial barrier to data portability. This is a direct attack on hyperscaler storage lock-in strategy and compounds competitive pressure on AWS/Azure/GCP to match egress pricing.

The global cloud storage market is forecast to grow from $173B in 2026 to $380B by 2031 at a 17.1% CAGR, with object storage outpacing at 24.4% CAGR and already representing 51% of total cloud storage revenue. Public cloud accounts for 63.7% of deployment. The growth drivers are AI/ML workload expansion, hybrid cloud adoption, and enterprise data governance requirements. Critically, roughly half of public cloud storage spend currently goes to non-capacity fees (egress, API calls, retrieval charges) — a structural inefficiency that alternative providers like Wasabi, Backblaze B2, and Cloudflare R2 are actively monetizing against the hyperscaler model.

Enterprise AI initiatives are colliding with data gravity: AI training workloads require massive data co-location with GPU compute, but multi-cloud data estates scatter data across providers. Moving a petabyte between clouds costs up to $80,000 in egress fees alone, creating a hard economic constraint on multicloud AI strategy. Organizations that assumed they could freely move data between AWS, Azure, and GCP for AI training are discovering that data portability is theoretical — practical portability requires either S3-compatible object storage (Wasabi, Cloudflare R2) or explicit egress waiver programs like CoreWeave 0EM. Data architecture decisions made 3–5 years ago are now creating expensive AI bottlenecks.