S3’s object-size ceiling has jumped 10x, from 5 TB to 50 TB, purpose-built for the massive training-dataset objects foundation-model teams increasingly want to store as single units rather than sharded across thousands of smaller objects. It is a quiet, infrastructure-level change rather than a headline product, but it removes a real friction point for AI teams whose object sizes were pushing against the old ceiling.

Query processing charges for S3 Vectors indexes above 10 million vectors fell as much as 80% earlier this summer, sharpening AWS’s pitch that native vector storage inside S3 beats a standalone vector database on cost for large-scale RAG workloads — AWS claims up to 90% total cost reduction versus dedicated vector DB services. Analysts remain split on whether it complements or cannibalizes the vector-database market, but the pricing trajectory shows AWS treating this as a strategic front, not a side feature.

gp3 remains 20% cheaper per GB than gp2 while including 3,000 IOPS and 125 MB/s of throughput at no extra charge, a baseline still undercutting Azure Premium SSD v2 and GCP pd-balanced on list price. Not new news, but a useful anchor point given how much attention this cycle has gone to Azure and Google’s AI-storage announcements — AWS’s core block-storage economics haven’t moved and remain the comparison floor.

Enterprises can now consolidate data directly from Blob Storage into Foundry IQ for grounding enterprise knowledge, fine-tuning models, and serving low-latency context to inference — collapsing a step that previously required a separate ingestion pipeline. It is the clearest sign yet that Microsoft is positioning Blob as core AI infrastructure rather than a commodity object store, and it raises the switching cost for Azure-anchored enterprises building AI pipelines.

The Azure Storage Blobs SDK changelog shows a dense cluster of governance-adjacent features landing together: Blob Smart Tier, directory-level and user-delegation SAS, content validation via structured message, conditional-tier blob deletion, and server-side encryption rekeying. Individually minor, collectively it reinforces the same theme as Storage Discovery’s GA — Azure is building out access control and lifecycle governance depth ahead of raw throughput.

Microsoft says the underlying Blob Storage architecture has been evolved to deliver exabytes of capacity, tens of Tbps of throughput, and millions of IOPS feeding GPU clusters directly — infrastructure-level work aimed squarely at large-scale AI training rather than general enterprise storage. Context for the Foundry IQ integration above: Microsoft is investing at the architecture layer, not just the product layer, to compete for AI training workloads.

Multi-region Nearline at-rest pricing rose from $0.010 to $0.015 per GB-month — a 50% increase — while multi-region Archive pricing fell as much as 40%, to $0.0024/GB in the US and EU and $0.0030/GB in Asia. The net effect reshapes the lifecycle-tiering math for any workload straddling Nearline and Archive: Archive is now cheaper relative to Nearline than it has ever been, and policies that route cooling data through Nearline first should be re-priced against the new curve.

The new GCSFS release dynamically predicts and background-fetches sequential read patterns by default, boosting single-file throughput 5x and scaling up to 21 GiB/s when paired with Rapid Bucket — enough to saturate the NIC. Aimed directly at eliminating GPU data-starvation in PyTorch training pipelines, it is a narrow release with an outsized effect on effective training-cluster utilization.

In buckets with hierarchical namespace enabled, the Object Lifecycle Management Delete action will start removing empty folders once a lifecycle rule’s conditions are met — a small housekeeping change, but one that matters for teams relying on folder structure for downstream tooling and should be checked against existing lifecycle rules before it lands.

Google has stopped charging data-transfer fees for UK customers moving data to a competing cloud provider, moving ahead of the EU Data Act’s January 12, 2027 deadline that will bar all switching fees, egress included, for EU customers. Google was also first to waive full-exit egress fees back in January 2024; this UK move suggests the hyperscalers are increasingly competing on portability terms ahead of the regulatory floor, not just meeting it. Worth tracking whether AWS or Microsoft follow with similar early moves before year-end.

A survey of 1,700 IT decision-makers worldwide — the fourth consecutive year of Wasabi’s Global Cloud Storage Index — found storage-related fees now make up about 49% of the average bill, essentially matching raw capacity spend. TCO models anchored to per-GB list price alone are understating real cost by roughly half; fee structure, not sticker price, is now the bigger lever for enterprise storage budgets.

The cloud storage market has reached roughly $173B in 2026, with object storage the fastest-growing segment at a 19.1% CAGR, driven primarily by AI training data plus backup and archive volume rather than traditional application storage. The divergence is a clean signal that AI workloads, not conventional enterprise IT, are now the marginal buyer shaping hyperscaler storage roadmaps.