AWS S3 Tables now supports Intelligent-Tiering natively, automatically moving Apache Iceberg table data between access tiers based on query patterns — reducing storage costs by up to 80% without performance impact. Combined with streamlined IAM permissions across S3 Tables, Glue Data Catalog, and compute (Athena, Redshift, SageMaker), AWS is building a complete “lakehouse storage layer” that self-optimizes. The strategic play: make S3 Tables so frictionless that it becomes the default for structured data, displacing both self-managed Iceberg and competing analytics platforms.

Amazon ECS now supports mounting EBS volumes directly to containers in AWS GovCloud Regions, enabling storage-intensive workloads (ETL, media transcoding, ML inference) to run on serverless Fargate with persistent block storage in sovereign environments. This closes a gap that previously forced GovCloud customers to use EC2 for any workload requiring persistent volumes, and signals AWS’s intent to make Fargate viable for data-intensive government and regulated workloads.

AWS launched a new Local Zone in Istanbul, enabling organizations to store and back up data locally within Türkiye to meet data residency requirements. Local Zones provide EBS and S3 access with single-digit millisecond latency to local end users. For cloud storage buyers, this matters because data residency is increasingly becoming a hard constraint — not a preference — and Local Zones are AWS’s answer to the “data must stay in-country” mandate without sacrificing the full AWS service portfolio.

Microsoft introduced a minimum billable object size of 128 KiB for Azure Blob and Data Lake Storage cool, cold, and archive tiers. Objects smaller than 128 KiB are now billed as 128 KiB at the tier’s rate. This materially impacts workloads storing millions of small files (IoT telemetry, log fragments, metadata) in cooler tiers. The change reflects Microsoft’s position that cold/archive economics assume large-object patterns — teams with small-object architectures need to batch before tiering or accept increased per-object costs.

Azure Blob Storage’s SFTP endpoint now supports Microsoft Entra ID-based access in public preview, replacing the previous local-user-only model (password or SSH key). This enables organizations to authenticate SFTP connections with their existing identity provider, eliminating credential management overhead for file transfer workflows. For enterprises operating SFTP-dependent data pipelines (healthcare, financial services, supply chain), this removes a significant security friction point and brings Blob Storage SFTP in line with Azure’s broader zero-trust identity model.

Microsoft’s grace period ended: Azure Unmanaged Disks were fully retired on March 31, 2026 (extended from the original September 2025 deadline). Customers who did not migrate to Managed Disks are now seeing service impact. The forced migration pushes organizations into Premium SSD v2 and Ultra Disk territory — newer tiers with independent IOPS provisioning but also more complex sizing and billing models. Lucidity’s analysis notes that forced migrations almost always result in more provisioned capacity (teams overprovision to avoid breaking things during cutover), making post-migration right-sizing essential.

Alluxio released AI v3.9 with a POSIX write cache delivering 7.6 GiB/s peak single-node write throughput (20 GiB/s across 3 nodes) with sub-2ms P99 latency, plus RDMA read support achieving 49.5 GB/s on 400 Gbps NDR InfiniBand (99% of line rate) with sub-100µs P99 on small reads. The update works transparently with all major training frameworks and falls back to TCP if RDMA hardware isn’t present. This directly addresses the checkpoint-write bottleneck that gates AI training step time — the storage layer is no longer the constraint at 400G network speeds.

At Google I/O 2026, Google Cloud reinforced its storage-as-AI-infrastructure positioning by integrating Storage Intelligence with the broader Agentic Data Cloud. The reimagined platform provides zero-configuration dashboards, aggregated activity views, and enhanced batch operations that make storage management agent-ready. Combined with the MCP server GA’d at Next ’26, Google now has the most complete “agent-accessible storage” story across the three hyperscalers — storage that AI agents can discover, read, write, and analyze without human orchestration.

NAND flash maker Kioxia reported record revenues driven by AI-related storage demand and announced plans for a US stock listing. SK Hynix’s stake in Kioxia (via convertible bonds) has reached ~$40 billion in value. The broader signal: AI storage demand is real enough to underwrite IPOs. Enterprise SSD and high-capacity QLC deployments for AI training data (exemplified by KAYTUS’s all-QLC system delivering 10 TB/s aggregate bandwidth at exabyte scale) are driving the next wave of NAND investment and pricing power for storage manufacturers.

Wasabi’s annual survey of 1,700 IT decision-makers reveals that nearly half of organizations exceeded their cloud storage budgets in the past year, with 91% citing fee-related complexity (API calls, retrieval charges, egress) as the cause — not raw storage capacity. Additional findings: 72% of organizations estimate at least 25% of their storage is “dark data” (stored but never analyzed), 60% plan to increase AI infrastructure spending, and 64% are deploying hybrid storage architectures for AI workflows. The report confirms that cloud storage cost management remains structurally broken for most enterprises.

A cross-cloud analysis from Lucidity documents how all three hyperscalers’ H1 2026 storage announcements (higher IOPS ceilings, new tiers, forced migrations) are collectively increasing enterprise spend despite marketing as “free performance.” Key insight: average enterprise block utilization remains 15–30%, meaning 70% of provisioned storage sits idle. Higher performance ceilings give DevOps teams license to provision bigger volumes “just to be safe,” while forced migrations (Azure Unmanaged → Managed) almost always increase provisioned capacity during cutover. The diagnosis: faster storage doesn’t save money; smarter storage does.

At Dell Technologies World, Eli Lilly showcased their 15-year Dell partnership now extending into AI at scale: Dell PowerStore feeds “LillyPod,” their Nvidia DGX SuperPOD with 1,000+ GPUs, at nearly 2 TB/sec of read bandwidth. Meanwhile, Mazda deployed Dell PowerScale to consolidate 30 years of engineering data into a 10 PB AI/GenAI data lake, achieving 90% cost reduction per storage unit while eliminating tape offload. These reference architectures demonstrate that on-prem enterprise storage at AI scale is not just viable — it’s delivering economics that challenge cloud-only approaches for steady-state workloads.