Objects in S3 can now carry up to 1GB of mutable, queryable JSON/XML/YAML context attached directly to the object, letting agentic workflows discover and reason about data without maintaining a parallel metadata layer. Announced at the AWS Summit in New York, Annotations extends the same land grab as S3 Vectors and S3 Files — AWS keeps pulling more of the AI stack’s plumbing into the bucket itself. For architects, it’s another reason the marginal cost of leaving S3 keeps rising: metadata, vectors, and files are now co-located, and none of it travels cleanly to a competing object store.

While Wasabi and Backblaze have both raised pay-as-you-go cold storage rates this year, S3 Standard and EBS gp3 pricing haven’t moved, even as enterprise SSD contract costs have risen more than 50% quarter over quarter. AWS can absorb the squeeze longer than independents because storage is a small slice of a much larger compute-and-services bill, but the gap between AWS’s flat list price and its rising input costs is a margin story, not a cost-immunity story. Watch Q4 renewal cycles for signs the line stops holding.

Customers can now store and restore EBS snapshots entirely within the Hanoi Local Zone instead of routing them through a parent AWS region, closing a compliance gap for Vietnamese data-residency requirements. It’s a single-market launch, but it’s part of a steady pattern of AWS pairing Local Zone expansion with local snapshot support — a template it will likely repeat as more sovereign-cloud rules land across Southeast Asia.

Storage Discovery now gives a single-pane view across every Blob and Data Lake account in a tenant, with Copilot handling natural-language queries like “which containers haven’t been accessed in 90 days.” The real story is what it feeds: Discovery identifies the cold, orphaned, or overprovisioned data, and Storage Actions — also now GA — executes the tiering or cleanup. Microsoft has effectively productized the FinOps storage-audit workflow consultants used to bill for.

Storage Actions is now generally available as a serverless task engine that runs conditional tiering, deletion, and cleanup operations directly against Blob and Data Lake Storage, billed by objects scanned rather than a managed compute footprint. Paired with Storage Discovery’s GA the same week, Microsoft is bundling detection and remediation into one workflow — a direct answer to the manual tiering audits AWS and GCP customers still mostly do by hand.

Blob Storage’s SFTP endpoint now authenticates through Entra ID instead of locally-scoped credentials, centralizing access control and closing an audit gap that legacy SFTP left open — local users weren’t visible in the same identity and logging plane as the rest of the tenant. It’s a security-hygiene fix more than a feature, but it’s the kind of gap enterprise security reviews flag every audit cycle.

Rapid Bucket is now generally available, running on Colossus — the same distributed storage system behind Gemini and YouTube — and delivering throughput and request-rate ceilings well above standard GCS buckets. It’s Google’s clearest signal yet that AI training and inference workloads get a dedicated, faster storage tier rather than a faster version of the general-purpose bucket, mirroring the tiered-performance path AWS and Azure are also building toward.

The new Dynamic tier prices low-latency parallel-filesystem storage specifically for the burst-read, burst-write pattern of model checkpointing, undercutting what teams previously paid to over-provision a static Lustre allocation for the same workload. Combined with the 10 TB/s throughput ceiling Google announced earlier this year, it’s a direct pitch at the GPU-idle-time problem — storage fast enough to stop being the bottleneck during checkpoint restores.

Smart Storage now auto-generates metadata annotations and exposes them to AI agents through the Model Context Protocol, letting agentic tools query and act on GCS objects the same week AWS shipped a similar capability for S3. The two hyperscalers converging on agent-queryable object metadata within days of each other suggests this becomes table stakes fast — the differentiator won’t be whether a provider offers it, but how well it’s integrated with each provider’s own agent framework.

Wasabi raised pay-as-you-go pricing to $7.99/TB-month on July 1 and Backblaze moved B2 to $6.95/TB-month in May, both citing hardware and infrastructure cost increases as enterprise SSD contract prices have risen 25% to nearly 60% quarter over quarter since late 2025. The cause is AI: Samsung and SK Hynix are reallocating fab capacity from NAND to the far-higher-margin HBM used in AI accelerators, and cloud providers have signed multi-quarter contracts to lock in what’s left. For buyers, the assumption that cold-tier storage gets 8-12% cheaper every year no longer holds for cost-leader providers — budget models built on that assumption need a rewrite.

Storage Mover can now move data straight from S3 into Blob Storage without staging it through an intermediate location, cutting both migration time and the duplicate egress charges that used to make cross-cloud moves expensive. It’s a small technical change with a strategic purpose: every friction point Microsoft removes from leaving AWS storage is a friction point AWS doesn’t control, and multicloud buyers now have one less reason to stay put.

Forecasts now put the cloud storage market at roughly $179B in 2026 growing to $514B by 2031, a 23% CAGR that outpaces most other infrastructure categories — the acceleration is credited to generative-AI pipelines that treat storage as a performance-critical input rather than a commodity expense. That framing shift is already visible in the product roadmaps above: Rapid Bucket, Managed Lustre’s Dynamic tier, and S3 Annotations are all storage vendors competing on AI-workload fit, not just price per gigabyte.