With new legacy blob account creation already blocked and October 13, 2026 now exactly one month out, any enterprise still running legacy or GPv1 Blob Storage accounts is down to its final weeks to migrate proactively rather than let Microsoft’s default GPv2 settings — and default billing baseline — get applied automatically. Azure Storage Discovery, still in public preview, is the most direct tool for inventorying an estate before the deadline hits; teams that haven’t started that inventory this week are cutting it close.
The new release makes adaptive concurrent prefetching the default, boosting single-file read throughput 5x and scaling up to 21 GiB/s for PyTorch, Dask, Hugging Face Datasets, and Ray workloads — a direct answer to GPU data-starvation, which is a training-cost problem measured in idle accelerator-hours, not a storage line item. It’s a reminder that the real currency in this year’s AI-storage competition increasingly isn’t $/GB, it’s GPU utilization enabled per dollar of storage spend.
With Google’s GA Model Context Protocol server for Cloud Storage already live and running in production, and Azure Blob still without a native equivalent, re:Invent in Las Vegas is now the concrete, dated checkpoint for whether AWS closes that gap on S3 or lets Google keep the early lead in storage-to-agent integration standards. Enterprises building AI agent architectures around object storage have about eleven weeks before that question gets answered one way or the other.
Amazon’s ~$200B, Google’s $175–185B, and Microsoft’s $110–120B guidance figures are unchanged from prior weeks, which is itself worth noting: after several straight weeks of upward capex revisions across the sector, a quiet week where the numbers simply hold is a mild signal that the guidance cycle may be stabilizing rather than continuing to escalate quarter over quarter.
Google’s new Dynamic Default Storage Classes automate storage-class compatibility across mixed-generation VM clusters in GKE — a small feature next to Hyperdisk or GCSFS, but one that removes a genuinely common source of manual provisioning error for teams running heterogeneous node pools. It’s the kind of unglamorous fix that doesn’t move a pricing headline but does move real operational toil.
Newer market sizing has the cloud storage market growing from roughly $178.8B in 2026 to $688B by 2034 (an 18.2% CAGR), alongside the previously cited $514B-by-2031 estimate — different forecast horizons, but both pointing the same direction. With AWS, Azure, and Google still holding 31%, 25%, and 11% share of the broader infrastructure market respectively, the growth keeps arriving into a market structure that hasn’t meaningfully shifted in years, reinforcing why pricing experimentation (not share competition) is where the real 2026 action has been.
AWS Storage
AWSWith dates now confirmed, re:Invent becomes the specific checkpoint this digest flagged as the “most important cloud storage watch item on the calendar” for whether AWS ships a native S3 MCP server to match Google Cloud Storage’s GA offering. Roughly eleven weeks remain; no AWS signal on the question has surfaced publicly yet.
Event Details →No new AWS block-storage pricing move this week — gp3 remains the cheapest general-purpose baseline of the three clouds, 20% below gp2, with IOPS and throughput included free up to the standard limits. Worth restating precisely because Azure and Google have both moved their AI-tier pricing this year while AWS’s core economics have stayed still.
AWS EBS Pricing →Azure Storage
AZUREThe five-week runway flagged in prior editions is now down to roughly four: any legacy or GPv1 Blob account still standing on October 13 gets converted to GPv2 on Microsoft’s terms, including default tiering and redundancy, which Microsoft’s own documentation warns can raise the bill. Azure Storage Discovery (public preview) remains the fastest inventory path for teams that haven’t started.
Microsoft Learn →The preview combines premium low-latency tiers with Massive Blob’s scalability targets, continuing Azure’s 2026 pattern of layering AI-workload-optimized products on top of its existing Blob line rather than repricing the commodity tier. It’s a further sign that “Blob as AI infrastructure” is now Azure’s default storage marketing frame, not an occasional pitch.
Windows Mode →GCP Storage
GCPThe release targets GPU data-starvation directly across PyTorch, Dask, Pandas, PyTorch Lightning, Hugging Face Datasets, and Ray by dynamically predicting and background-fetching sequential read patterns. For AI training pipelines bottlenecked on storage I/O rather than compute, this is a meaningful throughput unlock without any infrastructure change on the customer’s part.
Google Cloud Blog →The feature removes a common source of manual provisioning error for teams running heterogeneous GKE node pools — a smaller, operational-toil-focused update rather than a capacity or pricing headline, but a real quality-of-life improvement for platform teams managing storage class sprawl.
Google Cloud Blog →Industry
CROSS-CLOUDBeyond the S3 MCP question specific to AWS’s own card above, re:Invent is also the traditional venue for cross-cloud interoperability announcements — worth watching given Azure’s stated intent to join the AWS-Google Cross-Cloud Interconnect effort “later this year,” a window that now points squarely at this event.
Event Details →New research citing surveyed IT leaders puts non-storage fees — retrieval, egress, API operations, replication — at nearly half of total cloud storage spend, corroborating the digest’s recurring point that raw $/GB rates increasingly understate real-world cost by a factor of two to five. Any storage cost review that stops at per-GB pricing is missing roughly half the bill.
Finout →| Metric | EBS gp3 | Premium SSD v2 | pd-balanced |
|---|---|---|---|
| Storage $/GB/mo | $0.08020% cheaper than gp2 | $0.100/GiBHourly billing | $0.100Default disk type |
| Baseline IOPS | 3,000Free — independent of size | 3,000Free — independent of size | 6/GBScales with disk size (500 GB = 3K) |
| Baseline Throughput | 125 MB/sFree — independent of size | 125 MB/sFree — independent of size | 0.28 MB/s/GBScales with disk size |
| IOPS Provisioning | Independent$0.005/IOPS/mo · up to 16K | IndependentSeparately provisionable · up to 80K | Tied to sizeNo independent IOPS control |
| Max IOPS (family) | 256,000io2 Block Express | 400,000Ultra Disk | 120,000pd-extreme |
| Snapshots | Incremental, S3-backedFast Snapshot Restore available | Instant Access (Pv2/Ultra)Instant creation + restore · free until Jul ’26 | IncrementalCloud Storage-backed · regional |
| Dimension | AWS EBS | Azure Managed Disks | GCP Persistent Disk |
|---|---|---|---|
| Price Leader | gp3 at $0.08/GB Cheapest | Pv2 at $0.10/GiB | pd-balanced at $0.10/GB |
| IOPS Flexibility | Independent (gp3) Winner | Independent (Pv2) Winner | Scales with disk size only |
| Latency | io2: sub-ms | Ultra: sub-500µs avg Winner | pd-ssd: single-digit ms |
| Snapshots | Incremental + Fast Restore | Instant Access (Pv2/Ultra) Winner | Incremental, Cloud Storage-backed |
| Shared/Multi-Attach | io2 multi-attach (16 instances) | Broadest: Ultra/Premium/Standard SSD Winner | Multi-writer (up to 2 VMs) |
| Cold Block $/GB | sc1: $0.015/GB Cheapest | Std HDD: ~$0.04/GB | pd-standard: $0.040/GB |
Azure Premium SSD v2 charges $0.0005/IOPS/month above its 3K baseline — that’s 130x cheaper than AWS io2 Block Express ($0.065) and 10x cheaper than AWS gp3 and GCP Hyperdisk ($0.005). At 64,000 IOPS, the gap is staggering: Azure Pv2 costs $30.50/mo for IOPS alone, while AWS io2 charges $4,160/mo and GCP Hyperdisk Extreme charges $640/mo. If your workload is IOPS-intensive and you’re not on Azure, you’re leaving significant money on the table.
| IOPS Level | AWS gp3 | AWS io2 BE | Azure Pv2 | GCP HDB | GCP HD-Ext |
|---|---|---|---|---|---|
| 3,000 | $0 (free) | $195 | $0 (free) | $0 (free) | $30 |
| 10,000 | $35 | $650 | $3.50 | $35 | $100 |
| 64,000 | $305 | $4,160 | $30.50 | $305 | $640 |
| 256,000 | N/A (80K max) | $16,640 | N/A (80K max) | N/A (160K max) | $2,560 |