Cloud Storage Briefing

S3 Vectors can now apply metadata filters before the similarity search rather than after, which AWS says lifts recall by up to 5x on filtered queries. The feature launched September 30 and reached GovCloud on October 9. Post-filtering was the main quality gap against dedicated vector databases for retrieval with tenant, date or permission filters. Closing it makes “keep embeddings in S3” a credible default for RAG at object-storage prices, and puts pressure on standalone vector stores.

S3 Tables now handle every Iceberg V3 data type, including variant and geospatial, and raise the limit to 100 table buckets per Region per account. On October 5 Snowflake added S3-compatible storage support for Iceberg tables in its Horizon catalog. Iceberg on object storage is becoming the neutral table layer that engines compete on top of. That shifts lock-in from the storage layer to the catalog, which is where AWS, Snowflake and Databricks are now fighting.

GitHub is replacing Spokes, which stored full repository copies on local disks and needed a quorum of replicas to acknowledge every push, with an architecture that writes each commit once to Azure Blob Storage and lets the object store handle redundancy. Internal tests show 35x higher write throughput. The trigger is agents: 7.38B commits in September, more than five times a year earlier. It is the clearest example yet of agent workloads pushing write-heavy systems onto object storage, with Microsoft’s own flagship developer platform as the reference customer.

New and upgraded Family and Premium subscriptions now share one 2TB pool instead of 1TB per person for up to six people. Existing subscribers move after their first renewal on or after May 2, 2027. Coverage ties the change to industry-wide storage shortages and rising hardware costs. It is a consumer plan, but it shows Microsoft re-pricing included capacity rather than absorbing higher flash and disk costs. Enterprise buyers should expect the same logic in fewer price cuts and tighter included quotas, not in list-price increases.

Customers reserving H4D instances through their Google account team can now include a Hyperdisk pool in the same or a later request, so block storage capacity is guaranteed next to the compute. As accelerator and HPC capacity is sold on reservations, storage that is not reserved alongside it becomes the bottleneck at launch. Google is the first of the three to package the two together on the reservation path. Expect AWS and Azure to follow for their HPC and GPU capacity blocks.

Z3 storage-optimized VMs with 8 to 44 vCPUs can now attach Hyperdisk Balanced HA volumes, which replicate synchronously across two zones. Before, regional HA block storage on Z3 required the larger shapes. Mid-size databases and stateful services on Z3 can now get zone-failure protection without over-buying compute, which narrows the gap with Azure’s zone-redundant disks and AWS’s Multi-AZ managed services.

Kioxia’s LD4 pairs eighth-generation BiCS QLC NAND with the E1.L ruler form factor. It is sampling at 15.36TB and 30.72TB, with the architecture validated to 122.88TB, and will be shown at the OCP Global Summit (October 12–15). It targets read-heavy AI data repositories and object storage. With flash contract prices still rising, density per rack unit is how hyperscalers hold down the cost per TB of their hot object tiers, and it is the main lever that can slow tiering to HDD.

K2, in public beta, stores a durable, ordered log as segment files on R2, uses R2’s atomic operations for ordering, and runs with no brokers. The trade-off is about one second of p99 produce latency. Pricing is $0.04 per GB produced and $0.04 per GB consumed, with egress free as on R2. It is a direct challenge to MSK, Event Hubs and Pub/Sub on cost for retention-heavy streams, and another case of object storage absorbing a category that used to need dedicated infrastructure.

TrendForce’s October 7 spot update shows 512Gb TLC wafer prices down 2.9% as module houses carry high inventory and weak retail demand, while DRAM spot held flat. The Q4 contract forecast is still up 15–20%, driven by enterprise SSD. The split matters for cloud economics: hyperscalers buy on contract, so their SSD costs keep rising even as spot softens. Do not read falling spot prices as a sign that SSD-backed tier pricing will ease in 2027.