A glowing cloud diagram links AI computing, databases, analytics dashboards, and server racks in a modern data center.
Microsoft has put a new Turbo tier for Azure Blob Storage into public preview. It adds SSD-backed performance to a normal Blob account, and it is aimed at AI, analytics and HPC workloads where storage has become the bottleneck. It is a preview, so it is for non-production use and testing. It is not a generally available service.

A glowing cloud diagram links AI computing, databases, analytics dashboards, and server racks in a modern data center. What Microsoft announced​

Microsoft announced the preview at Ignite on November 17, 2026. In its Azure Storage Blog post, Microsoft describes Turbo as built for AI and other data-intensive workloads that move data between hot, cool and cold tiers and Turbo based on usage patterns. That wording comes from the Azure pricing page. The pricing page also says data stored on Turbo is billed at the rates in its data storage prices section.

The pitch is that you no longer have to pick one of two options:

  • Put everything on expensive SSD-backed storage.
  • Run separate storage systems to get performance for only part of your data.

Turbo is an access tier inside a standard storage account. It is not a cache and not a separate high-performance account. Applications keep using the same account endpoints, namespace, Blob APIs, SDKs and tooling.

Microsoft's Learn documentation on access tiers has already added Turbo, marked as a preview. It says Turbo adds SSD-backed performance for low-latency, high-throughput workloads such as AI, machine learning and interactive analytics. It also says you can choose the tier for each blob inside the same general-purpose account.

The Azure Updates page tied to this item showed no entries when it was captured. The product details here therefore come from Microsoft's Azure Storage Blog post and the Learn documentation.

How Smart placement works​

Turbo comes with a feature called Smart placement. You set a Turbo capacity target on the storage account. That target sets the size of your high-performance working set. You can set it through the Azure portal, Azure CLI, Azure PowerShell or management-plane REST APIs.

Smart placement then manages eligible data in three ways:

  1. Land fast. When an account is configured for Turbo, new data lands directly in Turbo.
  2. Cool as demand changes. Eligible data moves back to HDD-backed tiers when it no longer needs Turbo performance.
  3. Accelerate when needed. Eligible data is promoted from HDD-backed tiers into Turbo as access patterns change.

If you already know which data needs Turbo, you can skip Smart placement. You can then move data yourself with explicit Blob tiering operations.

Microsoft says "eligible" repeatedly. It never claims that every blob qualifies for automated movement.

Don't confuse this with the separate smart tier feature that went to preview in November 2025. That feature is a fully managed, automated tiering solution aimed at optimizing storage costs. Microsoft's Learn page describes it as moving data between the hot, cool and cold tiers. Turbo is a different, performance-oriented tier.

Preview limits​

The restrictions are tight, and they matter more than the headline numbers:

  • Eligibility is limited to general-purpose v2 (GPv2) accounts.
  • Those accounts must use locally redundant storage (LRS).
  • The accounts must be in selected regions, subject to regional capacity.
  • You must request access to the preview.

Microsoft's announcement does not list the eligible regions. It also doesn't give quotas, a service-level commitment or a general availability date. The Learn documentation says eligible preview customers can get additional documentation and onboarding guidance.

Pricing is also unclear. The Azure pricing page mentions an enablement fee per Turbo-enabled storage account, but the figure shown in the search result was blank. I can't tell you what that fee will be. Check current pricing before you plan a budget.

Microsoft's benchmark claims​

All of these numbers are Microsoft's own, and Microsoft says they are specific to its test configurations.

  • Random reads. In a synthetic 4 KiB random-read test, Turbo reached 10.1 million transactions per second. Standard Blob on HDD-backed tiers reached 601,000, which Microsoft calls 16.8 times the rate.
  • Worker VMs. The test used 72 Standard D96s v6 worker VMs for Standard Blob and 256 for Turbo. Microsoft says the Standard Blob setup used fewer workers to stay within HDD-backed account limits. That makes the comparison less like-for-like than the headline ratio suggests.
  • Read bandwidth. In a separate benchmark, Turbo sustained more than 20 Tbps of read throughput. Microsoft ties this to checkpoint recovery, where large GPU clusters reload model weights ranging from several GiB to tens of TiB.
  • Caveat. Microsoft says the benchmarks measure storage performance, not end-to-end workload speedups. Results vary with workload, object size, concurrency, network conditions and account configuration.

The cost example​

Microsoft gives an illustration based on a 1 PiB dataset with roughly 20% active, high-performance data. That works out to about 205 TiB in Turbo and 819 TiB in Hot. In East US 2, for a full month, Microsoft compares two setups:

PlacementIllustrative monthly capacity cost
1 PiB all-SSD baseline$115,343
20% Turbo + 80% Hot$37,500
Illustrative reductionabout 67%

The all-SSD baseline uses $0.11 per GiB. Microsoft says this covers capacity cost only. It does not represent total workload cost, a pricing guarantee or a savings guarantee. Compute, transactions, networking and operations all affect the real bill.

The 67% figure also depends on a 20% working set. If your hot data is closer to 60%, the arithmetic changes. Microsoft's claim that "only a fraction" of data typically needs top performance comes from its own analysis of customer workloads, which we can't verify.

Who should test it, and how​

Microsoft names these use cases:

  • AI training and fine-tuning
  • Inference, with fast model loading and scale-out
  • Analytics and HPC pipelines that repeatedly process the active slice of a data lake
  • High-throughput reads, writes and checkpoints

A sensible evaluation plan based on Microsoft's description:

  1. Confirm storage is the bottleneck. If GPUs or compute sit idle waiting on data, Turbo is relevant. If they don't, it probably isn't.
  2. Estimate your working set. Turbo pays off when the hot data is much smaller than the total dataset.
  3. Check eligibility. Your account needs to be GPv2 and LRS, in a supported region with capacity.
  4. Request preview access and test with representative data and access patterns, not synthetic ones.
  5. Confirm pricing, billing and supported operations in Turbo-specific documentation before planning any production use.

One design point to note: the Learn documentation says tier changes apply only to block blobs. Append and page blobs don't support tier changes. Microsoft has not said how Turbo interacts with that rule, so check it before moving data.

Analysis​

Premium SSD-backed Blob storage already exists. Turbo's difference is that it sits in the same account and namespace as Hot, Cool and Cold. For teams that do not want to maintain separate accounts and copy pipelines, that could be a real simplification.

The preview restrictions are the main constraint. LRS-only means no zone or geo redundancy, so Turbo is not yet suited to data that needs that resilience. The tightly scoped, vendor-run benchmarks should also be treated as a ceiling rather than a forecast.

Turbo is worth a test if you run storage-bound AI or analytics workloads on Azure. Everyone else should wait for regional availability, pricing details and independent measurements.

 

References

  1. (In preview) Public Preview: Azure Blob Storage turbo tier Azure Updates 2026-11-17T20:00:32Z
  2. Prijzen voor Azure Blob Storage azure.microsoft.com
  3. Blob Storage Turbo Tier: High performance with object storage scale and economics | Microsoft Community Hub techcommunity.microsoft.com